The Paradoxes of AI
Nine contradictions in enterprise AI with guidance for leaders

Synopsis
Since 2023, hundreds of billions have been spent on enterprise AI. The returns are harder to find. AI has made intelligence dramatically cheaper, faster, and more widely available. Yet more intelligence has not reliably produced more value. This paper introduces nine new paradoxes from my research and practice. Some extend familiar economic arguments; others arise from the peculiar conditions of abundant machine intelligence.
In the paper’s words
- Racecar Paradox: the F1 analogy“Put your author in an F1 Ferrari and it will be neither safe nor fast. Yet enterprises keep upgrading the engine while giving too little attention to the drivers.”
- Cheap Intelligence Paradox: the 10/90 rule“With today’s frontier models so powerful, the rule of thumb I offer leaders is value is about 10% model power, and about 90% ‘the rest.’”
- Plug-and-Play Paradox: the part that’s yours“Nobody can sell you the finished workflow, because that idiosyncratic workflow is the part that’s yours.”
- Endowment Paradox: the verification tax“The team completes 3.43 units an hour, not 4.8: a threefold improvement in individual production becomes only a 2.14-fold improvement in system output.”
- Outsider Paradox: the plumber“Following me was a plumber who had dropped out of high school at fifteen. After a stint in the military, he had opened an HVAC business that, in under eighteen months, had grown to 44 technicians. He had built his entire operational infrastructure in Claude Code.”
The nine paradoxes
The paper draws on intellectual history (Coase, Jevons, Polanyi, Baumol, Solow, Paul David and others), on social science research, and on emerging evidence. Its paradoxes are signals from the frontier: propositions to test against the work organisations are doing now. None is yet definitive, but leaders do not have the luxury of awaiting clarity. So beneath each paradox sits a strategic heuristic that a chief financial officer, a chief information officer or a chief people officer can act upon.
- The Racecar Paradox: the better the machine, the more the driver matters.
- The Illusion Paradox: the first summit is not the mountain.
- The Cheap Intelligence Paradox: cognition got cheap. Getting value from it did not.
- The Plug-and-Play Paradox: the product is standard. The work is particular.
- The Participation Paradox: participation is part of the build, not an afterthought.
- The Temporal Shear Paradox: keeping up can leave you further behind.
- The Endowment Paradox: the richest clients have the least to gain.
- The Outsider Paradox: the next significant contribution may come from someone without the expected credentials.
- The Big Science Paradox: universities helped create the field; they now rent access to its frontier.
1. The Racecar Paradox
The better the machine, the more the driver matters.
Enterprises are upgrading the engine faster than they are redesigning the system around it and developing the people to drive it. AI-enabled expertise now includes verifying outputs, supervising delegated work and recognising failure, and those abilities need practice even among people with deep knowledge of their field. The paper cites the University of Melbourne and KPMG’s 2025 global study of about 32,000 employees in 47 countries: fewer than half reported receiving AI training, and two-thirds reported relying on AI output without evaluating it. A firm can buy models, licences and agentic tools in a quarter. It cannot acquire sound judgement, redesigned work and supervision capability at the same speed, and the larger cost is confident but unverified work.
What to do. Do not approve model and infrastructure spend without a matching commitment to systems redesign and role-specific human capability. Rewrite the definition of expertise used in hiring and promotion: in an AI-enabled organisation, expertise includes knowing how to supervise the machine. Invest in drivers; invest in tracks.
2. The Illusion Paradox
The first summit is not the mountain.
Early, real gains can make the larger frontier hard to see. Someone who gets quicker drafts and useful advice from an AI assistant can conclude that they “know AI”. What remains invisible is how much further skilled use can go through better context, verification, workflow design and delegation. Self-report surveys supply useful signals about trust, enthusiasm and friction, but people saying they are willing to experiment does not mean that they do. Leaders can mistake early fluency for mature capability, or enthusiasm for return. This is why the paper insists that leaders build: until they have experienced both the power of the technology and the friction of reorchestrating even simple workflows, they are asking others to do something they understand only in the abstract.
What to do. Don’t rely exclusively on self-reports of use and value. Use self-report to find enthusiasm, friction and emerging practice, then pair it with behavioural measures that show what changed in the work. Give people ways to see the next mountain: shared practice, examples of excellent use, peer comparison and operational evidence, plus a sandbox to experiment in.
