Two Clocks: Why Operating Models, Not Models, Will Decide Who Profits from AI
The technology is compounding on a cadence of months. Organisations change on a cadence of years. The gap between those two clocks, not any limitation of the models themselves, is what will hold back the full exploitation of AI. The evidence for this is now overwhelming, and the firms that act on it first will own the next decade.
The fast clock
Begin with what is not in dispute. The capability of frontier AI systems is improving at a pace with no precedent in enterprise technology. The research organisation METR, which measures AI capability by the length of task a system can complete autonomously, found in 2025 that the task horizon of leading models had doubled roughly every seven months since 2019. When METR updated its estimates in January 2026, the post-2023 doubling time had compressed to around 131 days. A model that could reliably complete a fifty-minute engineering task in early 2025 has successors whose measured horizons now run to a working day or more. Whatever methodological quarrels one has with any single benchmark, and there are legitimate ones, the direction and the steepness of the curve are not seriously contested.
Set that against the clock on which large organisations actually change. A policy estate is reviewed annually. An operating-model redesign in a regulated financial institution is a two-to-three-year programme. Role definitions, approval thresholds, delegated authorities, four-eyes controls and audit-evidence standards were written for a workforce composed entirely of humans, and they are rewritten at the speed of committees. Clayton Christensen (1997) could assume that an incumbent threatened by disruption had years to respond; the standard prescription of standing up a separate unit takes longer to execute than a full doubling of the technology it is meant to exploit. The asymmetry is structural. Technology compounds; organisations amortise.
The conclusion most boards have drawn from the fast clock is that AI is a technology adoption problem: pick the model, build the platform, run the pilots, harden the integration. This essay argues the opposite. The binding constraint on AI value is not the technology and will not be the technology at any point in the planning horizon that matters. The constraint is organisational: workflows, decision rights, role definitions, governance, culture and cost discipline. The technology is the forcing function, not the solution. The companies that win will be the ones that treat AI as a redesign problem and use the models to force the redesign.
What the failure data actually says
If model capability were the constraint, enterprise results should be improving as the models improve. They are not. S&P Global Market Intelligence’s 2025 Voice of the Enterprise survey, drawing on just over a thousand respondents across North America and Europe, found that the proportion of companies abandoning most of their AI initiatives before production rose from 17 per cent in 2024 to 42 per cent in 2025, with the average organisation scrapping 46 per cent of its proofs of concept. RAND’s analysis puts the AI project failure rate above 80 per cent, roughly twice the rate of non-AI technology projects. That differential is the diagnostic figure. The general difficulty of corporate IT explains the baseline; only something specific to AI explains the doubling, and the candidate explanations all point at the organisation rather than the model.
MIT NANDA’s much-quoted GenAI Divide report claimed that 95 per cent of enterprise generative-AI pilots deliver no measurable profit-and-loss impact against an estimated thirty to forty billion dollars of spend. The headline number has been credibly challenged on methodology, notably by Futuriom, and should be handled with care. But the direction is corroborated independently: Capgemini found in 2023 that 88 per cent of pilots failed to reach production, and BCG’s survey of a thousand executives across 59 countries found 74 per cent of companies struggling to achieve and scale value, with only a quarter having moved beyond proofs of concept. Four research streams with different methods, samples and incentives converge on the same shape of result. Spend is up, capability is up, and realised value at the level of the income statement is stubbornly flat.
The most telling number in the NANDA work is not the 95 per cent. It is the shadow-AI gap: only around 40 per cent of firms had an official large-language-model subscription, while roughly 90 per cent of their employees reported daily personal use of consumer AI for work. Individuals have crossed the divide; their institutions have not. The technology is demonstrably usable, because people are using it, unsanctioned, at scale, every day. What has failed to keep pace is everything around the technology: the workflows it should sit inside, the controls that should govern it, the roles that should change because of it.
