On a narrow class of work, one person with a good AI tool can now outrun an organisation a hundred times its size. I hold this view loosely. It is an inference from a single experience, not a law, and large organisations still win decisively on most things. But it came from something concrete. I recently built a browser-native AI platform myself, with AI assistance and no prior coding experience. I cannot unsee what it implied. If it is even partly right, it changes where leaders should look. The effort a large organisation pours into building the thing may no longer be its binding constraint. For anyone setting out to become "AI-first", the constraints and risks that now deserve that energy lie elsewhere.
The same tool, very different returns
Give two organisations the same tool, the same model, the same licence, and you get very different returns. If the tool were the differentiator, that would not happen. Yet it happens constantly, so the differentiator sits somewhere else. None of that is new: it was true of every technology wave before this one. The tool was never the thing that mattered. The orchestration around it always was — how a company arranges its processes, decisions and value chain to get the best from it. Firms that treated technical implementation as the finish line have always trailed the ones that built the organisational context around it. That difference has always shown up as ROI: buy the same system as a rival, reshape the business around it while they merely install it, and you earn several times the return they do. What is new with AI is the size of that spread. Good orchestration compounds the return faster than any earlier tool allowed. Bad orchestration does more than waste the spend. It amplifies the weakness already there. And it often lodges the damage somewhere in the value chain no one is watching, so a negative return can run on unnoticed until it is large.
The usual metaphor points the wrong way. AI is not an escalator that lifts everyone the same distance for the price of a ticket. It is an amplifier. It multiplies whatever signal you feed it, and it does not care about the quality of that signal. Feed it clear judgment and it compounds clear judgment. Feed it a hesitant, committee-averaged decision and it compounds that instead. The output is plausible, finished-looking and confidently wrong, produced at a speed the old process would have filtered out. Some of that filtering was mere organisational friction. But some of it was experienced people sitting at each stage of the work. They noticed when something looked off and stopped it before it moved on. As AI runs more of the process end to end, those people sit further from it. That quiet, distributed checking thins out just where it used to earn its keep — at the middle of the chain, before a mistake reaches the end. In a weak system racing to be AI-first, the failure is not that nothing happens. It is that the wrong thing happens faster, and looks more finished while it does.
Two mechanisms, both about the operating model
Why does the same tool amplify so differently across organisations? Two mechanisms, both matters of operating model rather than technology. Neither says the people inside large organisations have worse judgment. They do not. Structure decides how much of that judgment survives the journey to the tool. That is the thing a leader can actually change.
The first mechanism is a shift in where the bottleneck sits. For decades the scarce resource in knowledge work was production capacity: the ability to build the thing. That scarcity acted as a throttle. You could not ship much bad software. It was slow and costly to produce, and the cost filtered out a great deal of poor work before it saw daylight. AI removes that throttle. Once it is gone, quality is gated by whatever is left: the judgment and the incentives of whoever is steering. The bottleneck moves from production capacity to judgment capacity. Most organisations spent years optimising the first, through headcount, process, delivery pipelines and system integrations. Very few deliberately built the second: taste, domain-embedded decision-makers, empowered contrarian voices, verification loops, accountability, and incentives that reward being correct rather than merely shipping.
The second mechanism is older. Ronald Coase explained why firms grow large. For complex production, coordinating work inside a firm was cheaper than transacting for it in the open market, and size was justified by the coordination it bought. AI collapses the cost of production towards zero for a widening class of knowledge work. It does nothing for the cost of agreeing. Once production becomes cheap, the firm's remaining costs start to dominate: coordination, alignment, sign-off, consensus, internal politics. Overhead that was once an asset becomes a tax. That is why a single motivated operator can now outpace a large enterprise on build-and-ship work. Not because the individual is more capable, but because they carry no coordination cost, so their judgment reaches the tool undiluted.
How a weak operating model launders judgment out
Large organisations do not lack judgment; they are full of it. The risk is that a weak structure launders it out before it reaches the tool. A decision passes through risk committees, competing incentives, diffused accountability, and the ordinary wish not to be the one who authorised the mistake. By the time it steers the AI, it can be a committee average rather than the sharp, owned call that produces good work. The amplifier then works on a degraded input, and the result will be worse — not because the people are worse than the solo operator, but because the aggregation is. None of this dilution is new; it happened long before AI. What AI changes is speed and reach: it makes the diluted decision act faster, and in an automated workflow its effects can surface in a part of the value chain no one was watching.
Where the soft factors turn mechanical
The factors leaders already discuss stop being soft here and become mechanical. Risk appetite, leadership style, culture: each has a specific effect.
Risk appetite gates conversion. AI produces plausible output quickly. But you only realise the value by shipping it, accepting the risk of being wrong, and correcting. A low-risk-appetite organisation spends all the time AI saved sitting in approval queues. The speed advantage evaporates in governance, and the investment shows up as cost with no matching output.
