The Coordination Premium
AI made cognition cheap. The value moved to everything that turns output into outcomes
Something strange is happening inside companies that have adopted AI in earnest. The cost of producing a draft, an analysis, a block of code or a customer reply has fallen to almost nothing. And yet the organisations doing this rarely report a matching fall in cost or rise in output. They report more review, more meetings about which outputs to trust, more infrastructure, more governance, and a curious sense that the bottleneck has moved somewhere they cannot quite see.
This essay is an attempt to say precisely where it moved, and why. I’ve written before about the six ways enterprise AI value leaks away between a working model and a moving P&L; this is the deeper question sitting underneath that observation. Why should all six leaks appear at once, across firms that share almost nothing else? The answer is that they are not six problems. They are six symptoms of a single shift in relative prices, and that shift is the subject here.
The starting point is a distinction that most discussion of enterprise AI glosses over: the difference between cognitive output and organisational outcome. A model can now generate plausible cognitive output, whether text, code, analysis or recommendations, in effectively unlimited quantity at a marginal cost approaching zero. An outcome is different. An outcome is output that has been specified correctly, checked, integrated into a process, and made someone’s responsibility. Output is abundant. Outcomes are not.
Basic economics tells you what happens next. When one input to a production process becomes cheap and abundant, the value in that process migrates to whichever complementary inputs remain scarce. Cheap steel made the value in a railway move to the things steel could not supply: rights of way, scheduling, capital, and the managers who ran them. Cheap cognition does the same thing. It pushes value toward the factors that convert raw output into trusted outcomes, such as objectives, context, verification, coordination, incentives, accountability, and the accumulated organisational capital that holds them together.
I will call the return on those scarce complementary factors the Coordination Premium: the growing share of enterprise value that accrues not to the intelligence itself but to whatever governs it.
A word on that term, because it is easy to misread. Coordination here does not mean meetings, or management overhead, or the soft business of getting people aligned. It means something more specific and more mechanical: the set of systems that define objectives, supply context, verify outputs, route exceptions, and assign accountability. Coordination in this sense is infrastructure. It is as real a factor of production as capital or labour, and like them it can be scarce, invested in, and owned.
The outcome equation
It helps to be concrete about what has actually changed in the cost structure. The cost of a trusted outcome, meaning not a token and not a draft but a piece of work an organisation is willing to act on, breaks down roughly as follows:
Cost of trusted outcome = generation + specification and context + verification + implementation + expected cost of residual error.
For most of the software era, and all of the pre-AI era, the first term dominated. Producing the analysis, writing the code, drafting the contract was the expensive part. Checking and using it was comparatively cheap, because a competent human who produces work also tends to certify it as they go. AI has inverted the weighting. The generation term has collapsed toward zero. Every other term has stayed roughly where it was, and one of them, the expected cost of residual error, has arguably risen, because AI output fails in ways that are harder to see.
This equation is the organising idea of the essay. Almost everything that frustrates enterprise AI adoption is a case of one cheap term, generation, surrounded by four expensive ones that did not move. The model got cheaper. The job did not. The rest of this piece is an examination of those four stubborn terms: why they stayed expensive, which economic forces govern them, and why the firms that learn to drive them down will capture the Coordination Premium while everyone else funds it.
Railroads forced management to become a technology
We have some idea how this plays out, because a version of it has happened before. As George Sivulka recently wrote in a thread on managing AI as if it were a workforce, the instructive case is not a story about a new machine but a story about what a new machine demanded of the organisations that ran it.
Between 1830 and 1840, American railroad mileage grew from a few dozen miles of track to nearly three thousand. This was two orders of magnitude in a decade, and the largest infrastructure build-out the world had then seen. The engineering was the visible achievement. The invisible problem arrived close behind it. On 5 October 1841, two Western Railroad passenger trains met head-on in western Massachusetts, killing a conductor and a passenger and injuring seventeen. The railroad had been running trains in both directions over more than a hundred miles of single track on fixed timetables alone, with no reliable way to know where any given train actually was. Nothing was wrong with the locomotives. What was missing was a system to coordinate them.
The response is the part worth attending to, because it did not come quickly and it did not come from engineering. It took more than a decade. As Alfred Chandler documented in The Visible Hand, it was Daniel McCallum at the New York and Erie Railroad in the mid-1850s who assembled the answer: defined operating divisions, fixed reporting lines, explicit responsibility at each level, and, the genuinely novel part, a discipline of daily and hourly telegraphic reports that let a central office see the state of the whole system in something close to real time. His 1855 organisational chart is generally recognised as the first of its kind.
