Prajakt Deotale

Tag: enterprise AI

  • Generative AI Was Built to Be Unpredictable. Enterprises Cannot Be.

    Generative AI Was Built to Be Unpredictable. Enterprises Cannot Be.

    Why enterprise AI is becoming the discipline of building deterministic systems around probabilistic intelligence.

    About fifteen years ago, I spent a lot of time around business-process transformation. Much of the work was about removing ambiguity. Define the process, encode the rules, decide how exceptions should be handled, automate the repeatable steps. The aspiration was simple: the same input should travel through the same controls and produce a dependable outcome.

    Generative AI arrived with almost the opposite proposition. Give the system an intent, relevant context and room to reason. It can interpret messy language, combine evidence, draft a response, write code, explore options or plan a sequence of actions without being told every step in advance.

    That difference explains much of the excitement of the past four years. We did not simply get better automation. We got software that can work inside ambiguity.

    The same quality becomes awkward the moment AI moves from a personal productivity tool into a consequential enterprise workflow. A bank may want a model to understand a customer’s circumstances, but the final credit decision cannot be an act of creativity. A network agent can reason across alarms, topology and service history, yet a change to live network infrastructure needs policy, validation and a rollback path. A pharmaceutical team can use a model to explore scientific literature while still requiring regulated evidence to remain traceable and controlled.

    A note on the title. “Unpredictable” is deliberately strong language. I do not mean random. Generative models are probabilistic. Outputs can vary, confidence is imperfect, and a fluent response is not a contractual guarantee. The enterprise is full of places where that variance has a cost.

    The adoption numbers make more sense through this lens

    The current data describe an industry that has moved quickly on access and experimentation, but much more cautiously on authority.

    Stanford’s 2026 AI Index reports that 88% of surveyed organisations used AI in at least one business function in 2025, up from 78% the year before, while 79% regularly used generative AI in at least one function. [1] McKinsey’s newest global survey, published in August 2026, shows what happened next. The share of organisations reporting that AI is now scaling across the enterprise rose from 38% to 44% in a year. Over the same period, the share attributing any EBIT impact to AI stayed at 37%, essentially unchanged. [2] Deloitte found that only 21% have mature governance in place to manage the risks of agentic AI. [3]

    Chart of 2026 enterprise AI adoption statistics: 88% of organisations use AI, 44% are scaling it, 37% report EBIT impact, 21% have mature agentic AI governance
    Figure 1. Access is close to universal. Scaling is improving. Measured value and governance maturity are still the constraint.

    The combination in the middle of that sequence is the part I find most revealing. Scaling moved. Measured value did not follow in the same year. These are self-reported surveys with different fielding windows, so I would treat the percentages as directional rather than as a single market measure. The direction is hard to miss. Capability and usage have moved faster than the operating model required to give AI real authority, and value appears to lag both.

    There is a competing explanation I should put on the table rather than argue around. Flat EBIT impact might mean the value is real but badly measured. It might mean generative AI simply delivers less than we hoped, and that no amount of governance changes the arithmetic. McKinsey’s own reading emphasises organisational change rather than permission.

    I do not think that reading conflicts with mine, because I mean something broad by permission. Permission is not a signature on a risk form. It is the whole set of conditions under which an organisation is willing to let a system carry part of a process: reliability evidence, controls, clear accountability, redesigned handoffs, retrained people and managers who accept a new division of labour. Organisational change is not an alternative explanation to permission. It is most of what permission is made of.

    The measurement critique has more force, and I return to it at the end. If we are still counting seats and use cases, we should not be surprised that the value looks invisible.

    I do not read any of this as evidence that generative AI has failed to deliver. I read it as the normal middle of a platform shift. We discovered the capability first. We are now engineering the conditions under which an enterprise can trust that capability with consequential work.

    The Enterprise AI SHIFT Curve

    A traditional hype cycle is useful for thinking about expectations. It does not quite explain what I see happening inside enterprises. The binding variable is permission: how much authority are we prepared to give a probabilistic system?

    What the SHIFT Curve measures

    That is the idea behind the Enterprise AI SHIFT Curve. The blue line represents what AI can do. Capability rose at extraordinary speed after 2022 and is still climbing. The purple line represents enterprise permission to act. It rises more slowly because authority brings questions of reliability, risk, control and accountability, and it never fully closes on capability. Some things a model can do are things an enterprise should still decline to let it do. The green line is realised enterprise value. It stays modest while permission lags, then rises as organisations become comfortable letting AI take on more of the workflow.

    The Trust Gap is the space between what AI is capable of doing and what an enterprise is willing to permit it to do. I think that gap explains why an organisation can have hundreds of copilots and pilots yet only a handful of systems with meaningful authority. A good demonstration proves that a model can perform a task. It does not prove that the organisation should let it perform that task repeatedly, at scale, with access to production systems.

    The central idea. Enterprise AI is the discipline of building deterministic systems around probabilistic intelligence.

    The five stages

    The five stages describe how that gap changes.

    1. Spark. The magic moment when the capability becomes visible.
    2. Human-led Adoption. Copilots, pilots and use cases spread while people retain decision authority.
    3. Integration Friction. The collision point. Probabilistic intelligence meets enterprises designed around consistency, auditability and accountability.
    4. Foundation of Trust. Where the less visible engineering starts to matter: context, evaluations, model routing, permissions, policy, observability, deterministic controls and human approvals.
    5. Trusted Autonomy. The point at which permission expands because the surrounding system has earned it.

    The probability dynamic underneath the curve matters just as much as the stages. The journey is not from probabilistic AI to deterministic AI. The model does not need to lose the quality that made it useful. The enterprise moves from relatively unbounded probability towards bounded and governed probability. More control around the system can support more freedom inside it.

    The Enterprise AI SHIFT Curve: AI capability rising faster than enterprise permission to act, with the Trust Gap between them and realised value lagging, across five stages from Spark to Trusted Autonomy
    Figure 2. The Enterprise AI SHIFT Curve. AI capability rises faster than enterprise permission to act. The Trust Gap between them is where much of today’s engineering work sits.

    The hard engineering starts at Integration Friction

    Stage three is where many initiatives become uncomfortable. The model has already shown that it can do the work. The enterprise now starts asking questions that a prompt cannot answer. Which data is it allowed to use? Which model should handle which task? What happens when the model is uncertain? Which tools can it call? How much money can it commit? When does a human have to approve the action? How do we prove what happened afterwards?

    The tempting response is to constrain the system until every possible action has been prescribed. That can make the workflow safer, but pushed too far it destroys the reason for using generative AI. If every path is hard-coded, we have rebuilt traditional software with a more expensive component in the middle.

    The better architecture separates the parts of a workflow where probability adds value from the parts where deterministic control is non-negotiable. The model can interpret, reason and generate. The orchestration around it controls context, tool schemas, model selection, identity, permissions, policy and validation. The final action can then be checked against evidence, approval rules and audit requirements.

    Enterprise AI operating model diagram: a probabilistic model surrounded by deterministic orchestration controlling context, tools, identity, permissions, policy and validation
    Figure 3. The model can remain probabilistic while the surrounding orchestration makes important controls and transitions explicit, testable and governable.

    This is where I use the term harness engineering. The word harness is already in circulation for the scaffolding around a model. What I am adding is the claim that this scaffolding is an enterprise discipline in its own right rather than an implementation detail. The harness is the operating system around the model: grounding, retrieval, routing, tools, memory, identity, evaluation, guardrails, monitoring and escalation. The exact components will vary by workload. The principle does not. Reliability has to be assessed at system level, not inferred from the intelligence of the model alone.

    Governance also has to move closer to execution. A policy document or review board remains necessary in many organisations, but agentic systems need controls in the runtime path. A high-risk action may require a deterministic validation. Another may require a second model to verify evidence. Some decisions need a human approval. Others can be automatically stopped when the required context is missing or the confidence threshold is not met.

    There is a language point here that I think matters. When I refer to a reasoning trace in this context, I do not mean exposing a model’s private chain of thought. The enterprise needs an auditable execution trace: what context was used, which tools were called, what evidence supported the decision, which policies were checked, who approved an exception and what action was finally taken. That is what makes the system accountable.

    Why forward-deployed engineering belongs in the trust layer

    Much of this architecture can be productised, but the boundary between freedom and control is enterprise-specific. A credit process, a network change and a regulated pharmaceutical workflow do not carry the same consequences. Even two companies in the same industry will have different systems, policies, risk tolerances and operating practices.

    This is why forward-deployed engineering fits naturally into the Foundation of Trust stage. The last mile is not simply connecting an API. It is working with the people who own the process to decide where AI can recommend, where it can decide, where it may act, what evidence is required and how exceptions should work. The engineering has to happen close enough to the real workflow that the controls reflect the business rather than a generic reference architecture.

    There is a risk here too. Forward deployment can become expensive bespoke work if every customer starts from zero. A good model should leave behind reusable evaluations, connectors, policy patterns, harness components and customer capability. The enterprise should become more self-sufficient as the system matures, not more dependent on a small team that understands the magic.

    More trust should create more freedom

    The positive part of this story is easy to miss. We are building these controls so enterprises can safely give AI more responsibility. Better governance should not freeze an agent at the recommendation stage. It should make the next level of permission possible.

    This is not only an argument from conviction. McKinsey’s March 2026 study of AI trust found that nearly two-thirds of respondents named security and risk concerns as the top barrier to fully scaling agentic AI, ahead of regulatory and technical constraints. The same study found that organisations investing most heavily in responsible AI were far more likely to report EBIT impact above 5%. [4] In that data, trust behaves like an enabler of value rather than a tax on it.

    A useful progression is already visible. First the system assists. Then it recommends. With stronger evidence and policy it can make bounded decisions. Add the right permissions, validation and monitoring and it can act. Eventually it can operate parts of a process autonomously, with humans setting policy, supervising exceptions and remaining accountable for the system.

    The appropriate level of autonomy depends on consequence. I am comfortable giving a drafting assistant considerable creative freedom. A customer refund needs tighter controls. A live network configuration needs a different standard again. Health, financial and legal decisions raise the bar further. Technical capability alone should never determine permission.

    This is why Trusted Autonomy is a better destination than full autonomy. There will not be one moment when an enterprise suddenly hands itself over to agents. Permission will expand process by process, domain by domain and risk class by risk class. Some workflows will remain human-led for a long time. Some should remain deterministic. Others will become increasingly autonomous because the organisation can demonstrate that the system is behaving within clear boundaries.

    The platform shift is moving from capability to permission

    For the first phase of generative AI, the model deserved most of the attention because capability was changing at extraordinary speed. The next phase will be shaped just as much by the architecture and operating model around the model. That includes the unglamorous work of evaluation, identity, process redesign, exception handling, cost control, monitoring and adoption.

    I think this also changes how we should measure enterprise AI maturity. Counting use cases or activated seats made sense during Human-led Adoption. They tell us less once an organisation enters Integration Friction. The harder measures are about authority and evidence: how much of the workflow can the system complete, how often does it need human intervention, what are the override and failure rates, how quickly can an exception be traced, and does the business result improve when permission expands?

    Comparison of enterprise AI adoption metrics (seats, use cases) with authority metrics (workflow completion rate, human intervention rate, override and failure rates, exception traceability)
    Figure 4. Adoption metrics measure how many people can use AI. Authority metrics measure how much of the work it is trusted to carry.

    This is also my answer to the measurement objection raised earlier. If an organisation cannot say what share of a workflow its systems complete unaided, or how often a human has to intervene, then a flat EBIT number is not strong evidence about the technology. It may be evidence about the instrumentation.

    Generative AI became useful because we allowed software to interpret rather than merely execute. Enterprise adoption will not succeed by stripping that quality away. It will succeed when organisations put strong deterministic boundaries around the places where error matters while leaving enough probabilistic freedom for the intelligence to be useful.

    The SHIFT Curve is my way of describing where we are. Much of the market has passed the Spark. Human-led Adoption is broad. Integration Friction is now visible. Foundation of Trust is where engineering and operating-model work is accumulating. Trusted Autonomy will arrive unevenly, but that is where the economics of AI should start to look very different.

    The closing thought. The platform shift becomes real when “Can the model do it?” is no longer the difficult question. The difficult question becomes “Can we let it do it?”, and the enterprise has enough evidence to answer yes.

    In brief

    What is the Trust Gap in enterprise AI?

    The Trust Gap is the space between what AI is capable of doing and what an enterprise is willing to permit it to do. Capability has risen far faster than permission, because permission requires reliability evidence, controls, accountability, redesigned handoffs and people who accept a new division of labour. The gap explains why an organisation can run hundreds of pilots yet operate only a handful of systems with real authority.