3. The Cheap Intelligence Paradox
Cognition got cheap. Getting value from it did not.
The price of a unit of machine intelligence has collapsed. The paper cites Stanford’s 2025 AI Index, which shows GPT-3.5-level intelligence falling from twenty dollars to seven cents per million tokens between November 2022 and October 2024. Lower prices make more uses economic, so total spending can still rise: the Jevons effect. And a cheaper token does not bring an equivalent reduction in the cost of redesigning a process, developing a workforce or clarifying who makes a decision. AI and human capability are complements, and the organisation around the model can remain the limiting factor. Brynjolfsson, Rock and Syverson’s productivity J-curve explains why weak early returns can reflect investment still taking effect, as well as poor deployment; Paul David’s account of electrification shows why cheaper power alone was insufficient. The false economy is buying what has become cheap and visible while underinvesting in what still limits value.
What to do. Spend on what limits value: systems redesign, human capability and leadership. Approve an AI proposal when it changes the system in which cheaper thinking becomes useful work.
4. The Plug-and-Play Paradox
The product is standard. The work is particular.
The same model can serve thousands of organisations, but a dependable workflow must fit the organisation using it. Two companies of the same size, in the same industry and country, differ in technology stacks, staffing, objectives, decision rights and exceptions: who can approve a discount, which record is authoritative. Packaged tools can be immediately useful to individuals. An agentic process that acts across company systems, handles exceptions, respects permissions and leaves someone accountable is a different undertaking, and reliable operation requires design and testing in situ. The paper cites McKinsey’s 2026 State of AI survey, in which nearly three-quarters of AI high performers (respondents attributing at least 5 per cent of EBIT to AI) reported fundamentally redesigning workflows, against one-quarter of other respondents. The survey cannot establish causation, but it gives weight to the distinction between buying access and changing how work gets done. A licence buys access; omit the costs of integration, local design, testing and supervision and they return as repeated pilots, consultant dependency and manual workarounds.
What to do. Buy useful software and reusable components, but budget explicitly and sufficiently for making workflows work in your organisation. Choose a consequential workflow and involve the people who perform it, including those who receive its outputs. Define exceptions, permissions, quality measures and responsibility before extending the system. You can buy the components. You still have to design how the work gets done.
5. The Participation Paradox
Participation is part of the build, not an afterthought.
Involving users in technology strategy and workflow design has long been recommended, with roots in organisation development and sociotechnical research, and has long been treated as optional. A workshop held, a town hall delivered, a communications plan completed: participation theatre could tick the “user engagement” box. A largely prescribed SAP or Salesforce process could be mandated even when users disliked it. When a team is redesigning judgement-heavy work around AI, the important questions remain open: who decides, which exceptions matter, what needs checking and how the work reaches the next department. Practitioners bring the knowledge that answers them, so their involvement becomes part of creating the system they are expected to adopt. When participation is postponed, the organisation pays for design decisions before discovering whether they fit the work.
What to do. Build Adoption Pods™ around specific workflows: the practitioners doing the work, the person accountable for its outcome, a forward-deployed engineer and a forward-deployed change practitioner. Give the pod time and authority to design, test and improve the process, and assess it through improvements in the work and sustained, appropriate use. Participation belongs in the delivery budget from the outset.
6. The Temporal Shear Paradox
Keeping up can leave you further behind.
The work needed to make today’s capability useful can become an obstacle to using tomorrow’s. The paper describes five clocks: model releases measured in weeks, agentic advances in months, budgeting in quarters, infrastructure replacement in years, and institutional change sometimes in decades. Decisions made on one clock must work with commitments made on another. A firm may integrate a capability, validate its behaviour and train people to use it, only to find that a new model makes some of that work unnecessary; the paper cites Anthropic’s engineers describing a workaround that a stronger model turned into overhead. The term comes from the AI Leadership Delta™ framework: connected organisational plates moving at different velocities. The organisation can pay both for moving too quickly and for committing too rigidly.
What to do. Give different decisions appropriate lifetimes. Keep technical components replaceable, and test changes against the work they must perform. Preserve useful operating practices long enough for people to become capable, while revisiting assumptions that newer models may have overtaken. Review subscriptions and workarounds as deliberately as new purchases.
7. The Endowment Paradox
The richest clients have the least to gain.