The usual gloss is that McCallum “invented management.” I believe that is too strong, and the overstatement hides the real lesson. Administration, hierarchy, and delegation are ancient; armies and religious orders had them for millennia. What the railroads did was force management to become an industrial technology, something codified, standardised, teachable, and reproducible across an enterprise dispersed in space and running faster than any individual could personally supervise. Management existed before McCallum. What did not exist was management as a formal system you could install.
And the sequence matters. The engineering shock came first. The organisational technology that made the shock safe and profitable lagged it by more than ten years. During that lag the railroads were not worthless, but their organisational capability visibly trailed their physical capability, and the gap showed up as accidents, waste, and disputes. Only once the coordinating technology caught up did the industry become what it became: from the Civil War until 1900, railroads made up more than half of total US stock market capitalisation, and the majority of trading on the New York exchange. The locomotives were necessary. They were not sufficient. The coordination was the rest.
Enterprise AI is somewhere in that lag today. Firms have laid down enormous new productive capacity, handing every employee something close to unlimited cognitive throughput, while running it on organisational systems built for a world in which cognitive output was expensive and therefore naturally rationed. The throughput is real. The coordinating technology to make it safe and profitable is mostly not built yet. That is not a failure of the models. It is the ordinary shape of an adoption lag, and it is precisely the interval in which the Coordination Premium is set.
Deterministic execution, stochastic generation
To see why the non-generation terms of the outcome equation stayed expensive, it helps to be exact about how AI differs from the software that preceded it. The popular framing is that traditional software was reliable and AI is unreliable. This is wrong, and the wrongness matters.
I have written countless lines of traditional software over my career, and let me assure you that it was never bug-free. It was, however, deterministic. Given the same input and state, it generally produces the same output every time. Its failures are mostly reproducible, which means they can be isolated, diagnosed, and fixed once. A test that passes today passes tomorrow for the same reason. Verification, in that world, is mostly a fixed cost: you check the logic once, and the machine honours it indefinitely. This is why classical software behaved economically like capital, an upfront cost followed by near-zero marginal cost, running unattended.
Generative AI is stochastic. The same prompt can yield different outputs. Quality depends on context that may be present or absent, complete or polluted. Crucially, the output is optimised for plausibility: it is fluent, well-formatted, and confident whether or not it is correct. Its failures are not reproducible in the way a logic bug is. They are contingent, contextual, and often invisible on the surface. This changes the economics of verification from a fixed cost paid once into a variable cost paid per output. You cannot check a stochastic generator once and trust it thereafter, the way you can a deterministic function. Assurance has to be re-established, to some degree, every time.
That single shift, verification moving from fixed to variable and from once to recurring, is the mechanism behind most of the disappointment in enterprise AI. It is also why the naive productivity calculation misleads. Measuring AI output by tokens generated is like measuring a factory by the electricity it draws. Tokens are what you spend. Accepted work is what you make, and acceptance is now the expensive part.
There is a second-order effect worth naming, because it has a clean economic analogue. When generation is free, so is polish. Every output arrives fluent and well-presented regardless of whether it is right. George Akerlof’s classic analysisof markets with hidden quality, the market for lemons, can be invoked loosely here, but the precise and defensible version is narrower. AI does not literally create a market for lemons inside the firm. What it does is destroy polish as a signal. In a world where good work looked polished and rushed work looked rough, surface quality carried real information, and a reader could triage on sight. Once polish is free, it correlates with nothing, and the cheap heuristic that let managers allocate scarce attention stops working. The reader is forced back onto expensive substantive verification for everything. The lost signal is not a metaphor. It is a concrete rise in the verification term of the equation.
The verification-cost hypothesis
If verification is the term that did not fall, then the tasks where AI creates value soonest should be the tasks where verification was cheap to begin with. This turns out to be a surprisingly powerful lens, and it explains the single most conspicuous fact of the current cycle: the breakout commercial use of AI, by a wide margin, is software engineering. The usual explanation is that models happen to be good at code. The sharper one is that code comes with an unusually deep stack of cheap, automatic verification, from compiler to type checker to test suite to runtime, that catches most errors before an expensive human ever looks. Code is not cheap to deploy AI against because it is simple. It is cheap because decades of tooling built an assembly line of graduated, almost-free checks. A legal memo or a clinical assessment has no such layer: the first line of verification is also the most expensive one, a qualified human reading the whole thing with judgment. Same models, same capability, radically different cost of trust, and therefore radically different adoption.