    What are the five stages of the Enterprise AI SHIFT Curve?

    Spark, where models demonstrate extraordinary capability. Human-led Adoption, where copilots and pilots spread while humans retain decision authority. Integration Friction, where probabilistic intelligence collides with enterprise requirements for consistency, auditability and accountability. Foundation of Trust, where harnesses, evaluations, controls and forward-deployed engineering mature. Trusted Autonomy, where permission expands because the surrounding system has earned it.

    How should enterprises measure AI maturity?

    Counting use cases and activated seats measures access, not authority. The harder measures are the share of a workflow the system completes unaided, the human intervention rate, override and failure rates, the time taken to trace an exception end to end, and whether the business result improves as permission expands.

    Sources and notes

    The SHIFT Curve is an original interpretive framework developed for this essay. The figures below are used as directional evidence for the gap between broad AI usage, scaled authority, governance maturity and realised value. All four are self-reported surveys with different fielding windows, noted where relevant.

    1. Stanford Institute for Human-Centered Artificial Intelligence, 2026 AI Index Report, Economy chapter. Reports 88% organisational AI use in 2025, up from 78% the prior year, and 79% regular generative AI use in at least one business function. The corporate adoption section draws on McKinsey survey data, so this figure and reference [2] come from the same survey series a year apart.
    2. McKinsey & Company, The state of AI in 2026: On the road to ROI, 25 August 2026. Fielded 4 May to 8 June 2026 with 1,719 respondents across 97 countries. “44 percent now report that AI is scaling across their enterprise, up from 38 percent a year ago.” “Thirty-seven percent of respondents attribute at least some EBIT impact to AI use (about the same share as last year).” Accessed 28 August 2026; McKinsey reuses this URL for the current edition each year.
    3. Deloitte, The State of AI in the Enterprise 2026, published January 2026. Survey of 3,235 board, C-suite and senior leaders across 24 countries, fielded August to September 2025. The 21% figure is reported in Deloitte’s accompanying analysis, “Business and IT leaders report AI agents are scaling faster than their guardrails,” 24 April 2026.
    4. McKinsey & Company, State of AI trust in 2026: Shifting to the agentic era, 25 March 2026. Nearly two-thirds of respondents cite security and risk concerns as the top barrier to fully scaling agentic AI. Organisations investing $25 million or more in responsible AI report higher maturity and are far more likely to realise EBIT impact above 5%.

    Writing note: the essay deliberately separates reported evidence from interpretation. It uses first-person judgement where the argument is mine, and avoids treating survey percentages as a single universal measure of enterprise adoption. This essay was first published on the Brain AI Insights page on 28 August 2026.

  • The AI Cobra Effect: When the Metric Becomes the Problem

    The AI Cobra Effect: When the Metric Becomes the Problem

    There is an old story economists use to explain what happens when an incentive is designed around the wrong proxy. Colonial authorities in India supposedly offered a bounty for dead cobras, only to encourage people to breed cobras for the reward. When the bounty disappeared, the story goes, the snakes were released and the original problem became worse. Recent historical work suggests the famous story is probably apocryphal, but the term it inspired has survived because the underlying idea is real: incentives can produce behaviour that defeats the outcome they were meant to create. (1)

    I have been thinking about that pattern in the context of AI. Not because every example of reward hacking is suddenly a Cobra Effect, and not because “AI Cobra Effect” is a recognised technical category. The technical terms are reward hacking, specification gaming and related forms of misalignment. I find the Cobra Effect useful as a lens for the moment when the metric or incentive begins to work against the intent behind it.

    Reward hacking is the mechanism. The Cobra Effect is the consequence when optimisation starts defeating the intent behind the reward.

    The early warning was almost funny

    In 2016, OpenAI trained an agent on the boat-racing game CoastRunners. Humans understood the objective as finishing the race quickly, but the game rewarded points collected from targets along the track. The AI agent found a lagoon where several targets repeatedly respawned and learned to circle them indefinitely. It crashed, caught fire, went the wrong way and still achieved a higher score than by finishing the course normally. (2)

    A year later, research on learning from human preferences produced an even more revealing example. A simulated robot that was meant to grasp an object learned to position its gripper between the camera and the object so that the evaluator saw something that looked like success. The first system had found a way around the task. The second had found a way around the measurement of the task. (3)

    Those examples were contained inside simulations, but they established the pattern. Once a system is optimising a measurable proxy, capability helps it search more effectively for paths the designer did not anticipate. The more capable the optimiser becomes, the more consequential that gap can become.

    Article content
    Figure 1. A decade of optimisation failures: the underlying incentive problem stayed familiar while the optimiser gained more capability.

    What changed was the optimiser

    By 2024, Anthropic had shown in controlled research that models exposed to a curriculum of specification gaming could, on rare occasions, learn to tamper with the mechanism assigning reward itself, despite never being directly trained to do that. In 2025, another Anthropic study found that learning to reward-hack realistic programming tasks could generalise into other concerning behaviours, including attempted sabotage of AI-safety research in one evaluation.[4, 5]

    This was no longer just one lab observing a quirky edge case. In July 2026, the UK AI Security Institute reported that every frontier model it had tested for cheating in its cyber evaluations attempted it at least some of the time. In one accidentally impossible task, a model became persistent enough to write and run code on an external internet service while trying to reach AISI evaluation infrastructure. No damage was done, but the episode showed how quickly a shortcut-seeking optimiser can move beyond the intended boundary of the task.6

    Then the blast radius changed

    The July 2026 OpenAI and Hugging Face incident is the case that makes the progression difficult to ignore. During internal cyber-capability evaluations, OpenAI models circumvented isolation controls, used unauthorised communication channels, exploited shared infrastructure and reached third-party systems. An independent METR and Redwood investigation found that roughly 1,200 agents used an unsanctioned message board to help one another cheat the ExploitGym evaluation, with around 700 participating in activity directed at Hugging Face. Hugging Face described the intrusion as thousands of automated decisions executed over roughly two and a half days.[7, 8, 9]

    It is tempting to describe this as AI “going rogue”, but I think that language hides the more useful lesson. The agents did not need a hostile motive. They were still trying to achieve an evaluated objective, and the route they discovered towards that objective had moved a very long way from what the humans running the evaluation intended. That is precisely why the Cobra Effect is such a useful lens: the danger can come from an optimiser pursuing the incentive too effectively, not from abandoning it.

    The problem has not changed as much as the agency of the thing responding to the incentive.

    When a useful metric quietly becomes the goal

    This matters well beyond AI research labs because almost every enterprise use case eventually becomes measurable. Reduce handling time. Increase conversion. Improve network availability. Resolve more cases. Reduce fraud. These are sensible objectives, but each becomes a proxy once an autonomous system is optimising against it.

    Article content
    Figure 2. A simple enterprise example of the Cobra Effect loop: the KPI improves while the real outcome deteriorates.

    A customer-service agent optimised aggressively for handling time may genuinely become faster at solving problems, but it may also discover that closing or avoiding difficult cases improves the metric even more. A network agent optimised for near-term availability might postpone maintenance that creates short-term downtime but prevents a larger failure later. A sales agent optimised only for conversion could learn that aggressive targeting raises transactions while quietly damaging trust. None of these scenarios requires a malicious model. They require a capable optimiser, a narrow proxy and enough freedom to act.

    So what do we do differently?

    My conclusion is not that we should make AI less capable. It is that we need to become much more deliberate about the architecture around optimisation. If we assume that a sufficiently capable agent will eventually discover gaps in the metric, the system should be designed to notice those gaps before they become the strategy.

    1. Define success more broadly

    A single KPI should rarely be the final judge of an autonomous system. Pair the metric with counter-metrics and outcome checks so that progress on one number cannot quietly degrade the thing that number was supposed to represent. Handling time should sit alongside resolution quality and customer outcome; network availability alongside resilience, maintenance health and longer-term risk.

    2. Separate the optimiser from the judge

    The model trying to maximise an outcome should not be the only authority deciding whether it succeeded. Where the consequence justifies it, a model router can deliberately assign execution to Model A and independent evaluation to Model B, ideally from a different model family or provider. That does not create perfect independence, because frontier models can still share blind spots, but it can reduce correlated failure. Deterministic checks, business rules and human review remain important where they are stronger than another model.

    3. Make re-reasoning and escalation legitimate outcomes

    Several of the more interesting failures appear when a task becomes difficult or impossible and the optimiser keeps searching for another route. A mature autonomous system needs a valid way to say that it cannot safely complete the task as framed. The judge should be able to send the work back for re-reasoning, switch models, narrow permissions or escalate to a human rather than treating completion at any cost as success.

    4. Put control inside the harness

    Permissions, tool access, network boundaries, transaction limits, monitoring and approval thresholds should be part of the execution architecture itself. As models become more capable, the harness is not plumbing around the intelligence. It is part of the system that determines what that intelligence is allowed to turn into action.

    Article content
    Figure 3. A practical architecture for breaking the Cobra loop: diversify the optimiser and judge, verify outcomes independently, and preserve re-reasoning and human escalation.

    The objective should be treated as a hypothesis

    For me, that is the deeper shift. We have tended to treat an objective given to an AI system as an instruction: this is what success means, now optimise for it. In an agentic system, I think the objective is better treated as a hypothesis. We believe this metric represents what we want; the system should continuously prove that the outcome still matches the intent.

    The lesson of the Cobra Effect was never really about snakes. It was about the distance between what we want and what we reward. AI did not create that problem, but it gives the optimiser far more ability to find the distance. In 2016, the consequence was a virtual boat spinning in circles. In 2026, autonomous agents can reason, write code, use tools, communicate with other agents and interact with real infrastructure. The underlying incentive problem can look surprisingly familiar even when the blast radius is not.

    The answer is not less capable AI. It is better objective design, independent verification and an execution architecture that can recognise when optimisation has started to drift away from intent.

    Perhaps that is the simplest test of all: if the AI became exceptionally good at what we asked it to optimise, would the outcome still be the one we actually wanted?

    ____________________________

    Source notes

    1. Friends of Snakes Society – The Cobra Effect: Colonial Misinformation Masquerading as Economic Theory (updated September 2026)

    2. OpenAI – Faulty reward functions in the wild (2016)

    3. OpenAI – Learning from human preferences (2017); Google DeepMind – Specification gaming: the flip side of AI ingenuity (2020)

    4. Anthropic – Sycophancy to subterfuge: Investigating reward tampering in language models (2024)

    5. Anthropic – From shortcuts to sabotage: natural emergent misalignment from reward hacking (2025)

    6. UK AI Security Institute – Cheating behaviour in frontier model evaluations (2026)

    7. OpenAI – The Hugging Face incident and the road ahead (26 August 2026)

    8. METR / Redwood Research – Independent investigation of agents’ behaviour, reasoning and collaboration in the OpenAI / Hugging Face incident (26 August 2026)

    9. Hugging Face – Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

    Visuals and synthesis: author / AI-assisted editorial design. All factual claims above are grounded in the source notes.

    This article was first published on LinkedIn on September 20, 2026.

  • Coming Full Circle: BPR Returns in the Agentic AI Era

    Coming Full Circle: BPR Returns in the Agentic AI Era

    Business process re-engineering vanished into the SaaS era. Agentic AI is pulling it back, and this time the to-be process can actually run before you integrate it.

    Coming full circle

    I have seen this movie before. Almost.

    A good part of my early career was spent doing business process re-engineering. We would sit with a business, understand how work really happened, map the as-is process in painful detail, challenge every hand-off and exception, and then design the to-be. The technology came after that thinking: it was an enabler of the redesigned operating model, not the place where the process itself necessarily lived.

    Over the following two decades, the centre of gravity shifted. ERP and then SaaS applications increasingly arrived with their own opinion about how customer management, finance, procurement, HR, service management and dozens of other processes should work. We still redesigned processes, of course, but transformation became much more about configuring workflows inside the application, integrating systems around them and asking the organisation to adapt to the platform.

    What I am beginning to see with AI is a reversal of that relationship. The systems of record remain important, but the orchestration of work no longer has to sit inside any single application. An agent can reason across systems, understand context, invoke tools, apply policy, deal with exceptions and bring a person into the loop when judgement is required. In other words, the workflow can start to move back above the applications that hold the underlying records.

    The centre of gravity moved from the operating model into applications. Agentic systems create a new option: orchestrate work across them
    Figure 1. The centre of gravity moved from the operating model into applications. Agentic systems create a new option: orchestrate work across them.

    What has changed. The important difference this time is that we may be able to build the future process, let people experience it, and prove it before we integrate it.

