Series · Accountability in Regulated AI
The system is getting more efficient, and I need more budget.
The Named Owner / Judgement Muscle — cheap models will swallow most of the work. They will also send the bill up. The leaders who cannot explain why both are true at once are the ones who lose control of their spend.
Cheap models will swallow most of the work. They will also send the bill up. The leaders who cannot explain why both are true at once are the ones who lose control of their spend.
I have said that sentence to a Chief Financial Officer, out loud, across a table. The system is getting more efficient, and I need more budget. The pause that follows is the whole problem. To the person who signs the cheque it sounds like a contradiction, possibly a sleight of hand, the sort of thing a vendor says when the invoice has run away from them. It is neither. It is the most honest description of AI in production I know, and the people who cannot say it cleanly are the people who quietly lose the plot on cost.
A recent argument from UncoverAlpha, built on Epoch AI's cost data, makes the case crisply. Most enterprise AI tokens will end up running on cheap, distilled models rather than frontier systems, because the cost of reaching any given level of capability is collapsing. Epoch's own study puts the median decline at roughly fifty times per year across benchmarks, with a range running from nine times to nine hundred times depending on the task, though the fastest of those rates only appear after early 2024 and may not hold. To make it concrete, the price to match GPT-4 on PhD-level science questions fell by about forty times per year. Yesterday's frontier becomes this year's commodity. The conclusion is correct. The reading most people take from it is not. They hear most tokens run on cheap models and they file it as run everything cheap, and the saving will show up on the bottom line. That is the expensive misunderstanding.
01. The paradox is not a paradox
Efficiency is measured per task. Spend is unit cost times volume times scope, and only one of those three is going down. When a task gets cheap enough to be reliable, two things happen in the same quarter. You can finally point it at work that was uneconomic last year, and the organisation, having watched it work, asks for more of it. Success is what drives the bill up. The thing that is going right is the thing inflating the number.
This is not AI behaving strangely. It is Jevons paradox, first observed in coal in the 1860s, observed again in every cloud invoice that ballooned precisely because compute got cheaper. Name that precedent when you are in the room. It tells finance you are not pleading a special case for a hyped technology, you are describing a known economic shape they have already lived through once. That reframes you from optimist to operator in a single sentence.
02. Stability is what unlocks the saving, not the other way round
The right-sizing move, dropping a workload from a frontier model to a cheaper distilled one, is the engine of the whole cheap-model thesis. What goes unsaid is that you can only make that move on a system that has already stabilised. You cannot distill an unstable workload, because you do not yet know what good enough means for it. The eval harness does not exist, the failure modes are not mapped, the floor is unknown. You cannot right-size against a target you have not measured.
Which inverts the usual framing. Moving into stable production is not a phase where cost passively drifts down. It is the phase that first makes deliberate cost reduction possible. Maturity is not the point at which the system becomes cheap. Maturity is the point at which you have earned the right to make it cheap, on purpose, with evidence. That distinction is the difference between a saving you can defend and a gamble you took on the bill.
03. Two traps on the way down
The first is optimising the edge into mediocrity. The cheap-model logic holds for undifferentiated plumbing, the stable, specified, commodity work that simply has to run. It does not hold on the edge where you are actually competing. Frontier spend belongs there. Collapse the distinction, cost-optimise everything to the cheapest model that passes, and you economise your way into being average on the exact capability a competitor just bought a step-change on. Good enough is a bar the market sets and keeps raising, not one you set once and bank for three years.
The second is the silent regression, and in a regulated world it is the one that bites. A distilled model does not fail less, it fails differently, usually at the edges, and the eval that justified the downgrade was built on last quarter's distribution of inputs. A quality regression nobody notices is not a saving. It is a liability sitting on a desk, accruing, until it surfaces somewhere expensive. The cheap-model decision looks like a finance call. It is a governance call wearing a finance disguise, and treating it as the former is how the cost reappears later with interest.
04. The prize is legible cost, not lower cost
Here is the part most people miss, and it is the part worth standing in the room to say. The prize of stable production is not a smaller number. In aggregate the number goes up, and anyone who has actually run one of these systems already knows it. The prize is a number you can see coming. An unstable workload has spend that is unbounded and unforecastable, a line nobody can plan against and nobody can defend. Stabilising the system is what converts raw efficiency into a figure finance can put into a model and take to a board.
A volatile line cannot be underwritten. A forecastable one can. So the budget ask was never give me more because it got cheaper, which does sound like a contradiction. The ask is, stabilising this is what lets me hand you a number you can stand behind. That is not the awkward request to make of a CFO. Once you say it that way, it is the only rational one.
05. Someone has to own the curve
None of this manages itself. The efficiency was always coming, the Epoch curve delivers that on its own, you mostly have to wait for it. What does not arrive on its own is the discipline to spend the saving well. Someone has to own the trajectory. Where the line sits between plumbing and edge. Which downgrades are safe and which are quietly accruing risk. Who carries the regression when it surfaces, because it will surface. That is a named owner doing real work, not a name on a slide nobody reads.
And the call itself, plumbing or edge, save here or spend there, hold the line or move it, is judgement of the kind that does not survive being handed to a dashboard. The dashboard can tell you the cost per token fell. It cannot tell you whether the place you cut was the place you were competing. That is a person, sitting with the trade-off, willing to be wrong in public.
The bill is going up either way. Falling unit costs guarantee it, success accelerates it, and no amount of cheap-model evangelism changes the arithmetic. The only real question is whether anyone in your organisation can stand across the table and defend the shape of that bill. Find that person. Name them. Give them the authority to hold the line. The efficiency is free. The judgement is not.
Where this sits
Part of The Named Owner and The Judgement Muscle, with a foot in the Accountability Turn.