That points to a hypothesis, which I state carefully because it is a claim about the world that could be wrong: the rate at which AI is adopted in a domain is governed less by how well models perform there than by how cheaply their outputs can be verified. It is a hypothesis, not a law, and it reframes the whole enterprise-AI problem. The constraint is not model quality, which improves on its own and is available to everyone, but verification cost, which is specific, buildable, and, as we will see, ownable. A test suite confirms that code matches intent as written, but if the intent was wrong, perfectly verified code ships the wrong product flawlessly. Behind every cheap check sits an expensive question, what should this do?, and that question is a matter of specification and context, the second term in the outcome equation. Even a domain with free verification hands the hard problem back to knowing what you actually want. Which returns us to coordination.
Evals are how you buy down verification cost
If verification cost is the binding constraint, then the central act of enterprise AI is not deploying a model. It is manufacturing cheap verification where none existed, building, for legal or credit or clinical or operational work, the equivalent of the compiler and the test suite that coding already enjoys. In the AI world this machinery goes by the name of the evaluation suite, or eval: an inner layer of automatic checks that encodes what good looks like in a domain that never had a compiler, catches the cheap failures instantly, and reserves scarce expert attention for the hard cases.
The strategic point is what makes evals matter here. Models are converging toward commodity. Any capability a frontier lab offers you this quarter is available to your competitor next quarter on identical terms, and raw model improvements diffuse faster than almost any input in the economy. Evals do not diffuse that way, because the highest-value ones encode firm-specific policy, risk appetite, and judgment. What counts as an acceptable credit decision at your institution, a safe adjustment on your factory line, a compliant contract under your regulator and your risk committee: these are not general facts a model can supply. They are yours. An eval suite that captures them is a private asset a competitor cannot buy off a price list and a model cannot infer from the open internet. So the durable answer to commoditisation is not a better model. It is the accumulated, firm-specific machinery for judging output, the verification capital that turns a generic model into your outcomes rather than anyone’s drafts. Rent the intelligence; own the judgment.
Coordination as capital
Step back and the individual pieces, the specifications, context, evals, decision rights, and institutional memory, resolve into a single category. They are all forms of organisational capital: durable, firm-specific assets that raise the productivity of everything else the firm does, and that sit on no balance sheet.
Economists have long known this category matters, but AI sharpens it to a point. Consider what happens as the model itself trends toward commodity pricing. If the intelligence is available to everyone at similar cost, then it cannot be the source of advantage. By definition, a commodity input confers no edge. Whatever advantage remains has to live in the complementary assets that the model cannot supply and the market cannot price uniformly. Those assets are precisely the coordinating infrastructure: the specifications that tell intelligence what to do, the context that makes its output relevant, the verification that makes it trustworthy, the routing that handles exceptions, the accountability that makes someone answerable for the result.
This is where Chandler and Coase turn out to be describing the same thing from two directions. Chandler showed that the great enterprises of the industrial era won on administrative capability, the visible hand of professional coordination, not merely on technology, which diffused. Coase explained why such capability lives inside firms at all: because some forms of coordination are cheaper to perform under a roof, by direction, than to buy repeatedly in a market. AI presses on exactly this seam. It lowers some transaction costs dramatically. Searching, drafting, summarising, first-pass analysis all get cheaper, which in isolation should push activity out toward markets. But it raises verification and assurance costs, which pushes the opposite way, because trusting a stochastic external output is now expensive in a way trusting a deterministic vendor’s software was not. The net effect is not that AI simply dissolves or simply reinforces firm boundaries. It makes the boundary a live question again, redrawn task by task according to where verification is cheapest to perform, and it puts a rising premium on the internal coordinating capital that Coase’s firms exist to house.
The commodity is the intelligence. The moat is the machinery around it.
Why the office feels busier
If coordination is becoming the scarce factor, the symptom should be visible on the ground, and it is. The most common report from firms deep in AI adoption is not that work has vanished but that it has multiplied. Three mechanisms, all consequences of the same abundance, explain why.
The first is a relative-cost effect that William Baumol described in the 1960s. When productivity races ahead in one activity and stalls in another, the stalled activity’s relative cost rises inexorably. The string quartet still takes four players and the same forty minutes it always did, so as manufacturing productivity soars, live music becomes relatively dear. AI produces an internal version of the same phenomenon. Generation races ahead; specification, verification, and judgment barely move. So the share of the firm’s total cost and attention consumed by those slow activities has to rise, not because coordination got worse, but because everything it sits next to got radically cheaper. The committees, reviewers, and governance boards proliferating around enterprise AI are not simply bureaucratic bloat. They are the relative-cost signature of coordination becoming the binding factor: the cost disease of judgment, telling you where the value went.