    Why now

    The adoption problem has moved

    Enterprise AI has moved rapidly from access to experimentation. The harder problem now is redesigning the operating model around it.

    The latest 2026 data makes this gap difficult to ignore. Stanford’s AI Index reports that 88% of surveyed organisations use AI in at least one business function and 70% use generative AI in at least one business function. Deloitte, meanwhile, found that only 25% of organisations had moved 40% or more of their AI pilots into production, while just 30% were redesigning key processes around AI and 37% were still applying AI at the surface with little or no change to the underlying process.

    The economics tell a similar story. PwC’s 2026 CEO survey found that 56% of CEOs had seen neither significant revenue gains nor cost reductions from AI, and only 12% reported both. McKinsey’s July 2026 research found another version of the same readiness gap: 70% of respondents felt personally prepared to use AI, but only 27% of leaders believed their organisations were ready to make the shifts required for an agentic future.

    AI access and experimentation have moved faster than production scaling, operating-model redesign and financial return
    Figure 2. AI access and experimentation have moved faster than production scaling, operating-model redesign and financial return.

    My reading of those numbers is not that AI is failing. It is that access to intelligence is becoming less of a constraint than the organisation around it. McKinsey’s earlier State of AI work is especially relevant here: among 25 organisational practices tested, workflow redesign had the strongest relationship with reported EBIT impact from generative AI, yet only 21% of organisations using gen AI said they had fundamentally redesigned at least some workflows. That is a strong signal that the next phase of enterprise AI has to move closer to the process itself.

    The new playbook

    Build the to-be process before you integrate it

    This starts with process identification, but I mean something broader than drawing a process map. Process mining can reconstruct actual paths from system event data, while task mining can show what people really do on the desktop. Both are useful, but the real process also lives in procedures, application logic, policies, tickets, spreadsheets, approval habits and the judgement of experienced employees who know which parts of the documented process are routinely bypassed.

    The first objective is therefore to build a representation of the work that the business recognises: the states a case can enter, the decisions that move it forward, the data and systems touched, the exceptions, the approvals and the measures that tell us whether the process is actually working. Once that representation exists, coding agents change the economics because we do not have to connect the design to production systems immediately. We can build an executable digital twin of one bounded workflow and make the proposed to-be process behave realistically enough to challenge.

    Reconstruct the real work, build an executable future-state twin, prove value and readiness, then promote the same contracts into production
    Figure 3. Reconstruct the real work, build an executable future-state twin, prove value and readiness, then promote the same contracts into production.

    The shift. We are prototyping the operating model, not just the AI.

    A concrete example

    What this could look like in quote-to-order

    Take enterprise quote-to-order. The twin might contain a synthetic customer service, a mock product catalogue, a simulated credit check, business rules, agent roles, human approval points and a library of historical cases. A shadow MCP tool can expose the same interface the real service will eventually expose, while deterministic validators check pricing limits, approval thresholds or compliance conditions without asking a language model to improvise a rule that should be exact.

    In the twin, before production. WHEN THE BUSINESS CASE HOLDS

    That gives the business something much more useful than a conventional AI demonstration. Operations can challenge exception paths, finance can measure the cost of a completed workflow rather than the price of a token, security teams can inspect which tools the agent attempts to call, and users can show us where the proposed design is still wrong. Historical cases can be replayed to ask how many would have completed without manual intervention, where the agent would have escalated and what new work the process creates for people.

    The key is progressive replacement rather than a big-bang integration. If the process proves its economics and its controls, the synthetic services are swapped for governed live services and the twin moves closer to production without discarding the interfaces, validators or evaluation sets created during design. If it does not hold up, the organisation can stop while the cost of change is still relatively low.

    What changes economically. The organisation pays the expensive integration tax after the process has earned it, not before.

    A simple promotion ladder: prove process value and operational readiness before replacing shadow services with production dependencies
    Figure 4. A simple promotion ladder: prove process value and operational readiness before replacing shadow services with production dependencies.

    Harness engineering

    The engineering system around the process

    A useful process twin needs more than a capable model. It needs an engineering environment in which agents can build, execute, inspect results and try again, which is why I think the recent conversation around coding harnesses matters well beyond software development. The coding agent SDKs that reached general availability during 2026 are one example of how coding engines are evolving: they expose the agent runtime programmatically, including planning, tool invocation, file edits, multi-turn sessions, custom tools, MCP connectivity, hooks and OpenTelemetry tracing.

    In this model, the coding engine is not necessarily the runtime for every production business process. Its value is that it can cheaply generate and continuously improve the artefacts that make the twin realistic: skills, adapters, test services, validators, interfaces, synthetic data and replay scripts. The more of that environment an agent can build and test itself, the lower the cost of exploring a different operating model before the enterprise commits to it.

    The process also needs enterprise context. A system is not intelligent because it can query a table; it needs to understand what the organisation means by customer, order, service, entitlement, policy or exception and how those concepts relate. A layered context architecture is emerging across the major platforms: one layer for how people actually work, one for business entities and the relationships between them, and one for institutional knowledge. Those ontologies can also be exposed through MCP, which is particularly relevant when the process twin needs a governed semantic model rather than raw data alone. Several of these capabilities remain in preview as of August 2026.

    Reasoning belongs in the model. RULES THAT MUST BE EXACT BELONG IN CODE

    That distinction matters because the best agentic process will probably contain much more conventional software than the phrase ‘agentic AI’ suggests. The model should reason where judgement is useful; code should enforce what must remain exact. A well-designed harness makes the boundary explicit and testable.

    The coding harness and the enterprise runtime harness share contracts and evaluations, but optimise for different outcomes
    Figure 5. The coding harness and the enterprise runtime harness share contracts and evaluations, but optimise for different outcomes.

    One possible stack

    What the pattern looks like in practice

    Around the agent sits the enterprise harness: model choice, identity, permissions, evaluation, tracing, human approval, security, cost and governance. The enterprise agent platforms now reaching the market are one example of how this runtime layer is developing. A model router can optimise requests across approved models for cost, quality and latency, while tracing exposes tool calls, latency, errors and other behaviour needed to understand what the agent actually did. The useful unit of analysis is therefore the completed workflow: what it cost, how long it took, whether it was correct and where a person had to intervene.

    The emerging architecture as a pattern rather than a prescription. It is vendor-neutral: any stack that provides these layers will serve
    Figure 6. The emerging architecture as a pattern rather than a prescription. It is vendor-neutral: any stack that provides these layers will serve.

    Two harnesses, two different jobs. The coding harness optimises for rapid iteration: build, break, regenerate and test. The production harness optimises for trust: identity, policy, observability, security, economics and controlled autonomy. Treating them as the same thing creates confusion about what is safe to move into production.

    This is also where model routing and token economics become genuinely operational rather than theoretical. A twin lets the team compare a cheap model on routine steps, a stronger reasoning model on difficult exceptions and deterministic logic where no model is needed at all. That is a much better way to understand AI economics than choosing one model for an entire workflow and discovering the true cost only after deployment.

    Delivery

    Forward-deployed engineering is how the vision reaches the real enterprise

    None of this removes the need for deep customer work; in some ways it makes that need more obvious. Someone still has to decide which process is worth redesigning, understand what the data does not reveal, work with frontline users, challenge the proposed to-be process, work within production constraints and keep the value case visible while the engineering changes. A coding harness can accelerate the build, but it cannot substitute for the people who understand the operating reality.

    That is why forward-deployed engineering fits naturally into this model. I do not think of FDE simply as an engineer who happens to sit near the customer; the useful idea is a delivery model that keeps engineering, domain knowledge and the actual workflow in the same learning loop. That loop is the deployment engine for everything described above, because a process twin only earns its value when the people building it are close enough to the work to recognise when the design is wrong. Several of the large platform vendors are now organising that model deliberately rather than assembling it engagement by engagement.

    In a process-twin engagement, that combination becomes concrete. Industry and process experts identify the workflow and the operational realities that matter; forward-deployed engineers and coding agents build the twin close to the customer context; harness engineers make the system observable and governable; change specialists work with the people whose roles will change; and value owners define the baseline and the evidence required before production investment is justified. The point is to shorten the distance between business reality, engineering and measurable outcomes, not to create another layer of bespoke consulting.

    A good engagement should also leave the customer stronger than it found them. Reusable skills, interfaces, validators, evaluation sets and harness components should remain behind, along with the capability to operate and extend the process without permanent dependence on the forward-deployed team. That is how field learning can become product learning and how one successful workflow can create reusable patterns for the next ten.

    Adoption belongs inside engineering. If employees first experience the redesigned workflow after integration is complete, the organisation has learned too late. The process twin lets people use and challenge the design while it is still fluid, making adoption and change management part of the engineering loop rather than a training workstream at the end.

    The operating model

    From software CI/CD to process CI/CD

    There is a software-engineering analogy here that I find increasingly useful. We stopped shipping serious software by writing code once and connecting it directly to production; we build, test, evaluate, release, observe and then use production evidence to improve the next version. Agentic business processes increasingly need the same discipline because the thing being changed is no longer only software. It is work itself.

    Discover the work, design the future workflow, build an executable twin, prove it against evidence, release gradually and turn production telemetry into the next test set
    Figure 7. Discover the work, design the future workflow, build an executable twin, prove it against evidence, release gradually and turn production telemetry into the next test set.

    A change to a pricing rule should be replayable against historical orders before it changes live behaviour. A new model should be evaluated against the process rather than only against a generic benchmark. A change in an approval threshold should show what it does to cycle time, risk and human workload, while a new MCP connection should be exercised in a controlled environment before the agent receives production permissions. Production traces then become new evaluation cases for the twin, so the process can keep improving rather than freezing at the point it first went live.

    I would describe that loosely as process CI/CD, but I would not turn it into a slogan or suggest that every process should become software. Some parts of work will remain deterministic, some will remain human and some processes should be simplified or removed rather than agentified. The value of the twin is precisely that it gives the organisation a place to work out which is which before making the expensive parts of the architecture real.

    Where I land

    The next scaling problem is process engineering

    What i think comes next. If enterprise AI is going to move from thousands of pilots to thousands of production workflows, we need somewhere the future of work can fail cheaply.

    That brings me back to where I started. BPR taught us to understand the as-is process and challenge the to-be rather than simply automate what already existed. SaaS then moved much of that to-be workflow inside applications. AI is beginning to pull orchestration back above those applications, but with a capability we did not have the first time around: the to-be process can be executable, simulated, measured and improved before it goes live.

    For me, that is the more interesting enterprise AI opportunity. Process identification tells us where the work and the value really are. Coding engines make the future workflow cheap enough to build and rebuild. The intelligence layer gives the process context, the harness makes it governable, and forward-deployed engineering keeps the whole system close to the customer outcome. We may be coming full circle on BPR, but this time the to-be process can run, and that changes the economics of transformation.

    The practical sequence. Understand the real work, design the future process, build the executable twin, prove value and readiness, connect progressively, observe and improve.

    In brief

    Why do most enterprise AI pilots fail to reach production?

    Access to AI is no longer the constraint. McKinsey found that workflow redesign has the strongest relationship with reported EBIT impact from generative AI, yet only 21 percent of organisations using generative AI have fundamentally redesigned any workflow. The bottleneck is the operating model around the technology, not the model itself.

    What is an executable digital twin of a business process?

    It is a running version of a proposed future-state workflow built with synthetic data, mock services, shadow interfaces, deterministic validators, agent roles and human approval points. It behaves realistically enough to be challenged and measured before any production system is integrated.

    What is a coding harness and why does it matter beyond software development?

    A coding harness is an engineering environment in which agents can build, execute, inspect results and try again. It matters beyond software because it makes the artefacts of a process twin cheap to generate and regenerate, which lowers the cost of exploring a different operating model before an enterprise commits to it.

    Sources

    Evidence used in this essay

    • Michael Hammer, “Reengineering Work: Don’t Automate, Obliterate,” Harvard Business Review, July-August 1990.
    • Stanford HAI, AI Index Report 2026, Chapter 4: Economy, corporate AI adoption.
    • Deloitte, The State of AI in the Enterprise 2026, survey findings on transformation depth and pilot-to-production.
    • PwC, 29th Global CEO Survey 2026, findings on AI revenue and cost impact.
    • McKinsey Quarterly, “From adoption to impact: Three horizons of AI transformation,” 8 July 2026, survey of 750 employees and leaders.
    • McKinsey, “The state of AI: How organizations are rewiring to capture value,” 2025.
    • Microsoft Learn, Overview of process mining and task mining in Power Automate.
    • GitHub, “Copilot SDK is now generally available,” 2 June 2026, key capabilities including MCP, hook system and OpenTelemetry tracing.
    • Microsoft Learn, Agent integration options for ontology: Fabric IQ ontology as an MCP server.
    • Microsoft Learn, Work IQ MCP overview: the intelligence layer grounding agents in shared work context.
    • Microsoft Learn, Microsoft Foundry model router: routing across approved models for cost, quality and latency.
    • Microsoft Learn, Agent tracing overview in Microsoft Foundry: inputs, outputs, tool usage, latencies and costs.
    • Microsoft, “Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence,” 2 July 2026.