We have been here before
None of this should surprise an economic historian. Erik Brynjolfsson, Daniel Rock and Chad Syverson (2021) described the productivity J-curve: general-purpose technologies require large complementary investments in intangibles, in process redesign, governance, retraining and organisational restructuring, before they deliver measurable productivity. Because those investments sit off the balance sheet, measured productivity dips before it climbs. The current gap between AI capability and AI returns is not an anomaly to be explained away; it is the predicted downward stroke of the J. The crucial caveat is that the upturn is conditional, not automatic. The organisations that make the intangible investments climb out of the curve. The ones that keep buying technology and waiting do not.
Paul David (1990) made the same point about electricity. The dynamo was demonstrably superior to steam by the 1880s, yet factory productivity barely moved for nearly four decades. The gains arrived only when manufacturers stopped bolting electric motors onto plants designed around a central steam shaft and rebuilt the factory itself: unit drive at each workstation, single-storey layouts, work organised around the flow of materials rather than the location of power. The technology was necessary; the redesign was decisive. Substitute tokens for kilowatts and the analogy is almost embarrassingly exact. Most enterprises today are wiring electric motors to the line shaft, automating fragments of processes whose shape was dictated by the constraints the new technology has just removed.
The contemporary numbers put proportions on the history. BCG’s analysis of what drives AI value, published in January 2026, attributes 10 per cent of success in AI transformation to algorithms, 20 per cent to technology and data, and 70 per cent to people and processes. Yet roughly 90 per cent of corporate spend, attention and board reporting concentrates on the technological 10 to 30 per cent. That mismatch is not a rounding error; it is the entire problem, stated as an allocation. McKinsey’s 2025 State of AI research adds the sequencing evidence: organisations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting modelling techniques. Most companies do it in the opposite order, choosing the tool first and discovering the workflow later. The sequencing is the methodology. Get it backwards and you industrialise the existing mess.
The mechanisms of the stall
“Change management” is too vague a name for what goes wrong. The organisational constraint bites through at least six distinct mechanisms, and the research of the past eighteen months lets us name each one.
The first is structural. Joanne Kenny and Marius Oosthuizen, writing in Harvard Business Review in September 2025, documented something counter-intuitive: AI is deepening organisational silos rather than dissolving them. Each function deploys AI within its own boundary, optimising locally, and the organisation as a whole becomes less coordinated. Their banking example is instructive: a risk function whose models flagged customers as too risky while the marketing function’s models targeted the same customers for growth. Aggregate performance can go into reverse even as every individual deployment succeeds on its own terms. California Management Review reached a complementary conclusion the same month: AI changes the economics of cross-functional coordination itself, so that integration roles long held by humans, the demand managers and supply-chain coordinators who form an organisation’s connective tissue, migrate to systems. The job survives; the jobholder changes. An organisation that has not mapped where its integration actually happens will discover the dependency only when it removes it.
The second is governance lag. Joint MIT Sloan and BCG research on the emerging agentic enterprise found 35 per cent of enterprises already deploying agentic AI and a further 44 per cent planning to, and concluded that the technology is spreading faster than leaders can redesign processes, assign decision rights or rethink workforce models. McKinsey’s companion finding carries the same diagnosis in its title: AI is everywhere, the agentic organisation isn’t yet. Who may delegate what to an agent, who verifies, who is accountable when an agent acts, what an approval threshold means when the thing seeking approval can generate a thousand proposals an hour: these are operating-model questions, and in most institutions nobody owns them. Stafford Beer, following W. Ross Ashby’s law of requisite variety (1956), would have recognised the situation immediately. AI injects variety into the organisation that exceeds the requisite variety of its existing control systems. The reflex response is attenuation, restricting models, constraining outputs, limiting use cases. The cybernetically sound response is amplification: new forms of coordination and new local autonomy. Most governance frameworks currently being written are exercises in attenuation, which is why they are simultaneously burdensome and ineffective.