The second is a demand effect. When the price of an activity collapses, the quantity demanded rises, and activities that were never economically viable become worth doing. This is a straightforward rebound, and in the limit a Jevons-style expansion in which cheaper cognition means more total cognitive work, not less. A firm that once wrote one tailored proposal per major client now writes one per prospect. A team that shipped quarterly now ships weekly. Much of this new activity is genuine value. Some of it is merely newly affordable, and someone still has to decide which is which, an act of coordination that did not get cheaper. Cheap generation manufactures options faster than the organisation can exercise them.
The third is the constraint on the other side of that decision. Herbert Simon noticed, decades before it was obvious, that a wealth of information produces a poverty of attention: what information consumes is the attention of its recipients. Every cheap draft is an option that some human must still read, judge, and either act on or discard. The supply of that attention is fixed. As generation multiplies the options, the shadow price of the attention needed to adjudicate them climbs, even as the price of producing them falls. This is the deepest reason the busy office feels busier rather than idler after adopting AI: the work of producing candidate outputs was automated, and the work of deciding among them, which is pure coordination, was not.
A note on where this leaves labour, since it is the question everyone asks. The pattern above is not obviously labour-saving in the near term. It is labour-reallocating. The scarce, well-paid work shifts toward specification, verification, and judgment, toward the people who can define what good looks like and confirm when it has been achieved. Whether that is a transient phase of the adoption curve or a durable equilibrium is a genuinely open question, large enough to deserve its own essay rather than a paragraph here. What is clear is narrower and sufficient for the argument: for now, the premium is moving toward coordination and away from raw production.
The operating system for enterprise intelligence
All of this implies a specific and, I think, underpriced prediction about where the next great software category comes from.
The current market is organised around the model and its immediate surroundings: foundation models, inference, the application layer, the tooling to prompt and fine-tune. But if the argument here is right, the model is the commoditising part, and the durable value sits in the coordinating layer that turns a commodity model into trusted firm-specific outcomes. That layer is not yet a product category. It is going to become one.
Concretely, the coordinating layer is the set of systems that do the following jobs. Objective specification: capturing what the organisation actually wants in a form intelligence can act on. Context assembly: supplying the relevant, current, permissioned information around each task. Evaluation: the manufactured verification discussed above. Observability: seeing what agents are actually doing across the estate. Workflow routing: directing outputs, and especially exceptions, to the right check or the right human. Governance and accountability: deciding what is permitted, by whom, with what audit trail and what recourse when it fails. And institutional memory: retaining the decisions, corrections, and hard-won context that would otherwise evaporate between sessions and staff.
Assembled, those functions are something recognisable: an operating system for enterprise intelligence, the layer that schedules a scarce resource, enforces permissions, mediates between raw capability and safe use, and remembers. The generation of engines that produce intelligence is largely built. The generation that governs it is not, and that is the larger prize, because it is where firm-specific advantage is defensible rather than rented.
This is also the honest reading of a signal the market has already sent. The organisations best placed to know where the hard problem lives, the model builders themselves, are pouring resources into deployment, forward-deployed engineering, and enterprise services. They are behaving as though the model were no longer the constraint. The railroad analogy tells you what they are really building: not faster engines, but the signalling.
The signalling, not the engine
Return, one last time, to McCallum. His enduring contribution was not the org chart that gets reproduced in textbooks. It was the reporting system beneath it, the discipline of daily and hourly telegraphic information that let a single office see the whole railroad at once, and therefore run trains fast without running them into each other. The chart was the visible artefact. The signalling was the technology. Speed was survivable only because the coordinating system caught up with the engines.
Enterprise AI is running fast trains on new track with, for the most part, no despatcher and no signals. The engines are magnificent and increasingly cheap. What is scarce is everything that lets an organisation see what its intelligence is doing, judge whether the output is right, route the exceptions, and stand behind the result. That scarcity is not a temporary inconvenience on the way to abundance. It is where the abundance gets converted into value, or fails to, and whoever owns the conversion owns the margin.
Which is the whole argument, compressed. The first generation of AI companies built engines for producing intelligence. The next generation will build the institutions that let organisations govern it: the specification, verification, routing, and memory that turn abundant cognition into trusted outcomes.
Intelligence is becoming abundant. Coordination is becoming scarce, and therefore valuable. The Coordination Premium is the return on closing the gap between the two, and it is where the next decade’s profits will accumulate.