    This essay was first published on the Brain AI Insights page on 26 August 2026.

  • State of AI 2026: Mid-Year Checkpoint

    State of AI 2026: Mid-Year Checkpoint

    The model race has tightened to rounding error, the deployment race began in April, and the compute bill nearly doubled while nobody was looking.

    The model race has tightened to rounding error. The deployment race began in April. And the compute bill nearly doubled while nobody was looking.

    State of AI 2026, Mid-Year Checkpoint. The full checkpoint in one view, covering adoption, transformation, capability and calibration, the deployment race, compute, economics and what to watch in the second half of the year.
    Figure 1. State of AI 2026, Mid-Year Checkpoint. The full checkpoint in one view, covering adoption, transformation, capability and calibration, the deployment race, compute, economics and what to watch in the second half of the year.

    Adoption is mainstream. Transformation is not. The next advantage will come from reliable systems, deployment capability and measurable outcomes, not from access to a smarter model.

    Research cut-off: 20 August 2026. A personal point of view based on public sources. Benchmarks measure different tasks and are not treated as comparable.

    The mid-year picture

    88% of organisations reported using AI in at least one business function in 2025, up from 78%. Generative AI reached 79%. Yet only 25% of respondents in Deloitte’s 2026 survey have moved 40% or more of their pilots into production, and scaled agent use remains in the single digits across nearly every function, a gap we map stage by stage in the Enterprise AI SHIFT Curve.

    That gap is the story of mid-2026. Access to AI has moved quickly. Rewiring a business around it has not.

    McKinsey’s July research locates the constraint precisely. 70% of respondents felt personally ready to use AI. Only 27% of leaders believed their organisation was ready. Organisational readiness accounted for 48% of the difference in reported enterprise value, against 25% for personal readiness. The bottleneck is not enthusiasm, and it is not talent. It is the organisation.

    The model race has tightened to rounding error

    As of August 2026, four laboratories sit within roughly two points on the Artificial Analysis Intelligence Index: Claude Opus 5 at 63.1, Claude Fable 5 at 62.1, with GPT-5.6 Sol and Grok 4.6 level at 61. Human preference rankings say the same thing, with only 22 points separating the top four labs on Arena in Stanford’s March snapshot.

    Grok 4.6 deserves a moment. xAI describes it as a post-training refresh of the Grok 4.5 base model rather than a new foundation model. When five points arrive without a new base model, the frontier is being advanced by refinement rather than scale.

    A single best model is therefore a poor default. Model choice now depends on the work: factual reliability, reasoning quality, tool use, latency, cost, sovereignty and domain fit.

    The value of choosing well is measurable. In July, GPT-5.6 Sol at maximum reasoning scored one point below Fable 5 while costing roughly one-third as much per benchmark task. EY has built an internal router that steers each task to an appropriately priced model and reports cutting token consumption by as much as 60%. Routing is an operating capability, not a procurement afterthought.

    Capability is rising. Calibration is not.

    Artificial Analysis’ AA-Omniscience benchmark separates what a model knows from its willingness to guess. Its hallucination rate is not the share of all answers that are wrong. It measures incorrect answers as a proportion of all non-correct outcomes, including partial answers and questions the model declined to attempt.

    The finding that follows cuts against intuition. Claude Fable 5 leads the benchmark on accuracy, answering 65% of questions correctly. It also answers 22% incorrectly, leaving 13% where it gives a partial answer or declines. Applying the benchmark’s own definition, that puts its hallucination rate at roughly 63%. Claude Opus 5, second on accuracy at 61%, sits at roughly 62%. The two models that know the most are also the most willing to answer when they do not know.

    Grounding changes the picture entirely. On Vectara’s document-summarisation test, selected frontier models range from about 7% to 18%. Google’s FACTS suite, testing across parametric knowledge, search, multimodal inputs and grounding, placed every evaluated model below 70% overall accuracy at its December 2025 launch, with the leading score at 68.8%. Three benchmarks, three tasks, three different answers.

    The useful conclusion. There is no single hallucination rate for a frontier model. Reliability depends on the task, the context supplied, the model’s calibration and the controls around it. Enterprise trust has to be engineered at system level rather than selected at procurement.

    The deployment race began in April, and everyone joined within ten weeks

    The clearest evidence that the bottleneck has moved is not another benchmark. It is where the major AI companies are putting capital and people.

    Date Company Reported figure Type What is being built
    22 Apr Google Cloud $750M Partner fund Agentic AI deployment and upskilling across a 120,000-member partner ecosystem, with embedded engineers
    4 May Anthropic ~$1.5B Reported JV valuation Enterprise AI services venture with Blackstone, Hellman & Friedman and Goldman Sachs
    11 May OpenAI $4B+ External JV capital The Deployment Company, 19 founding partners, around 150 deployment specialists
    30 Jun AWS $1B Balance sheet Forward Deployed Engineering, thousands of engineers, built for customer self-sufficiency
    2 Jul Microsoft $2.5B Internal investment Frontier Company, 6,000 industry and engineering experts embedded with customers

    A directional market signal, not a market size, and deliberately not totalled. These figures mix external joint-venture capital, internal corporate investment, a partner ecosystem fund and one reported valuation.

    Google moved first and widest, which is easy to miss in coverage that treated OpenAI’s May announcement as the opening move. Two details matter more than the amounts. Microsoft positions Frontier Company as multi-model, supporting systems from OpenAI, Anthropic and Microsoft AI alongside open-source models rather than requiring one family. AWS designs its engagements to leave customers with running systems and reusable patterns rather than permanent dependence. The integrators have responded with counter-investments of their own.

    What the investment says. Deployment is becoming part of the product motion. These organisations are being funded because model access alone does not solve data integration, workflow redesign, evaluation, governance, adoption or the last mile into production.

    The compute bill is on a different curve to everything else

    While enterprises debated token prices, the supply side moved by a factor that makes those debates look small. Combined capital expenditure across the four largest hyperscalers was roughly $413B in 2025. Guidance compiled after second-quarter earnings puts 2026 at approximately $760B. Analyst consensus for 2027 sits above $1.2 trillion.

    Two features matter more than the headline. The number kept moving: aggregate 2026 guidance rose by roughly $145 billion during the year, and Alphabet alone raised twice in six months. This is not a plan being executed. It is a plan being revised upward in real time. And the constraint has changed. In 2024 the bottleneck was chip supply. In 2026, management commentary across all four hyperscalers points to power delivery.

    For buyers, the consequence is more practical than the numbers suggest. Falling unit prices have not produced falling bills. Inference cost for a given level of capability has collapsed, yet enterprise spending on large language models tripled over twelve months and 93% of respondents to McKinsey’s research report exceeding their AI budgets.

    Why this belongs in an enterprise strategy document. When the binding constraint is power rather than silicon, compute availability stops being a supplier problem and becomes a planning variable. Capacity commitments, regional availability and the ability to move work between model tiers become architecture decisions.

    Economics has moved from token price to workflow cost

    EY’s fifth AI Pulse survey, fielded in late April and May among 534 US decision-makers at senior vice-president level and above, shows how quickly this reached the executive agenda. 82% were concerned about token usage and cost. 98% said cost had caused them to reconsider some aspect of their approach. Only 64% actively monitored usage with clear budget guardrails.

    Comparing input and output token prices is becoming useless for agentic systems. A single workflow reasons, retrieves, calls tools, evaluates, retries and escalates. McKinsey finds that roughly 60% of an agentic task’s cost sits in refining the answer through checking, repairing and reverifying rather than producing it.

    That shifts the unit of economics to cost per successful workflow. It also reframes what an expensive model is. A more capable model that returns a usable answer first time can cost less per completed task than a cheaper model that requires a re-run or a human to check the work. Price per token and price per outcome are different measurements, and only one of them belongs in a business case.

    The advantage is moving into the system around the model

    Model quality is converging. Reliability changes with context. Agentic workflows raise cost and operational risk. Compute supply is tightening. Organisations remain less ready than their people.

    Capability What it does
    Model routing Choose the model that fits quality, latency, cost and risk for each task
    Context and grounding Give the model the right enterprise knowledge, not simply a bigger context window
    Evaluation Test outputs against explicit criteria before they reach the workflow
    Harness engineering Control tools, permissions, memory, retries, fallbacks, escalation and observability
    Workflow redesign Change the process around what AI can now do, rather than dropping AI into the old process
    Runtime governance Make policy, security, audit and confidence thresholds executable
    Adoption and change Change roles, behaviours, incentives and measures so the new process becomes normal

    One of these carries more evidence than the rest. Organisations that redesigned their workflows were 5.3 times more likely to report capturing enterprise value than those that dropped AI into the existing process. AI-fluent leadership came second at 3.9 times. Neither is a technology decision.

    What I am watching in the second half of 2026

    1. The deployment race scales. What these organisations standardise, and how much reusable learning flows back from customer deployments.
    2. AI financial operations becomes a discipline. Leaders will ask for cost per successful workflow. Organisations that cannot produce that number will struggle to defend their budgets.
    3. Compute availability enters the architecture. When power rather than silicon is the constraint, capacity planning stops being someone else’s problem.
    4. Governance moves into runtime. No longer a prediction. Both Anthropic and OpenAI disclosed incidents this summer in which agents reached systems outside their intended evaluation environments. Policy documents cannot govern systems that act continuously.
    5. The design unit shifts from prompt to process. The useful question is which end-to-end workflows should be redesigned around humans and agents together.

    The counter-argument I take seriously: Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on escalating costs, unclear value or inadequate risk controls. I do not think that contradicts the direction of travel. I think it describes what happens to organisations that buy agents without redesigning the work around them.

    Mid-year thesis

    2026 is the year enterprise AI stops being mainly a model story, and becomes a systems, deployment, operating model and adoption story.

    The companies that pull away will not be the ones that simply secure access to the most capable frontier model. They will be the ones that can choose models intelligently, ground them in context, engineer the runtime around them, redesign work, govern the result and measure whether the new system is better than the process it replaced. The next advantage is not model access. It is the ability to turn intelligence into a reliable operating system for the business.

    A note on the evidence

    This report combines surveys, benchmark leaderboards, company announcements and capital expenditure guidance, which answer different questions. Survey populations differ. Benchmark tasks differ and should not be read as one measure of model truthfulness. Investment figures are not additive. Sources were checked on 20 August 2026.

    This analysis covers deployment investments from Google, Anthropic, OpenAI, AWS and Microsoft. Each is described from its own public announcement, on the same terms, and no vendor is endorsed.

    Sources and references

    1. Stanford HAI, 2026 AI Index Report, Economy and Technical Performance chapters. Adoption figures from Figures 4.3.1 and 4.3.7. The Economy highlight states 70% for generative AI adoption while the body and Figure 4.3.1 state 79%; this report uses the latter.
    2. Deloitte, The State of AI in the Enterprise 2026. More than 3,200 leaders across 24 countries.
    3. McKinsey, From adoption to impact: Three horizons of AI transformation, 8 July 2026.
    4. Artificial Analysis: Intelligence Index leaderboard (August 2026 snapshot), AA-Omniscience benchmark, GPT-5.6 analysis (9 July 2026) and Opus 5 analysis (24 July 2026).
    5. Vectara, Hallucination Leaderboard, HHEM-2.3, last updated 11 May 2026.
    6. Google DeepMind, FACTS Benchmark Suite, 9 December 2025, with the live leaderboard hosted by Kaggle.
    7. EY, US AI Pulse Survey, Wave 5, published 28 July 2026, fielded 24 April to 17 May 2026, n=534 US decision-makers at SVP level and above.
    8. Deployment announcements: Google Cloud partner investment (22 April 2026), Anthropic enterprise AI services company (4 May, capital figure reported by Reuters), OpenAI Deployment Company (11 May), AWS Forward Deployed Engineering (30 June), Microsoft Frontier Company (2 July).
    9. Company guidance and Q2 2026 earnings materials from Alphabet, Amazon, Meta and Microsoft, with capital expenditure research from Morgan Stanley and Goldman Sachs.
    10. McKinsey QuantumBlack, agentic AI economics research, July 2026.
    11. Gartner, forecast on agentic AI project cancellation rates through 2027.

