The assumption nobody wrote down

Almost every AI business case written in the last few years rests on a line that never appears in the document. It is the belief that AI inference cost per unit of useful work will keep falling — that whatever a transaction costs to run today, it will cost meaningfully less by the time the volume that justifies it actually arrives. Nobody argues for this in the deck. It isn't a claim. It's the ground the claim stands on.

And it has been solid ground. Cost per unit of capability has fallen steeply, and it has fallen more than once. Models got more efficient at the same quality. Competition between providers pushed in the same direction. Hardware improved, and the software wrapped around it improved faster. Three separate forces pointed the same way for long enough that treating decline as the default was not optimism — it was the empirically supported position. In the programmes we've watched up close, the teams who priced on continued decline built things the teams who insisted on closing the case at today's prices never got to build. They were right, and they were right repeatedly. We would rarely have told them otherwise.

So this is not an argument that the assumption is wrong. It is an argument about where it lives. The assumption is load-bearing, and it is undocumented — and those two properties together are what create the exposure.

A premise nobody owns

Go looking for it in an actual business case and you will not find a line that says we assume unit costs decline. You will find it dissolved into other numbers. A per-transaction figure, measured in a pilot, extrapolated across a year. A sentence that says the workflow becomes viable at scale, where scale is a volume not yet reached and viability quietly depends on a rate not yet paid. A margin that works in the model and would not work at the price you are paying this morning. The assumption is everywhere in the arithmetic and nowhere in the prose.

That matters because of what documentation actually does. A premise written down as a premise acquires two things: a person whose job it is to watch it, and a stated consequence if it fails. An undocumented premise acquires neither. Nobody is monitoring the input, because nobody has named it as an input. Nobody has run the case at a materially different unit price, because there is nothing in the document that says the price is a variable. This is the same failure we described in strategy documents that describe a destination rather than a way of deciding — the reasoning that produced the conclusion is not preserved anywhere, so when the conditions move, there is no way to re-derive anything. Undocumented assumptions do not get monitored. They get discovered.

Cheap work is thin work

The asymmetry is what makes this more than a bookkeeping complaint. Think about which workflows falling costs actually unlocked. Not the small number of high-value judgements — those were affordable at almost any plausible price, because the value per decision dwarfed the compute behind it. What declining costs unlocked was the other kind: high-volume, thin-margin-per-transaction work. Document triage at industrial scale. Every inbound message classified rather than a sample. Enrichment applied to the whole file instead of the top decile. These are precisely the workflows where a change in unit price lands hardest, because there is almost nothing between the cost and the benefit to absorb it.

Now notice what you did operationally to make those workflows pay. You didn't run the new path alongside the old one. You retired the manual process, because keeping it was the expense you were removing. You reassigned the team. You let the vendor contract lapse, and the institutional knowledge of how the work used to be done left with the people who used to do it. That was correct — the savings were the point. But it means cost exposure and reversibility now move in opposite directions. The workflows most sensitive to unit price are the ones you have least ability to step back from. That is the combination that turns a pricing question into an operational one, and it is invisible on any spreadsheet that treats the two as separate.

We won't predict AI inference cost

Here is where we decline to be useful in the way you might want. We do not know whether costs continue to fall, and we think anyone expressing confidence in either direction is overreaching. The efficiency curve has real momentum behind it. Demand for compute is growing in ways that have not resolved. Energy and infrastructure constraints are genuine and their timing is not knowable from where anyone is standing. Competitive dynamics can compress prices and can also, once positions consolidate, stop compressing them. These point different ways, and we have not seen an argument that reconciles them which isn't mostly temperament wearing the clothes of analysis.

The point is not that prices will rise. The point is narrower and harder to dismiss: you have taken an unhedged position on a variable you did not know you were positioned on. If someone had asked you last quarter to name your major exposures, this would not have been on the list. It is on the list. It has been the whole time.

Levers you build early

Two capabilities determine whether a change in unit economics is an adjustment or a crisis, and both have to exist before the pressure arrives.

The first is the ability to move work between model tiers. An organisation that has already done the work of finding the cheapest model that is right enough for each task has a lever it can pull — it knows which workloads have headroom, which are running on more capability than they need, and what quality actually costs at each rung. An organisation that standardised on one tier because standardising was simpler has no lever at all. Building that flexibility takes months of evaluation work and cannot be acquired quickly under cost pressure, which is exactly when it is wanted.

The second is knowing where your value actually sits. If you can only report an aggregate return, a cost shock forces you to cut uniformly — trimming a percentage across the estate, which cuts the decisions that carry your value at the same rate as the ones that carry none. An organisation that has done the harder work of establishing that value concentrates in a small number of decisions rather than spreading evenly can do the opposite: concentrate spend where it earns and withdraw it everywhere else. Same cost change. Entirely different outcome. The difference was decided long before the price moved.

Write it down, then stress it

What we'd ask a client to do is unglamorous. Find the assumption in your model and state it as a sentence, with a name against it — someone whose job includes noticing if it changes. Then run the case at materially different unit costs, not to forecast anything, but to locate the point where it stops working, and then measure how far you currently are from that point. That distance is a number worth knowing. Almost nobody has it. And in the workflows where the economics are thin, keep one alternative path alive — degraded, expensive, slow, but real — because the cost of maintaining it is the price of not being cornered.

You may well be right about the price. The decline may continue, your margins may improve on their own, and the people who stress-tested the case will have spent effort on a risk that never arrived. We would take that trade, because the exposure was never that the assumption is wrong. The exposure is that it is invisible — and invisible assumptions have a particular way of surfacing. You find out it mattered at the same moment you find out it changed, and by then it is not a modelling question. It is an operational one, and the alternative path you would need is the one you already dismantled.