The third is cultural, and it is measurable. BCG’s AI at Work research describes a silicon ceiling: leadership enthusiasm that has not crossed into frontline practice, with only around half of frontline employees regularly using AI tools. INSEAD’s January 2026 analysis describes the texture of the stall: employees nod along in meetings, complete the training, and quietly return to old habits. Herminia Ibarra of London Business School puts the point at its sharpest: the issue is not the technology itself but the organisation’s capacity to learn, which makes AI a leadership challenge rather than a technology one. Culture here is not a soft residual. It is the gating constraint on whether the pilots that do work ever generalise.
The fourth mechanism is identity. Much of what gets labelled resistance to change is nothing of the sort. When a tool can produce in minutes what a professional spent a decade learning to produce, the unspoken question in the room is: if the tool can do this, what exactly am I for? Anthony Giddens (1984) called the underlying need ontological security, the sense that the world is stable enough to act within, maintained largely by routine. AI transformation disrupts precisely the routines through which professionals know themselves. Treating the resulting anxiety as obstruction to be managed, rather than as a rational response to be designed for, is one reason adoption mandates and performance-review metrics for AI usage so consistently fail to move the needle. Adoption is an emergent property of system conditions, not an individual behaviour to be incentivised.
The fifth is economic. The marginal cost of agentic workflows is no longer servers and seats but tokens, and token economics behave like nothing in the existing IT cost model. Version-to-version cost differences of five to twenty times are routine; poorly architected prompt pipelines produce cost growth that is exponential in workflow complexity. The FinOps Foundation’s work on FinOps for AI is explicit that the discipline cannot be owned by IT alone; it requires a cross-functional triad of engineering, finance and the business owner of the use case, with cost-per-unit-of-work mapped to business value. An organisation whose financial controls cannot see, attribute or forecast token consumption has an operating-model gap, not a procurement problem. The same logic applies on the income side: every AI use case needs a profit-and-loss line and a P&L owner, not a technology owner, or AI spend reverts to the classic enterprise-IT pattern of high cost and diffused accountability.
The sixth is the policy estate itself. Approval thresholds, segregation of duties, data-handling and retention rules, model and vendor governance, incident-response runbooks, audit-evidence standards: every one of these documents was written on the assumption that decisions are made, proposed and accelerated only by humans. Each must now be re-examined against an operating model in which agents participate, and either preserved, modified or retired. This is frequently misfiled as regulatory compliance. It is not. It is the systematic redesign of the organisation’s own rulebook, and it outlasts any single regulation.
The regulatory clock has slipped, the argument has not
Regulation was always going to force this conversation in Europe. The current timetable: the EU AI Act’s high-risk regime was due to bind from 2 August 2026. On 7 May 2026 the European Parliament and Council reached provisional agreement on the Digital Omnibus, deferring the obligations for stand-alone Annex III high-risk systems, which include credit scoring and employment tools, to 2 December 2027, and for AI embedded in regulated products to 2 August 2028. The deferral takes legal effect only on formal adoption, expected before August 2026, and the Article 50 transparency obligations remain largely on the original schedule.
It would be a serious misreading to treat the deferral as a reprieve from the operating-model agenda. Three reasons. First, the obligations themselves are unchanged in substance: risk management, data governance, human oversight, logging, post-market monitoring. These are operating-model requirements wearing regulatory clothing, and sixteen additional months merely converts a scramble into a programme for those who start now. Second, the market is already pricing governance regardless of the statutory date. Q1 2026 deal data shows AI Act exposure surfacing in M&A as specific indemnities and withdrawn assets, with documentation quality, not model performance, driving the discounts, and at least one European lender achieving a valuation premium for governance documentation it had maintained since 2024. Third, financial services never had the luxury of waiting in the first place: DORA, the FCA and PRA’s expectations on AI, and the US supervisory letters all reach the same questions about accountability, oversight and explainability through sectoral routes that the Omnibus does not touch.