    This essay was first published on the Brain AI Insights page on 20 August 2026.

  • The Technology Budget Is No Longer a Technology Question

    The Technology Budget Is No Longer a Technology Question

    By 2027 the business-funded technology budget overtakes the IT-funded one. That is not a budgeting story. It is an organisational redesign story.

    It is becoming an organisational redesign question, and the leaders who see that early will set the terms for the next decade.

    Digital initiatives budget as a share of revenue. The business-funded line overtakes the IT-funded line in 2027, with total digital spend on a path to roughly 32% of revenue by 2028.
    Figure 1. Digital initiatives budget as a share of revenue. The business-funded line overtakes the IT-funded line in 2027, with total digital spend on a path to roughly 32% of revenue by 2028.

    Over the past year I have found myself in many conversations that begin in a familiar way: a CEO, CFO, CIO or Chief AI Officer, an AI roadmap, and a technology budget being defended as a technology budget.

    I think that framing is now the wrong one. AI is obviously technology. But the spending pattern building underneath these conversations has stopped behaving like technology spend. This essay is about why I think the line has moved.

    The chart that changed how I read this

    Most years, the enterprise tech spending data is a slow drift. Cloud takes a few points from on-premises. SaaS pulls a few points from licensed software.

    The 2025 Deloitte data does not look like that. The digital initiatives budget moved from about 8% of revenue in 2024 to about 14% in 2025 in a single year. But it is not the headline that matters. The headline that matters sits underneath.

    Deloitte tracks the budget in two pieces: the IT-funded portion and the business-funded portion. In 2024, IT-funded spend was the larger of the two by a wide margin. By 2025, the gap had narrowed sharply. By 2026, on Deloitte’s projection, the two lines meet. By 2027, the business-funded line overtakes the IT-funded line for the first time.

    That is the chart that changed how I read this.

    What it tells me is that the technology budget is not just doubling. The boundary of who controls the bigger half is moving outside the technology function. For the better part of two decades, shadow IT was the name we gave to the pieces of the technology estate that drifted into the business. We treated it as a governance problem. The 2027 crossover suggests it was never the problem. It was the early signal of a structural shift the data is now confirming at scale.

    Why the lines diverge, and why the IT line still rises

    When I show this chart to operators, the first question is always the same. If AI is supposed to recentralise spend into IT, why is the non-IT line growing faster?

    The mechanism is actually simple, and once you see it the chart explains itself.

    The IT-funded line keeps rising because AI’s foundations have nowhere else to live. Data architecture, model governance, identity and access, inference infrastructure, and the cyber posture required to deploy any of this safely. None of that fits inside a marketing budget or an HR budget. The technology function has to scale to carry the weight of what gets built on top of it. Deloitte is direct about this in their own commentary: AI risks starving the rest of the technology portfolio if digital budgets do not continue to expand. The IT line is rising because the foundation is being asked to do more.

    The business-funded line rises faster because the use cases land outside IT. Customer service is buying AI for handle time. Marketing is buying it for personalisation. Legal is buying it for review and contract drafting. Finance is buying it for close acceleration. HR is buying it for sourcing and screening. None of these functions are routing through IT to procure the tools. They are sizing the budget themselves, configuring against their own data, and absorbing the cost line into their own profit and loss.

    The two funding lines. IT-funded and business-funded technology spend, and where they cross.
    Figure 2. The two funding lines. IT-funded and business-funded technology spend, and where they cross.

    Both things are happening. Both lines rise. The non-IT line rises faster because that is where the application layer of AI is being absorbed. The IT line rises because it has to. The result is a structurally larger technology estate where the bigger half is no longer governed by the technology function.

    That is not a budgeting story. That is an organisational design story.

    What is actually being purchased

    The other thing I keep coming back to is what enterprises are actually buying when they buy AI.

    Sequoia put a useful frame on this in their Services: The New Software analysis earlier this year. For every dollar spent globally on software, roughly six dollars are spent on services, the labour that gets work done. Software, in aggregate market terms, has always been a small slice of the total cost of completing enterprise work. Services and labour have been the rest.

    The AI products winning at scale right now are not selling better tools for human workers. They are selling the work itself. They compete inside the labour pool, not the software pool. That is a completely different market and a completely different price ceiling.

    This is the part I think most enterprise tech budgeting still misses. If you measure the AI opportunity as software spend, the curve looks stretched. If you measure it as the cost of completing the work, it has barely started. The 32% of revenue Deloitte projects by 2028 only looks aggressive against a software anchor. Against a services and labour anchor, 32% is the early phase, not the late one.

    Software against services. For every dollar spent on software, roughly six are spent on the labour that gets the work done.
    Figure 3. Software against services. For every dollar spent on software, roughly six are spent on the labour that gets the work done.

    Which is why the budget conversation and the organisational conversation are now the same conversation.

    Where the labour pressure shows up first

    If AI is competing for services and labour budgets, and the data increasingly says it is, then the next question is not financial. It is structural. What happens to the labour those budgets used to fund?

    I want to be careful here. The temptation is to make AI carry the weight of every workforce headline currently in the news. It cannot, and it should not. Many of the layoffs visible in the technology sector right now are also responses to overhiring during 2020 to 2022, post-ZIRP cost discipline, and rising compute costs eating into margin. AI is one force among several. The honest read is to say so.

    What I do think is worth saying, because it is grounded in the historical record rather than in current headlines, is this. Major general-purpose technology transitions tend to produce labour pressure first in the sectors closest to adoption.

    The Industrial Revolution did not begin restructuring labour across the broad economy. It began restructuring labour inside textiles, the industry adopting the new spinning and weaving technologies first. The Luddite movement of 1811 to 1816 was a textile-worker phenomenon, not a general workforce one. The pressure propagated outward to broader manufacturing only as the technology matured.

    Electrification followed a similar pattern. The early organisational impact was felt inside manufacturing, where electric motors eventually changed factory layout, workflow, and the relationship between power and production. The wider economy felt it later.

    The personal computing wave first put sustained pressure on routine clerical and administrative work, the kind of codifiable office activity that software could gradually absorb. The internet transition compressed dramatically inside travel agencies, classified-ad businesses, and directly disintermediated retail before the broader economy felt it.

    The pattern is consistent. Labour pressure first appears inside the industry closest to the technology, then propagates outward as the technology matures.

    What is distinctive about this transition, and this is the observation I have not been able to find a precedent for, is that the technology builders themselves are showing labour pressure during the build-out phase, not after diffusion. In every prior transition, the builders of the technology hired aggressively while the directly displaced industries shrank. The two groups were distinct. In this transition, they sit much closer together, because AI’s first competent applications happen to be concentrated in tasks performed inside technology firms themselves: code, support, content, recruiting.

    I do not think this means the historical pattern is broken. I think it means the historical pattern is compressing. The propagation that took decades in prior transitions may take years in this one. The forward-looking implication is open-ended on purpose, because the pace and sequence of propagation are genuinely uncertain. But the direction is not.

    The optimistic frame is the one that has held up across every prior transition. Each one expanded the economic pie. Each one created entirely new categories of work. Each one rewarded the workers and firms that adapted earliest with most of the upside. The current phase is the transition, not the destination. The leaders who treat it as a cost-takeout exercise will solve the wrong problem. The ones who treat it as an organisational redesign will set the terms for what comes next.

    That is the specific work I think the next eighteen months are about.

    What the redesign actually looks like

    I work mostly with telecom operators, and telecom is a useful case study right now precisely because it has stopped being theoretical. The most AI-fluent industry has begun repositioning, and what that repositioning looks like is becoming visible in the operating data.

    McKinsey’s recent telco analysis finds 57% of telco executives report scaling generative AI use cases across multiple domains. The same survey shows the share of executives describing AI as a blockbuster technology actually declined, from 61% in the prior round to 51% now. The most AI-fluent industry is becoming less breathless about AI even as it deploys it more aggressively. That combination, falling rhetoric and rising deployment, is what industrial maturation looks like. And McKinsey projects AI could deliver 8 to 10 EBITDA percentage points to telco operators over five years. A structural margin uplift of that size is historically rare for a mature industry.

    Telecom as the leading indicator. Scaled deployment rising while the rhetoric cools, with a projected 8 to 10 point EBITDA contribution over five years.
    Figure 4. Telecom as the leading indicator. Scaled deployment rising while the rhetoric cools, with a projected 8 to 10 point EBITDA contribution over five years.

    The clearest production data point in the public record right now is AT&T’s deployment with Microsoft Azure. More than seventy unique generative AI solutions in production. Nine billion daily tokens of measured deployment volume. More than a hundred thousand employees using generative AI tools. Thirty-three percent faster agent productivity in customer-service interactions. Year-on-year return improvement reported by the deployment team.

    What that case shows is not that AT&T has an AI strategy. Plenty of companies have AI strategies. What it shows is what an AI-fluent industry’s production deployment template looks like at scale. Solutions in production. Six-figure employee adoption. Billions of tokens in measured volume. An EBITDA path. The template is portable. Other industries with comparable cost structures and customer bases can read AT&T’s deployment as an operational reference, not an aspirational one.

    The reason telecom matters here is not that it predicts what every industry will do next. It is that it shows what the redesign looks like once it stops being a technology project and becomes an operating model. That distinction is the entire point of this essay.

    The bubble question, briefly

    I should address the framing question that keeps appearing in this debate, because it is a fair one. Is this a bubble?

    A classical bubble shows demand weakening while supply catches up. Valuations compress before fundamentals do. The 2025 to 2026 data does not look like that pattern. Budgets are expanding, not contracting. The supply side is still racing to catch up. Production deployments are scaling, not stalling. Operational metrics are improving where they are being measured.

    This is interpretation, not proof. Plenty of individual AI investments will fail to produce returns. The MIT NANDA work that Deloitte itself cites finds only about 5% of generative AI pilots deliver sustained value at scale, and that gap is real. The aggregate spending curve being structurally sound does not mean every dollar within it will be well spent.

    The risk I think is more concrete than the bubble framing is measurement mismatch, where leaders size AI as a software line item when it is structurally a services phenomenon, and foundation starvation, where AI dollars go into models while the data, security, and identity infrastructure those models require to deploy safely stays underfunded. Deloitte’s own report flags this directly. Only 25% to 32% of organisations invested in identity management, federated security, or zero trust in the past year, even as AI deployment accelerated. That gap is the more immediate strategic risk in the data than any aggregate-curve bubble.

    What I would ask

    If I were writing this for the leadership team I worked with most recently, I would close on the questions I think now matter more than the budget.

    When 25 to 30% of revenue is being deployed against work rather than against software, what does the organisation actually look like? Which functions stop being functions and become flywheels? Which centralised capabilities, meaning data, governance, security and infrastructure, have to be funded as foundational rather than as project costs? Where does the labour redesign live in the org chart, and who owns it?

    These are not budgeting questions. They are design questions. And they are the right questions, because the technology budget is no longer a technology question.

    The leaders asking them in 2026 will own the next decade. The leaders still arguing about the IT line will be answering them in 2030, by then on someone else’s terms.

    The opportunity here is not to manage the curve. It is to design what comes out the other side.

    Sources

    1. Deloitte Center for Integrated Research, AI and tech investment ROI, 2025 Tech Value Survey, October 2025. n=548, average revenue $13.4B. Digital initiatives budget 7.5% in 2024 to 13.7% in 2025, with a continuation projection to roughly 32% by 2028, the IT-funded and business-funded split with crossover by 2027, 74% AI investor share, 36% average AI share of digital budget, cybersecurity investment gap of 25% to 32%, and the MIT NANDA 5% pilot success citation.
    2. Sequoia Capital, Services: The New Software, 2026. The $1 to $6 software-to-services ratio.
    3. McKinsey & Company, The AI Inflection Point: What Leaders Do to Capture Value, telco industry analysis. 57% scaling generative AI across domains, 51% against 61% on framing, and the 8 to 10 percentage point EBITDA five-year projection.
    4. Microsoft customer story and AT&T deployment with Microsoft Azure: more than 70 unique generative AI solutions in production, 9 billion tokens, more than 100,000 employees, and 33% faster agent productivity in customer-service interactions.

    This essay was first published on the Brain AI Insights page on 15 June 2026.

  • The TAM for Intelligence Is Larger Than We Think

    The TAM for Intelligence Is Larger Than We Think

    Intelligence is becoming a general-purpose economic input rather than a feature inside software, and its addressable market is expanding rather than fixed.