A modest AI budget can add more capability to a resource-poor founder than a large one adds to an already well-equipped firm. Supplying eighty percent of a missing capability can be worth vastly more than adding another source of one already abundant. But faster individuals do not make a faster system. In the paper’s two-person illustration, tripling each person’s production while serious human checking still takes fifteen minutes per hour of output lifts completed work from 1.6 to 3.43 units an hour: a 2.14-fold gain, not threefold. Checking grows from a fifth of each person’s time to almost half. Frontier AI does not remove the bottleneck; it shifts it elsewhere in the production system.
What to do. Identify where AI adds a capability you lack and where it improves one you already have. Give a small team responsibility for an outcome from beginning to end, with access to the people and information it needs. Measure the full cost and value of the work, including review and coordination. The question for a large firm is how much of the freedom and immediacy enjoyed by a capable founder it can reproduce safely inside its own operations.
8. The Outsider Paradox
The next significant contribution may come from someone without the expected credentials.
Demis Hassabis, a chess prodigy and games developer who trained in neuroscience, was jointly awarded half of the 2024 Nobel Prize in Chemistry, with John Jumper, for using AI to predict a protein’s three-dimensional structure from its amino-acid sequence, a fifty-year-old problem. The paper’s more everyday example is the HVAC founder above, who had never studied algebra, never mind coding, yet built the operations layer of a 44-person business with AI. This is another dimension of People-First AI™: opportunity at the bottom of the economic pyramid, where fewer good ideas fail simply because their authors lack the expected route into a field. When credentials become a substitute for examining the work, the loss is businesses never attempted, ideas never tested and capable people never considered.
What to do. Create routes through which people can demonstrate capability: working prototypes, practical assessments and research contributions that others can examine and test. Pair access to AI with domain mentorship and rigorous verification. Keep the standards required for safe and dependable work, while widening the routes through which people can meet them.
9. The Big Science Paradox
Universities helped create the field; they now rent access to its frontier.
Frontier AI has outgrown university-scale resources, and private firms control much of the infrastructure on which academic researchers now depend. The paper cites Ahmed, Wahed and Thompson in Science (2023) on industry’s growing command of compute, data and talent, and Stanford’s 2025 AI Index: industry produced nearly 90 per cent of notable models in 2024, while academia remained the leading institutional producer of highly cited research. Universities have become customers of the firms driving the frontier, and the companies supplying research infrastructure may also own the models under investigation. Control of research infrastructure influences whose questions receive sustained attention and resources. For an enterprise buyer, the related risk is treating a provider’s roadmap as sufficient independent advice. A post-script draws the national implication: a country need not produce the leading model to benefit substantially from AI.
What to do. Assess technical capability and independent judgement separately. Examine a research partner’s access to relevant systems and its commercial dependencies. For enterprise decisions, combine hands-on technical knowledge with scrutiny that is independent of the supplier.
What to do Monday
The paper closes with nine rules for the next investment decision. Before approving the next AI spend, insist on these nine answers. A proposal that cannot answer them is not ready.
- No capability plan, no model spend. Fund the people, controls, and operating system needed to supervise the machine.
- Measure outcomes, not enthusiasm. Track cycle time, quality, rework, customer outcomes, and unit economics.
- Fund the bottleneck, not the model. Cheap intelligence is useless inside an unchanged operating system.
- No workflow owner, no automation. Define responsibility, exceptions, testing, and maintenance before buying the licence.
- Put practitioners inside the build. Give the people who know the real work authority in its design.
- Make components replaceable. Do not hard-wire today’s model into a system that will outlive it.
- Measure the whole outcome. Faster tasks do not count if work still sits in queues or creates a verification tax.
- Test capability, not credentials. Look for demonstrated work, practical judgement, and rigorous verification.
- Never let a supplier grade its own system. Separate technical access from independent judgement before placing the bet.
In the full paper
The 28-page PDF sets the nine paradoxes against ten pre-AI ideas that still explain why technical progress and enterprise value can move in opposite directions, among them the Jevons Paradox, Polanyi’s Paradox, Baumol’s cost disease, the Abilene Paradox, Goodhart’s Law, the Collingridge Dilemma, the ironies of automation and the Solow Paradox. It carries the full argument for each paradox, its costs, the worked examples and illustrations, and the selected sources.