What following the argument actually requires
If the diagnosis is right, five prescriptions follow. Each one cuts against the standard enterprise AI programme.
Redesign precedes selection. The McKinsey finding is a method, not a statistic: map the end-to-end workflow, decide where decisions move, and only then choose models. An organisation that begins with a vendor evaluation has already made the consequential mistake.
Governance centralises; execution federates. The emerging consensus across the major studies is an operating-model rule: centralised model management, vendor relationships and risk policy; federated ownership of use cases by the businesses that own the P&L lines they touch. The legacy reflex, federated governance with centralised execution, is precisely backwards, and it is what most organisations inherit from previous IT eras.
Conditions, not mandates. Because adoption is emergent, the work is to design the conditions under which people can renegotiate their roles: visible leadership use, multi-tiered education, explicit purpose that is not a euphemism for headcount reduction, and enough routine continuity that the transition does not feel like the dissolution of professional identity. Edgar Schein (1985) would add that culture is what the organisation has learned works; it changes through new experiences of success, not through communication plans.
Compliance becomes a design property. The institutions that will carry regulatory change cheaply are the ones that treat compliance as a designed-in property of the platform and the operating model, evidenced continuously, rather than a periodic audit performed on top. The same instrumentation that satisfies a regulator is the instrumentation that tells management whether the operating model is working.
Capability must be codified or it evaporates. The half-life of an unwritten practice in a fast-moving field is one reorganisation. What compounds is capability bound into reusable, evaluated assets: skills with explicit prompts, evaluation criteria, guardrails and policy shape, owned by the organisation and improved with use. Ikujirō Nonaka (1995) called the underlying move externalisation, the conversion of tacit knowledge into explicit form, and it is the one conversion AI cannot perform for you, because the tacit knowledge lives in your people.
The window
The two clocks generate a strategic timetable. Model capability will keep compounding whether or not any given enterprise is ready; that much is out of everyone’s hands. What is in management’s hands is the slow variable, and the slow variable is decisive precisely because it is slow: an operating model redesigned eighteen months before a competitor’s confers an advantage the competitor cannot buy, because the thing that creates it cannot be procured, only built. The contrarian framing of 2025, that AI is an organisational problem wearing a technological disguise, is visibly becoming the consensus of 2026; the research canon assembled above is the sound of that consensus forming. When it completes, the advantage will not go to the firms with the best models. Everyone will have the best models. It will go to the firms that rebuilt the factory while their competitors were still wiring motors to the line shaft.
The technology will be fine. It is the organisation that needs the work.
Sources: METR, Measuring AI Ability to Complete Long Tasks (2025) and Time Horizon 1.1 (2026); S&P Global Market Intelligence, Voice of the Enterprise: AI & ML Use Cases 2025; MIT NANDA, The GenAI Divide: State of AI in Business 2025, with the Futuriom methodological critique; BCG, Where’s the Value in AI? (2024), What Is Driving AI Value (2026), AI at Work 2025 and AI Transformation Is a Workforce Transformation (2026); McKinsey, The State of AI (2025) and AI Is Everywhere, the Agentic Organization Isn’t Yet; MIT Sloan/BCG, The Emerging Agentic Enterprise (2025); Kenny and Oosthuizen, Don’t Let AI Reinforce Organizational Silos, HBR (2025); California Management Review, The Silo Effect in the AI Age (2025); Brynjolfsson, Rock and Syverson, The Productivity J-Curve, AEJ: Macroeconomics (2021); Paul David, The Dynamo and the Computer (1990); INSEAD Knowledge, AI Transformation Is Not About Tech (2026); Herminia Ibarra, London Business School, Why AI Is a Leadership Challenge; FinOps Foundation, FinOps for AI; Deloitte, The Impact of Agentic AI in Software Engineering; WEF/BearingPoint, AI’s New Dual Workforce Challenge (2025); Council of the EU and European Parliament, Digital Omnibus provisional agreement (7 May 2026).