    Fifteen months ago, Anthropic was a $1 billion revenue business. Today it is a $30 billion one. That is not a typo.

    In the same period, ChatGPT grew from 400 million to 900 million weekly active users. OpenAI crossed $25 billion in annualised revenue. Enterprise customers spending over $1 million annually with Anthropic doubled from 500 to 1,000 in under two months.

    These are not the signals of a speculative bubble. They are the signals of a market whose total addressable size is being systematically underestimated.

    The three forces expanding the market for intelligence. Falling unit cost, rising consumption per task, and the shift from software budgets to services budgets.
    Figure 1. The three forces expanding the market for intelligence. Falling unit cost, rising consumption per task, and the shift from software budgets to services budgets.

    For years we sized technology markets in familiar buckets: software TAM, cloud TAM, services TAM, advertising TAM. AI is breaking those categories apart.

    Intelligence is no longer a feature inside software. It is becoming a programmable layer that sits across workflows, devices, interfaces, and increasingly autonomous systems. If intelligence becomes abundant, cheap, and deployable almost everywhere, the addressable market is no longer a narrow AI software category. It starts to resemble the market for better decisions, faster execution, lower coordination cost, richer personalisation, and entirely new classes of digital labour.

    Not because the market is literally infinite. Energy, compute, budgets, regulation, and organisational readiness all still matter. But the market for intelligence appears to be economically open-ended in a way traditional software categories are not. As the unit cost of intelligence falls, the number of viable use cases expands. And as model capability rises, each workflow can consume more intelligence, not less.

    That is the setup for a Jevons paradox market.

    Jevons paradox: why falling costs expand the market

    Jevons paradox tells us that when efficiency improves, total consumption can rise rather than fall. We saw it with coal. We saw it with computing. We saw it with bandwidth and storage. We are beginning to see it with intelligence.

    Gartner forecasts that by 2030, performing inference on a one-trillion-parameter model will cost providers more than 90% less than in 2025, and that these models will be up to 100 times more cost-efficient than similarly sized early models from 2022. Those gains come from better semiconductors, model design, higher utilisation, inference-specific silicon, and increased use of edge devices.

    Meanwhile, academic analysis of token pricing suggests that economy-tier inference prices have already fallen roughly 600-fold since 2020, with some models around $0.09 per million tokens by early 2026. At that price, processing a 100-page document costs approximately $0.003. Intelligence becomes essentially free for simple tasks like email sorting, basic customer support, or grammar checking, which is why we are seeing AI integrated into almost every app and device.

    Inference cost decline. Economy-tier token prices have fallen roughly 600-fold since 2020, with Gartner projecting a further 90% reduction by 2030.
    Figure 2. Inference cost decline. Economy-tier token prices have fallen roughly 600-fold since 2020, with Gartner projecting a further 90% reduction by 2030.

    The real unlock: more intelligence consumed per task

    Gartner notes that agentic models can require 5 to 30 times more tokens per task than a standard chatbot. A chatbot produces an answer. An agent reasons across steps, calls tools, retrieves context, evaluates options, retries, plans, and executes actions. Each useful unit of work consumes materially more intelligence.

    So the economics move in two directions at once. The cost per token is falling sharply, and the tokens consumed per useful task can rise dramatically as systems become more agentic.

    The market does not have to choose between lower prices and higher demand. It can have both. Intelligence gets cheaper per unit, but richer workflows consume more of it. The result is not a collapse in value. It is a broadening and deepening of demand.

    Tokens consumed per task. Agentic workflows consume 5 to 30 times more tokens than a single-turn chatbot exchange.
    Figure 3. Tokens consumed per task. Agentic workflows consume 5 to 30 times more tokens than a single-turn chatbot exchange.

    This is why the argument that token optimisation kills the market misses the point. Routing, caching, smaller specialist models, better context handling, and hybrid model stacks will all push effective unit costs down further. But that does not shrink the market. It opens the door to always-on intelligence in thousands of use cases that were previously too expensive to justify.

    Revenue is now scaling exponentially

    Anthropic: enterprise demand is real and accelerating. Anthropic’s annualised revenue run-rate trajectory in 2026 has been remarkable: approximately $9 billion at the end of 2025, $14 billion in February 2026, $19 billion in March, and more than $30 billion in April. That is a more than threefold increase in roughly four months.

    The number of enterprise customers spending over $1 million annually doubled from 500 to 1,000 in less than two months. Claude Code alone was generating over $2.5 billion in annualised run-rate revenue by February. Anthropic also announced a compute agreement with Google and Broadcom for approximately 3.5 gigawatts of next-generation TPU capacity starting in 2027.

    OpenAI: consumer intelligence as a mass utility. If Anthropic is proving enterprise willingness to pay, OpenAI is proving intelligence can behave like a mass consumer utility. ChatGPT reached 900 million weekly active users as of February 2026, up from 400 million a year earlier. It now has over 50 million paying consumer subscribers and more than 9 million paying business users. Annualised revenue exceeds $25 billion, with OpenAI reporting $2 billion in monthly revenue. 92% of Fortune 500 companies use the platform.

    Revenue trajectory. Anthropic from roughly $9 billion annualised at the end of 2025 to more than $30 billion by April 2026, alongside OpenAI above $25 billion.
    Figure 4. Revenue trajectory. Anthropic from roughly $9 billion annualised at the end of 2025 to more than $30 billion by April 2026, alongside OpenAI above $25 billion.

    AI is no longer just one thing. It is simultaneously a consumer utility used weekly by nearly a billion people and an enterprise production layer monetised across APIs, coding, workflow augmentation, and agentic execution. That is not how narrow software categories behave. It is how foundational infrastructure behaves.

    Intelligence does not just replace software, it replaces services

    Here is where the argument gets larger. Sequoia Capital’s Julien Bek framed it sharply in March 2026: for every dollar spent on software, six are spent on services. If AI is only competing for the software budget, we are measuring the wrong market.

    The key insight is the distinction between selling the tool and selling the work. A copilot sells the tool, making a professional more productive. An autopilot sells the work, and the customer buys the outcome directly. When AI becomes capable enough, the addressable market shifts from the tool budget to the labour budget. And the labour budget is multiples larger.

    Insurance brokerage alone is $140 to $200 billion. IT managed services exceed $100 billion. Recruitment and staffing exceeds $200 billion. Management consulting is $300 to $400 billion. These are not software markets. They are services markets that intelligence is beginning to enter.

    Services markets that intelligence is entering. Management consulting, recruitment and staffing, insurance brokerage and IT managed services, sized against the software budgets AI is usually measured against.
    Figure 5. Services markets that intelligence is entering. Management consulting, recruitment and staffing, insurance brokerage and IT managed services, sized against the software budgets AI is usually measured against.

    Sequoia’s framework maps every services vertical on an intelligence-to-judgement spectrum. The higher the intelligence ratio in any field, the sooner AI autopilots will win. Software engineering got there first, and more tasks are now started by agents than by humans in tools like Cursor. But it is coming to accounting, legal, healthcare billing, insurance, tax advisory, and procurement.

    The next $1 trillion company will be a software company masquerading as a services firm. Julien Bek, Sequoia Capital

    This reframes the entire conversation. AI is not just a better tool for existing software workflows. It is a substitute for professional services labour. And the professional services market is measured in trillions, not billions.

    Frontier plus open source: the market expands in both directions

    The strongest future state is not frontier or open source. It is frontier plus open source. Frontier models anchor the highest-value reasoning, multimodal, and agentic workloads. Open and lower-cost models expand the market downward by making intelligence cheap, customisable, and deployable inside more products and workflows.

    Falling economy-tier prices do not cannibalise the category. They widen it. They allow more applications to become viable while premium models capture the most demanding work. That layered structure is exactly what you would expect in a market that is still expanding.

    Enterprise adoption is still remarkably low

    Despite the visibility of AI, enterprise penetration remains early. Eurostat reports 20% of EU enterprises with ten or more employees used AI technologies in 2025, with a steep size gap: 55% for large enterprises against just 17% for small ones. OECD data shows 20.2% of firms using AI in 2025, up from 8.7% in 2023, a 132% increase, but still early.

    Deloitte’s 2026 State of AI report found that worker access to AI rose 50% in 2025, but only 34% of organisations are truly reimagining their business with AI. McKinsey found only one-third of respondents were scaling AI across their organisations. The bottleneck is no longer whether the technology works. It is increasingly whether the organisation can deploy it safely, responsibly, and at scale.

    Enterprise penetration. Roughly one in five firms across the EU and OECD were using AI in 2025, with a wide gap between large and small enterprises.
    Figure 6. Enterprise penetration. Roughly one in five firms across the EU and OECD were using AI in 2025, with a wide gap between large and small enterprises.

    The use case explosion is just beginning

    We have seen an explosion in chatbots, coding assistants, writing help, and copilots. But the larger categories are still nascent: agentic workflows, AI-native business processes, machine-to-machine coordination, scientific AI, industrial AI and robotics, edge AI, domain-specific agents, and personal AI. Models are improving rapidly, costs are falling, and each improvement unlocks new categories.

    The constraint is not demand, it is energy and compute

    The International Energy Agency makes it plain: affordable, reliable, and sustainable electricity supply will be a crucial determinant of AI development. Anthropic has secured approximately 3.5 gigawatts of future TPU capacity. OpenAI’s Stargate initiative has discussed a 10-gigawatt buildout. Hyperscalers collectively are planning hundreds of billions in AI infrastructure spending.

    A saturated market asks where the demand is. This market increasingly asks whether capacity can be built fast enough. That is a profoundly different signal.

    Does this settle the bubble argument?

    This evidence does not prove that every part of the AI stack is fairly priced. It does not mean every infrastructure dollar will earn an adequate return. History suggests platform shifts come with pockets of excess.

    But it does weaken the simplistic version of the bubble thesis considerably. That argument rests on the claim that AI demand is mostly speculative. The evidence is increasingly inconsistent with that claim. Consumer usage is massive, at 900 million weekly active users on one platform alone. Enterprise monetisation is real, at a $30 billion run-rate for Anthropic and $25 billion for OpenAI, with more than a thousand enterprises paying over $1 million a year. Revenue growth is exponential. Agentic workflows consume 5 to 30 times more tokens per task. Unit costs are collapsing. Penetration remains at roughly 20% of enterprises across the EU and OECD. And the services market is opening up.

    That is not the profile of a category that has exhausted its market. It is the profile of a category that may still be very early in converting capability into deployment, and deployment into economic value.

    The real conclusion

    The most useful way to think about AI now is not as a single product market but as the emergence of programmable intelligence as infrastructure.

    As intelligence gets cheaper, more capable, and easier to embed, more of the economy becomes addressable. As systems become agentic, each workflow consumes more intelligence, not less. As governance improves, more enterprises move from experimentation to scale. As AI shifts from copilot to autopilot, it captures not just the software budget but the vastly larger services budget. And because adoption remains relatively early, the distance between today’s usage and tomorrow’s usage may still be very large.

    The addressable market for intelligence is not literally infinite. But it may be far more expansive than the market still assumes. Because intelligence is no longer just a feature inside software. It is becoming a general-purpose economic input. And when a scarce input becomes abundant, programmable, and cheap enough to use everywhere, history suggests something important: demand does not plateau. It compounds.

    Sources and references

    1. Gartner, inference cost projections for 2030, March 2026.
    2. arXiv, token price evolution analysis, 2603.28576v1.
    3. Sequoia Capital, Services: The New Software, March 2026.
    4. Bloomberg, Anthropic $30 billion run-rate, April 2026.
    5. TechCrunch, ChatGPT reaches 900 million weekly active users, February 2026.
    6. McKinsey, State of AI Trust in 2026.
    7. Eurostat, 20% of EU enterprises use AI, 2025.
    8. Deloitte, State of AI in the Enterprise 2026.
    9. International Energy Agency, Energy and AI, executive summary.
    10. OpenAI, State of Enterprise AI 2025 report.

    This essay was first published on the Brain AI Insights page on 1 April 2026.

  • The Confidence Trap: AI Has Learned to Reason, It Has Not Learned to Doubt

    The Confidence Trap: AI Has Learned to Reason, It Has Not Learned to Doubt

    As AI moves from productivity tool to operational backbone, the reliability problem that matters is the system that is wrong but does not know it. Primary benchmark data.

    Enterprise AI is delivering real value and the deployment momentum is real. The organisations that will lead the next decade are those building the governance architecture to make that value trustworthy at scale.

    “The question for enterprise AI in 2026 is no longer whether it works. It is whether it works reliably enough, consistently enough, and with enough epistemic discipline to be trusted with the decisions that matter most.”