Leadership style sets how many amplifier points are well steered. Command and control centralises decisions, so the binding constraint becomes the decision throughput of a few people at the top. AI cannot relieve that, because the constraint was never production. Distributed authority puts judgment next to the work. It creates many points at which the amplifier can be pointed well, and in parallel.
Culture decides whether the verification loop runs at all. The loop that catches confident-but-wrong output depends on someone being able to say "not good enough, again" without career cost. In a blame culture that sentence goes unsaid. The plausible output ships, and the amplifier's speed becomes a liability.
The damage then outlives the incident. An amplified failure, shipped fast and punished, writes itself into institutional memory. A blame culture usually mislearns the lesson. It concludes that AI is dangerous, that taking risks with it burned the company, when the real fault was shipping without verifying. That mis-attribution is corrosive, because it hardens into blanket caution. A little caution after a failure is rational learning. Over-generalised caution, which blames the risk rather than the broken loop, is not. It leaves the organisation once bitten and twice shy, forgoing realistic upside for years. And it closes a doom loop: blame teaches the wrong lesson, the wrong lesson breeds risk-aversion, and risk-aversion breeds less experimentation. Which means less of the very upside the amplifier promised.
Build the judgment layer, not more pilots
So what should a leader do differently? Less than the vendor cycle implies, and harder. The reflex is to buy more tools and run more pilots. That is investment in production capacity, which AI has already made abundant. The higher-return move is to build the judgment layer the tool cannot supply. Put domain expertise as close to the AI as possible, so the steering signal is strong before it is amplified. Build cheap, fast verification loops, because the failure mode is confident-but-wrong, not slow. Move decision rights towards the work rather than up the hierarchy, so the amplifier has many good points to fire from. Align incentives to correctness and shipped outcomes — not to activity, sign-off, tokens consumed, or the appearance of diligence. None of these are procurement decisions; all of them are operating-model decisions.
Cognitive diversity: the error-correction layer
There is one more, and in an AI world it may matter most: hire deliberately for cognitive diversity, and protect informed dissent. AI output has a specific and dangerous shape. It is fluent, confident and homogenising. It regresses towards the consensus of its training data, and it is persuasive. Its most common failure is output that is plausible and wrong, sliding through because it matches the assumptions and biases already in the room. The reliable defence is a mind that does not share those assumptions and biases: someone with different scars, willing to say "that is plausible, and I still do not believe it". Homogeneous teams share blind spots, and an amplifier fires straight into a shared blind spot without resistance. So cognitive diversity is not an HR nicety here. It is the error-correction mechanism for a homogenising machine. The value of a well-placed contrarian rises as the tools improve, rather than falling. The hiring corollary is clear, if unfashionable. Tool proficiency is commoditising by the month, so do not optimise for it. The scarce, decisive input is the varied experience that produces nuanced, contrarian, brave judgment, and the safety that lets it be spoken. Dissent has to be informed and grounded in real experience, or it is just noise. And diversity without the safety to voice it is inert, which is why this point and the culture point are really one point wearing two hats.
If there is a single message to take away, it is that AI amplifies the system it runs inside. It multiplies the judgment and the operating model already there. Given the same tool, the return is set by the context that supplies or starves the judgment which the tool cannot provide. Fix that first and the tool compounds it. Leave it, and the tool compounds its absence.
Further reading and intellectual debts
This is an opinion essay, not a research paper, and it quotes no source directly. It leans on the following ideas, whose originators are worth crediting.
- Ronald H. Coase, "The Nature of the Firm" (Economica, 1937), on coordination and transaction costs, and why firms exist.
- Frederick P. Brooks, "The Mythical Man-Month" (1975), on the communication overhead of collaboration.
- Melvin E. Conway, "How Do Committees Invent?" (1968), the origin of Conway's Law, that a system reflects the structure of the organisation that builds it.
- Eliyahu M. Goldratt, "The Goal" (1984), on the Theory of Constraints and shifting bottlenecks.
- Barbara Levitt and James G. March, "Organizational Learning" (Annual Review of Sociology, 1988), on superstitious learning and competency traps.
- Daniel Kahneman, "Thinking, Fast and Slow" (2011), on loss aversion, availability and negativity bias.
- Irving L. Janis, "Groupthink" (1982), on shared blind spots in cohesive groups.
- Scott E. Page, "The Difference" (2007) and "The Diversity Bonus" (2017), on cognitive diversity in problem-solving.
- James Surowiecki, "The Wisdom of Crowds" (2004), on the value of diverse, independent judgment.
- Amy C. Edmondson, "The Fearless Organization" (2019), on psychological safety and voice.
- Erik Brynjolfsson, "The Turing Trap" (2022), and David H. Autor, "Why Are There Still So Many Jobs?" (2015), on augmentation versus automation.
Ari Das-Purkayastha advises organisations on delivering and managing technology-led transformation. The build described at the top of this essay is documented separately, as a first-person account with its limitations stated.