The Marginal Cost of AI Is Not Zero
Edition 22 — Why boards should stop approving AI investment on assumptions borrowed from software, where the next unit was free, and scale improved the margin on its own.
Somewhere in the pack for your next board meeting, there is probably an AI business case. It will have a build cost, a licence cost, an integration cost, and an adoption target, and management will have done honest work on all four. What it almost certainly will not have is a line showing what one unit of work will cost once the system is running at full volume. Nobody removed that line. It was never there, because for thirty years nobody needed it. The software era trained every director now serving on a board to treat marginal cost as a rounding error: the expensive part was building the thing, and once built, the next user, the next transaction, the next report cost close to nothing. Under that assumption, adoption was a perfectly reasonable proxy for value, because every additional unit of usage spread a fixed cost across a larger base. The assumption was correct for two decades, which is exactly why it stopped being examined. It has now quietly become wrong, and the business cases have not noticed.
The error is easy to locate. It sits in the spreadsheet, in the implicit claim that cost is fixed and value scales. Agentic AI inverts the shape. These systems consume metered compute every time they perform a unit of work, and the meter does not care that the software is already built. A board approving an agentic deployment on software-era logic is applying the economics of one kind of asset to a different kind altogether, rather as if it had approved a power station on the cost model of a patent.
The obvious objection is that software’s marginal cost was never truly zero. Per-seat licences rose with every hire, and cloud bills have scaled with traffic for fifteen years, so a finance director could reasonably say that variable cost is nothing new. Both points are true, and neither rescues the old logic. A seat cost the same whether its occupant ran ten queries a day or ten thousand; the licence scaled with headcount, while usage inside it was free at the margin, which is why nobody built a business case around it. Cloud infrastructure scaled with volume, but sub-linearly, with unit costs falling year on year, and it rarely amounted to more than a sliver of the value the software delivered. What has changed is not the existence of variable cost but its behaviour: it now attaches to each unit of work instead of each person or server, it grows with the task’s ambition where the old costs shrank with experience, and it is large enough to be a first-order line in the P&L of the service it powers. The old variable costs were real but never load-bearing. This one carries the weight of the case.
What the meter is actually reading
The mechanics matter because they determine the slope of the cost curve. An agentic process reasons its way through each task, and reasoning is billed by consumption. A simple lookup is cheap. A task that requires the system to plan, call other systems, check its own output, and retry is not, and the more capable the deployment becomes, the more of that expensive reasoning it does. EY estimates that a straightforward automated customer interaction that cost a few cents in 2023 costs around thirty times more in its 2026 agentic form, because the agent does far more work per interaction. The direction of that comparison matters more than its precision: cost per unit rises with ambition, and ambition is the whole point of deploying agents rather than scripts.
The consequence boards have not yet priced is that success is what generates the bill. The pilot that fails is cheap. The deployment that works, scales across the function, and gets extended into adjacent processes is the one whose costs compound month after month, and those costs arrive as operating expenditure long after the approval that created them has been forgotten. This inverts the risk profile directors are accustomed to, where the danger in a technology investment was paying for something nobody used. The danger now includes paying more than anyone expected for something everybody uses.
The margin consequences are no longer speculative. ICONIQ’s survey of some 300 software executives puts expected gross margins on AI products at roughly 52 per cent for 2026, a full tier below the 75 to 85 per cent that defined mature SaaS, and the buyer side is meeting the same arithmetic. Uber exhausted its entire 2026 AI budget by April, four months into the year, after rolling out AI coding tools to around 5,000 engineers and ranking teams on internal leaderboards by usage. The company measured adoption, rewarded adoption, and got adoption; the bill followed. Its president has since conceded publicly that the link between all that usage and features customers value “is not there yet”. The technology did not fail; the engineers found the tools so useful they would not stop. The governance did. A consumption-priced asset was managed with a metric built for a per-seat world. A board that has only ever asked whether the AI is being adopted has no way to see this coming until it surfaces in the operating cost line, at which point the conversation is about an overrun, not a decision.
Scale stopped doing the work
The deeper problem is strategic rather than budgetary. In the software era, scale was self-reinforcing. More users meant the same cost base spread thinner, which meant better margins, which funded the next round of investment, which attracted more users. Much of the received wisdom about technology strategy, from platform economics to winner-takes-most market structures, rests on that loop. Metered economics breaks it. When each unit of work carries a real cost, volume no longer improves the unit economics by itself. Consumption discounts exist, and negotiating them well matters, but they flatten the curve without restoring a near-zero floor. The loop that made scale compound has lost its engine.
This changes what counts as an advantage. Being the largest adopter of AI in your sector confers no cost benefit on its own; it may simply mean you have the sector’s largest compute bill. The advantage instead accrues to whoever achieves the lowest cost per unit of work at acceptable quality, which depends on process design, model selection, and engineering discipline rather than enthusiasm or budget. That is an efficiency race, and efficiency races reward different capabilities from adoption races. They reward the unglamorous work of routing simple tasks to cheap models, reserving expensive reasoning for the cases that justify it, and measuring relentlessly. There is a further wrinkle: the price per unit is not fully within the company’s control, because the suppliers setting it have repriced repeatedly and will again. That dependency deserves its own treatment, and a later edition will cover it. For now it is enough to note that a cost curve you do not control is a strange thing to build a strategy on without at least knowing its slope.
The question that belongs in the approval pack
The remedy is a discipline rather than a framework, and it fits in one question: what does one unit of work cost, and what recovers it? The unit of work is whatever the process produces: a resolved claim, a reviewed contract, a closed case, a customer interaction brought to completion. Three changes to board practice follow from taking the question seriously. The first is an approval criterion. No AI investment is approved without a forecast cost per unit of work at target volume, alongside a stated view on whether price, savings, or margin recovers it. This is not an onerous demand; it is the same arithmetic the board would require before approving a factory, a call centre, or a logistics contract, and management teams that cannot produce it have revealed something worth knowing.
The second is a reporting demand. For services and processes that AI now materially powers, the board should see gross margin, or unit cost against unit value, as a standing management report. The purpose is to watch the direction of travel: a healthy deployment shows unit cost falling as the team learns to route work efficiently, while a deteriorating one shows cost rising as ambition outruns discipline. The third is a kill rule. An initiative whose unit cost is not on a credible path to below its unit value gets stopped, no matter how impressive its adoption numbers. Together, the three tests retire a habit the software era made harmless, and the AI era makes expensive: approval by adoption metric. Usage was a defensible proxy for value when the next unit was free. It is now a number that can grow in perfect step with a loss.
The discipline is older than the technology
None of this asks boards to learn anything new. Directors have applied unit economics to physical operations for as long as boards have existed; no one approves a plant without a cost per tonne, or an airline route without a cost per seat. The discipline lapsed for technology only because software genuinely was different, and a generation of directors reasonably internalised the exception. What has changed is that the exception has ended while the habit persists. Asking for the denominator is not hostility to AI investment. It is what makes the strong cases approvable with confidence and the weak ones visible before they scale, and it moves the board conversation from whether the organisation is adopting AI to whether the adoption is worth having. One caution for the next meeting: a unit cost forecast is only as good as the life of the asset behind it, and the assumed life of AI assets deserves more scrutiny than it is getting. That is where this series goes next.
If this is the kind of board-level analysis you want more of, subscribe to AI in the Boardroom. I write for directors, executives, and advisers who need to see through AI’s adoption theatre to its commercial substance: the costs, margins, and decisions that determine whether the investment was worth making. Future editions will keep working through the economics that boards are being asked to approve on faith.