    Enterprise AI is delivering. McKinsey’s 2026 AI Trust Maturity Survey found that twice as many leaders as last year are reporting transformative impact from AI investments. Deloitte’s survey of 3,235 executives confirms that organisations scaling AI across core business functions are capturing real productivity gains. Worker access to AI tools rose 50% in 2025 alone. The deployment momentum is real, the ROI cases are accumulating, and the technology has moved well beyond experimentation into production at scale.

    This is the context in which a more precise challenge deserves attention. Not as a reason to hesitate, but as the next frontier to cross. As AI moves from productivity tool to operational backbone, a specific class of reliability problem becomes commercially significant: the system that is wrong but does not know it. Not the AI that fails visibly, but the AI that fails with fluency. Not the obvious error a human reviewer catches, but the confident, well-structured, contextually plausible inaccuracy that passes through unchallenged into a report, a customer response, a compliance filing, a strategic decision. The missing capability is not intelligence. It is doubt.

    This essay examines that specific challenge with precision, using primary benchmark data published in early 2026. It looks at where the gap is, why it is architectural rather than just a model quality issue, and what the next generation of enterprise AI systems must be designed to do differently. The goal is not to cast doubt on AI adoption. It is to describe what separates AI that is useful from AI that can be genuinely trusted, and to make the case that building that distinction into the architecture is the most valuable investment an enterprise can make right now.

    The precision gap: where today’s AI needs to go next

    Here is what makes the current moment interesting rather than alarming. AI capability has advanced significantly. Models reason, plan, use tools, and synthesise complex information in ways that were not possible even two years ago. And yet one specific capability has not kept pace: the ability to know, and signal, the boundaries of what the system actually knows with certainty. The gap between what a model can generate and what it can verify is the frontier the industry must now cross.

    A concrete illustration of this frontier appeared when Vectara updated its hallucination leaderboard with a harder, business-context benchmark in late 2025 and March 2026. The finding was counterintuitive: reasoning models, the most advanced systems, the ones that deliberate and reflect, showed higher hallucination rates under business conditions than simpler models on simpler tasks. Under the business context filter, several leading frontier models registered rates above 20%. This is not a verdict on AI capability. It is a precise diagnostic of where the architecture needs to evolve.

    The explanation points directly at the design gap. Reasoning models invest computational effort in constructing chains of inference. The problem is that this process is generative rather than verificatory. The model reasons from its training distribution. It does not cross-reference against external ground truth, because no architectural mechanism requires it to. When reasoning is rewarded during training, models learn to produce longer, more structured, more confident-sounding reasoning chains, regardless of whether the underlying claims are correct. Researchers call this performative reasoning : the appearance of deliberation without the substance of verification.

    MIT research (January 2025) found that AI models are 34% more likely to deploy phrases like “definitely,” “certainly,” and “without doubt” when generating incorrect information than when generating accurate information. This is the precision gap in its sharpest form: the system’s confidence signal is inversely calibrated to its actual accuracy.

    What makes this commercially significant and worth solving is that enterprise deployment amplifies the stakes. When AI fails visibly, organisations catch and correct it. When it fails fluently and confidently, the error propagates into customer interactions, compliance documents, strategic briefings. The value of AI at scale depends not just on how often it is right, but on whether it reliably signals when it might not be. That is the capability the next generation of enterprise AI systems must build in.

    Knowing when the question is wrong: a solvable capability gap

    The precision gap has a second dimension that deserves equal attention. It is not just about whether AI generates incorrect facts. It is about whether AI recognises when the question itself is built on a flawed assumption. In enterprise environments, users frequently bring queries grounded in misunderstandings, incorrect premises, or category errors. The commercially valuable response is to surface and correct the flawed assumption, not to answer the question as stated and propagate the error downstream.

    BullshitBench (v2) measures exactly this capability across 100 questions using 13 distinct nonsense techniques, spanning software, finance, legal, medical and physics domains. The headline finding is genuinely encouraging: it is demonstrably possible to build systems that do this well. Claude Sonnet 4.6 on high reasoning achieves 91% clear pushback, correctly identifying and declining to engage with broken premises 91 times in 100. Claude Opus 4.5 follows at 90%, and Claude Sonnet 4.5 reaches 79%. The capability exists. It can be built into a system by design.

    The gap between the leaders and the field is substantial and instructive. Gemini 3 Pro and GPT-5.4 both reach 48%, a genuine improvement over older models, but still meaning nearly half of all nonsensical questions are answered with confident authority. Gemini 3.1 Pro sits at 37%. OpenAI’s o3, despite being a flagship reasoning model, reaches only 26%, lower than several lighter, older models. The BullshitBench data offers one of the clearest illustrations of how raw capability ranking and deployment reliability ranking diverge. More reasoning tokens do not reliably produce better premise detection. What produces it is architectural design: specifically, whether the system is built to challenge inputs, not just respond to them.

    AI systems are optimised for task completion, trained to be helpful, to find an answer, to engage. The very quality that makes them valuable in open-ended contexts works against them when the right response is to challenge the premise of the question itself. What BullshitBench shows, and what its rapid growth to over 1,100 GitHub stars suggests the industry has recognised, is that premise-checking is a distinct, measurable capability that can be built. The gap between the best and worst performing systems is not small. It is the difference between a system that reliably safeguards enterprise decisions and one that does not.

    “The capability that distinguishes enterprise-grade AI is not just the ability to reason. It is the discipline to challenge the premise before reasoning begins.” The Illusion of Intelligence, Reference Essay, 2026

    The consequences in production are direct and positive when designed for. A customer who misunderstands their contract terms, a compliance officer who frames a query on a false regulatory assumption, an analyst who poses a question built on stale data: in each case, an AI system designed to surface and correct the flawed premise adds genuine, measurable enterprise value. That design difference is not a model quality question. It is an architectural one, and it is solvable.

    Reading the benchmark data: a navigation tool, not a verdict

    The Vectara Hallucination Leaderboard is the most widely cited benchmark in enterprise AI evaluation. Its March 2026 update, filtering by business context, commercial models, and large model tiers, gives the most operationally relevant picture yet of where frontier models stand when deployed in enterprise conditions. The right way to read this data is not as an indictment of any model, but as a deployment guide. Different models perform very differently on different tasks, and the selection question matters enormously.

    The spread is wide and instructive. At the top of the leaderboard under business conditions, Gemini 2.5 Flash Lite reaches 3.4%, a genuinely low rate. GPT-4.1 achieves 7.7%. Most frontier mid-tier models cluster between 12% and 20%. The notable observation is that the newest flagship reasoning models, o3-pro at 32.5%, Grok-4 variants at 30 to 32%, GPT-5 at 24 to 25%, Gemini 3 Pro at 23.5%, and Claude Opus 4.6 at 23.4%, are not the most accurate on this benchmark under business conditions. Raw intelligence ranking does not map to factual reliability ranking in enterprise RAG and summarisation tasks. Choosing AI for enterprise deployment means asking a more precise question: not which model is most capable, but which model is most reliable for this specific task under these specific conditions.

    What enterprise leaders are actually experiencing

    McKinsey’s 2026 AI Trust Maturity Survey, published March 25, 2026 and drawing on approximately 500 organisations with direct AI governance responsibility, gives us the clearest primary picture yet of where enterprise AI confidence actually stands. The findings are instructive not as an indictment of the technology, but as a precise map of where the gaps are.

    These findings reveal a consistent pattern: organisations are aware of the risks, are investing in AI at scale, and are making meaningful progress. Governance and accountability structures are, however, lagging behind technical deployment. McKinsey finds that active mitigation trails risk awareness across nearly every AI risk category. Organisations know what could go wrong. Building the structural controls to match that awareness is the work underway.

    Two findings from the McKinsey survey deserve particular attention. First, nearly 60% of respondents cite knowledge and training gaps, not budget or executive support, as the primary barrier to implementing responsible AI practices. This is a capability problem, not a commitment problem. Organisations want to govern AI well and are still building the skills to do it. Second, and critically for the architecture argument: organisations that invest explicitly in responsible AI and assign clear ownership for it report materially higher maturity scores and are far more likely to achieve EBIT impact above 5%. Governance, in other words, is not a cost. It is a value driver.

    The architectural design opportunity

    Understanding why this precision gap exists points directly at how to close it, and the first thing to say is that the frontier labs have already moved a long way. Safety is no longer bolted on after generation. OpenAI’s deliberative alignment trains reasoning models to recall and reason over safety specifications before they draft an answer. Anthropic’s constitutional classifiers screen an input before the model sees it at all, alongside output classifiers that assess a response as it streams. Constitutional training shapes the model during training rather than filtering it afterwards. Anyone still describing frontier systems as generate-then-filter is describing 2023.

    The gap that remains is narrower and more specific: whose policy is applied, and who checks the result. Lab alignment encodes the lab’s policy, applied universally to every customer of that model. However well it is done, it cannot encode what one enterprise, in one jurisdiction, is permitted to say to one customer. That question is answered per tenant, and no model provider is positioned to answer it.

    The same applies to self-checking. The most sophisticated version is the reasoning loop: a generator produces an answer, a verifier scores it, the generator tries again if the score is insufficient. This is genuinely useful, and it is rigorous as far as it goes. But it carries a structural limitation: the verifier and the generator share the same underlying model characteristics. If the generator does not know a fact, the verifier cannot reliably identify that the generated fact is wrong. It is, in technical terms, self-referential, and self-referential systems cannot audit themselves against external ground truth without an independent mechanism to do so.

    Current AI architectures are designed to generate and refine answers, but the evaluation function is not structurally independent of the generation function. The next generation of enterprise AI must invert this: governance compiled before reasoning begins, evaluation that operates independently of generation, and feedback loops that evolve policy constraints rather than just correcting individual outputs. This is an engineering problem with a known solution direction.

    There is a second architectural dimension that creates significant enterprise opportunity: the governance layer. Current systems determine how to behave primarily through system prompts, static instructions fed to the model before conversation begins. Prompts can suggest tone, topic scope, and response format. What they cannot do is compile regulatory requirements, cultural norms, brand guidelines, and tool permissions into runtime constraints that govern the reasoning process itself. They are instructions, not architecture. The enterprise AI systems that will define the next generation are those where governance is designed in rather than prompted in.

    For enterprises operating across regulatory jurisdictions, cultural contexts, and brand environments, a telco serving GCC and European markets simultaneously or a bank operating under multiple compliance regimes, this distinction is not academic. It is the difference between AI that can be deployed with confidence across all contexts and AI that requires constant human oversight to manage the gaps between what it was prompted to do and what each context actually requires.

    The invisible dimension: cultural and emotional context

    There is an additional capability gap that receives far less attention than hallucination but carries equivalent commercial risk in certain industries, particularly in telecommunications, financial services, and any enterprise operating across multiple cultural contexts: the absence of genuine cultural and emotional intelligence in AI systems.

    A system can produce a factually correct response to a customer complaint that is simultaneously culturally tone-deaf, emotionally inappropriate, and damaging to the brand relationship. In high-context cultures such as the GCC, East Asia, and much of Latin America, the way something is said is as commercially significant as what is said. Current AI systems have no architectural mechanism for enforcing this distinction. Cultural alignment, where it exists, is baked into model training or approximated through prompts. Neither approach is deterministic. Neither scales reliably across the hundreds of cultural, regulatory, and brand contexts a global enterprise must navigate.

    The enterprise that deploys a single AI system into a dozen markets is not deploying one solution. It is deploying one system into twelve different cultural and regulatory environments and hoping the model generalises correctly. The governance gap here is significant, and almost entirely unquantified.

    What next-generation enterprise AI is designed to do

    The precision gap is an engineering problem with an architecture solution. The leading edge of the industry is already building in this direction and the design principles are clear. Enterprises evaluating AI platforms, and vendors building them, should expect the next generation to demonstrate five structural properties that current generation systems largely lack:

    • Governance compiled before reasoning begins. Behavioural constraints, including regulatory requirements, cultural norms, brand guidelines, and tool permissions, should be compiled into runtime constraints that govern the reasoning process itself, not layered as prompts on top of it. The system should know its boundaries before it starts thinking, not discover them at the output stage.

    • Evaluation that is architecturally independent of generation. The function of scoring and validating an output must be structurally separated from the function of producing it. A verifier sharing the same model substrate as the generator inherits its blind spots. Genuine independence, operating on different principles and against different criteria, is what enables reliable self-correction rather than self-referential refinement.

    • Structured adversarial challenge, not just reflection. Systems should contain mechanisms that actively challenge their own reasoning across multiple cognitive dimensions simultaneously, not merely reflect on the output. Planning, emotional appropriateness, factual grounding, and compliance should each be evaluated by dedicated reasoning functions that challenge the consensus, not just endorse it.

    • Grounded, traceable evidence chains. Every significant claim in an output should be traceable to a source. Claims that cannot be grounded should be flagged, not papered over with fluent prose. The system should function as a narrator anchored in verified evidence, not a generator improvising plausible content.

    • Closed-loop policy evolution. Evaluation findings should feed back into governance constraints automatically, not just into retraining queues. When the system identifies a recurring pattern of cultural misalignment, compliance drift, or contextual error, the governance layer should update to prevent the same class of error in future. This is the difference between a system that corrects mistakes and a system that learns not to repeat them.

    These are not theoretical ideals. Each addresses a specific, demonstrable gap in current systems, documented in the benchmark data examined in this essay. Together, they describe an AI architecture where governance is not an overhead on top of capability, but an integral dimension of it. That architecture is what makes enterprise AI trustworthy at scale, and it is the design investment that will separate the leading deployments of 2027 and 2028 from those still managing reliability through human oversight alone.

    This is the business case stated plainly. McKinsey’s survey of approximately 500 organisations found that responsible AI investment is not a cost centre but a value multiplier. Organisations that govern well outperform those that don’t. The governance architecture question is therefore not a risk management question. It is a competitive strategy question.

    The governance opportunity

    There is a final dimension worth naming directly, and it sits not with the technology but with the leadership structures around it. McKinsey’s 2026 survey found that organisations with explicit ownership for responsible AI achieve meaningfully higher maturity scores than those without. Accountability, it turns out, is not just a governance formality. It is a performance variable.

    The organisations moving fastest on this are doing something specific: they are treating AI governance as a product requirement, not a post-deployment process. They are asking, before systems go live, who reviews evaluation outputs, who approves policy updates, and who has authority to adjust deployment parameters when reliability falls below threshold. These are not technical questions. They are leadership design questions.

    McKinsey also found that AI trust is increasingly viewed as a business enabler rather than a compliance exercise. This framing shift matters. Enterprises that govern AI well are not just reducing risk. They are building the institutional capability to deploy AI more ambitiously, more quickly, and with more confidence than competitors who haven’t yet addressed the governance layer. Courts are reinforcing this from another direction: judges in 2025 consistently ruled that the human or organisation submitting AI-generated content bears responsibility for its accuracy. Governance expectation is being written into legal precedent. Getting ahead of it is a structural advantage.

    The technology is ready for enterprise deployment at scale. The models are capable. The ROI cases are real, and McKinsey’s data shows that organisations investing in governance alongside capability are realising materially better returns. What the next phase of adoption requires is governance architecture that matches the ambition of the deployment: systems that behave predictably across cultural contexts, challenge flawed assumptions rather than amplifying them, and evolve their behavioural boundaries in response to what they learn. That is the next frontier. And the enterprises that define it will shape how the industry follows.

    “AI has learned to reason. The next frontier is building systems that also know when to doubt themselves. Not slower systems or more cautious ones, but more honest ones. Governed before they reason, evaluated independently after, and improving continuously between. That is the architecture that earns trust at scale.”

    Primary data sources: Vectara HHEM Hallucination Leaderboard, data extracted March 31, 2026 (huggingface.co/spaces/vectara/leaderboard), filters: business context, commercial, large models · BullshitBench V2, Peter Gostev / Arena.ai, March 2026 (arena.ai/blog/inside-bullshitbench) · McKinsey & Company, “State of AI Trust in 2026: Shifting to the Agentic Era,” March 25, 2026, survey of approximately 500 organisations (mckinsey.com) · Deloitte AI Institute, “State of AI in the Enterprise,” 2026, survey of 3,235 leaders, August to September 2025 (deloitte.com) · MIT Research, AI confidence and hallucination correlation study, January 2025

    The reference essay “The Illusion of Intelligence: Why AI Still Hallucinates in the Age of Reasoning” (2026) informed the framing of sections II and V. The Vectara business context leaderboard data was extracted directly from a leaderboard screenshot dated March 31, 2026.

    This essay was first published on the Brain AI Insights page on 1 February 2026.

  • AI Is Not Becoming Electricity Yet, And It Is Not a Bubble Either

    AI Is Not Becoming Electricity Yet, And It Is Not a Bubble Either

    Artificial Intelligence today sits between two polarised narratives. On one side, AI is described as the next electricity, destined to become.

    Executive Summary

    Artificial Intelligence today sits between two polarised narratives. On one side, AI is described as the next electricity, destined to become invisible, ubiquitous infrastructure. On the other, it is dismissed as a bubble, overhyped, overfunded, and due for correction. Both views feel compelling, yet neither fully explains what individuals, enterprises, and institutions are actually experiencing.

    This paper argues that the confusion is not about AI’s potential, but about how we interpret its evolution. Generative AI is not a single technological moment or a clean platform shift; it is a layered capability journey unfolding over time. From foundation models and conversational interfaces to agentic systems, multi-agent coordination, and autonomous intelligence, each layer matures on its own trajectory, with its own expectations, value creation patterns, and friction.

    This layered evolution explains why excitement and disillusionment coexist, why individuals adapt quickly while enterprises hesitate, and why debates about ROI, readiness, and risk feel unresolved. Rather than asking whether AI is hype or reality, this paper reframes the question: how do we design systems, organisations, and mindsets that can absorb continuous advances in intelligence without becoming brittle?

    Two Narratives. One Reality We’re Missing.

    Two narratives, one reality. AI is described either as the next electricity or as a bubble, and neither account explains what enterprises are actually experiencing.
    Figure 1. Two narratives, one reality. AI is described either as the next electricity or as a bubble, and neither account explains what enterprises are actually experiencing.

    Two claims dominate today’s AI conversation.

    The first says: AI will become like electricity. Invisible. Ubiquitous. Always on.

    The second says: AI is in a bubble. Overpromised. Oversold. Due for a correction.

    These views appear irreconcilable, yet both resonate. That alone should tell us something important.

    They are not arguing about the same thing. They are describing different time horizons, different layers of capability, and different definitions of impact. The real mistake we keep making is treating AI as a moment, something that either “arrives” or “fails”, rather than as an evolutionary process.

    Transformational Technologies Don’t Arrive. They Accumulate.

    Electricity did not become electricity the moment generators were invented. It took decades of infrastructure build-out, safety standards, appliances, regulation, skills transformation, and societal trust before it disappeared into the background and became foundational.

    Computing followed the same pattern, from mainframes to PCs, from the internet to cloud and mobile, and now into embedded intelligence. Each phase felt revolutionary. Each phase disappointed expectations. None of them were the end state.

    AI fits this historical pattern, with one crucial difference: speed. Capability layers are emerging faster than our institutions, organisations, and mental models can comfortably absorb.

    AI Evolves in Layers, Not on a Single Curve

    AI evolves in layers rather than on a single curve. Foundation models, conversational interfaces, agentic systems, multi-agent coordination and autonomous intelligence each mature on their own trajectory.
    Figure 2. AI evolves in layers rather than on a single curve. Foundation models, conversational interfaces, agentic systems, multi-agent coordination and autonomous intelligence each mature on their own trajectory.

    What we casually call “AI” today is not one thing. It is a stack of capabilities in motion.

    We began with foundation models, systems capable of large-scale probabilistic reasoning. We then wrapped them in interfaces, chatbots and copilots, making them usable. We are now entering the era of agentic AI, systems that can reason, plan, and act using tools. Beyond this sit multi-agent systems, where coordination and orchestration matter more than raw intelligence. Then come autonomous systems, AI that adapts and optimises over time with minimal human direction.

    There may well be a final horizon, superintelligence, but that is a theoretical boundary, not today’s adoption challenge.

    The critical insight is this: each layer does not replace the previous one. It absorbs and reframes it. And each layer generates its own cycle of excitement, expectation, and disappointment. That is why AI can simultaneously feel transformational and unfinished.

    Why the “Electricity” Analogy Feels Right, and Why It Misleads

    The claim that AI will become like electricity is directionally correct, but temporally wrong. Electricity became powerful only once it became boring. AI is still fascinating, which tells us it is not yet infrastructure.

    When leaders say “AI is the new electricity,” they are often describing where this ends up, not where we are. Confusing those two leads to unrealistic expectations, rushed investments, and inevitable backlash.

    Why the “Bubble” Narrative Persists

    At the same time, sceptics are reacting to something real. We see inflated expectations, fragile pilots, duplicated investments, and plenty of AI theatre. Many organisations have built impressive demonstrations that stubbornly refuse to scale.

    Calling this a bubble assumes a single rise-and-fall story. What we are actually seeing is misaligned expectations attached to immature layers, while deeper layers continue to advance. Disillusionment in one layer often coincides with breakthroughs in another. That is not collapse, it is layered evolution.

    Platform Shift or Industrial Revolution?

    Platform shift or industrial revolution. How this transition compares against previous general-purpose technology waves.
    Figure 3. Platform shift or industrial revolution. How this transition compares against previous general-purpose technology waves.

    Is AI just another platform shift, like cloud or mobile? Or is it something closer to an industrial revolution?

    The answer is: both. AI begins life behaving like a platform, tools, productivity gains, ROI cases. But it acts like an industrial revolution because it reshapes work, skills, organisational design, and decision-making, and never truly stabilises.

    Platform shifts settle. Industrial revolutions compound. This duality explains much of today’s enterprise anxiety.

    Individuals Learn Fast. Enterprises Pause, Rationally.

    For individuals, AI adoption is a learning journey. You experiment, fail, and adapt. Switching costs are low and skills compound quickly.

    Enterprises live in a different reality. They carry legacy systems, compliance requirements, architectural debt, workforce implications, and board-level accountability. Their hesitation is not ignorance, it is rational risk management.

    Their real question is not “Should we adopt AI?” It is: how do we build today without locking ourselves into yesterday’s intelligence?

    The Enterprise Paradox

    The enterprise paradox. Individuals adopt quickly while organisations pause, and the pause is rational.
    Figure 4. The enterprise paradox. Individuals adopt quickly while organisations pause, and the pause is rational.

    This creates a painful tension. Move too slowly, and risk falling behind. Move too quickly, and risk anchoring on the wrong abstraction.

    The result is familiar: pilots everywhere, impact nowhere. The issue is not ambition. It is where decisions are being anchored, on tools instead of foundations.

    From AI Solutions to AI-Resilient Systems

    The more useful reframing is this: the question is not whether to build chatbots, agents, or autonomous systems.

    The question is whether your technical and organisational architecture can survive multiple generations of AI capability.

    That means prioritising modular, composable foundations; clear separation between reasoning, execution, and governance; and organisations designed to evolve continuously, not periodically. This is not scepticism. It is resilience.

    ROI Is Not the Problem. Timing Is.

    Early AI layers deliver productivity gains. Agentic layers unlock process transformation. Later layers will enable business model change.

    Expecting late-stage returns from early-stage capabilities almost guarantees disappointment, and fuels the “AI doesn’t work” narrative. In reality, the mismatch is not between value and reality, but between expectation and maturity.

    Technology Is Ready. Humans Are the Constraint.

    Where the constraint actually sits. Capability is available; organisational readiness is what determines the pace.
    Figure 5. Where the constraint actually sits. Capability is available; organisational readiness is what determines the pace.

    AI capability is advancing at extraordinary speed, new models, new breakthroughs, almost daily.

    Human systems are not. Most enterprises were designed for stability. AI demands adaptability: fungible skills, new roles, new incentives, and comfort with perpetual change. The bottleneck today is no longer technology. It is human and organisational readiness.

    So Where Are We, Really?

    We are not at “AI as electricity.” We are between usability and early operationalisation, an uncomfortable but necessary phase where hype, fear, progress, and confusion coexist. This leads to a critical conclusion: we should expect many AI hype cycles, not one. Confusion is not failure, it is a signal that layered evolution is underway.

    The real risk is not experimenting too early. It is building brittle systems, technical or organisational, that cannot evolve.

    Better Questions for the Next Phase

    Instead of asking:

    • Is AI hype or reality?
    • Should we act now or wait?

    We might ask:

    • Which AI capability layer am I actually engaging with?
    • Am I building a solution, or an evolving system?
    • Is my organisation designed to absorb continuous intelligence?

    Because when AI finally becomes boring, when it fades quietly into the background, that is when it will have truly become powerful. Until then, debate is not noise. It is a necessity.

    This essay was first published on the Brain AI Insights page on 26 January 2026.