Three Cents On The Dollar

Three cents.

That is roughly what one extra dollar of AI model spend is replacing, in routine knowledge work, at the firms most exposed to the shift. Among the most AI-exposed companies, each $1 drop in spending on online labour marketplaces corresponded to only about three cents in added AI model spend by the third quarter of 2025. The ratio comes from firm-level payments data analysed by Ryan Stevens of Ramp, and it has been doing the rounds in public-sector commentary this week.

If that number holds, the implication is not that AI is ‘cheaper’ in the comforting, marginal way most agency notes describe it. The implication is that the market price of a meaningful slice of first-pass, reviewable knowledge work has collapsed by roughly an order of magnitude. That is not a procurement story. It is a public finance story, and we should treat it as such.

What the 33-to-1 ratio actually says

A 33-to-1 substitution ratio is not, strictly, a substitution ratio. It is a ratio between two budget lines, captured at one moment, in one slice of the economy, by one provider’s payments data. Read with care, it tells you three things.

One: buyers have found a class of output where the first pass is now an AI deliverable rather than a human one, and the first pass is the part they used to outsource. Two: the unit cost of producing that first pass has fallen far enough that even a sceptical CFO is rewriting last year’s contractor budget. Three, and most important: the saved spend is not, mostly, being reinvested in a more expensive in-house model. It is leaving the books altogether.

For a government department, that last point matters most. In the private sector, ‘leaving the books’ eventually shows up as somebody’s lost job. In the public sector it shows up - or refuses to show up - as a backlog that does not grow, a clearance cycle that does not lengthen, a sanctioned post quietly left vacant. The change is invisible until you go looking for it.

Why government cannot just copy the ratio

It would be a mistake to read this number and rush to swap people for tokens inside a tax administration. Public-sector value is not what the private sector buys when it buys ‘first-pass output’. From inside a national tax administration, one learns quickly that the value of a routine notice is not in its first draft. It is in the fact that the draft is recorded, attributable, reviewable and defensible six years later under cross-examination.

None of that vanishes if AI does the drafting. But it changes who is accountable for what, and at which moment in the workflow. Records-management law, natural justice, statutory limitation, the discipline of audi alteram partem - these are not procurement constraints. They are the architecture of the trust the department holds with the citizen. Any honest AI strategy in this domain begins there and works backwards to the model, not the other way around.

The serious question for the head of any large administration is therefore not ‘where can I deploy a model?’ It is: where in my workflow is the first pass currently a bottleneck, who reviews it, and what does that review actually verify? If the review is genuine - a senior officer reading, weighing, signing - AI is a gift; the bottleneck moves to where it belongs, which is judgment. If the review is rubber-stamping, AI will reprice that work to zero and the institution will not notice until the courts do.

Where the repricing is already real

The categories where this repricing is real will be familiar to anyone who has worked the field. Drafting of routine intimations and standard show-cause notices. Cross-referencing returns against third-party information. Summarising voluminous assessment records before a hearing. Preparing first-pass replies to grievance petitions. Tagging and classifying taxpayer correspondence. Translating dense statutory text into plain-language guidance for citizens.

Each is a place where the marginal cost of a first pass has fallen sharply. Each is also a place where a human signature still ought to mean something. The design problem is precisely how to keep the signature serious while the draft becomes cheap.

Three operating shifts that would actually help

• Measure the bottleneck, not the licence. An honest dashboard tracks clearance cycle times, refund-issue latency, time-to-first-response on grievances. If those are not moving, the model is decorative.
• Promote the reviewer. Where AI does the first pass, the human downstream should be more senior, not less. The saving is bought by elevating the reviewer’s standard, and paying her for judgment rather than throughput.
• Budget for the audit trail before the model. Every AI-assisted output in a tax administration must carry a verifiable record of what the model saw, what it produced, what the officer changed, and why. The audit trail is the primary product. The draft is secondary.

The budget question we should be asking

Michael Ting’s course on public sector organisations at Columbia used to insist that the deepest puzzle in bureaucratic design is not how to motivate effort but how to allocate scarce attention. The Ramp number, read honestly, is a story about attention, not effort. Cheap first passes free attention; what an institution does with that freed attention is the whole game.

The danger a department most needs to guard against is the most boring one: cashing the saving without rebuilding the work. If sanctioned posts are quietly left vacant, if the freed hours are absorbed by more of the same routine work, and if the audit trail is patched on at the end, then the 33-to-1 ratio will become the public sector’s number too - and we will have repriced our own knowledge work without ever asking what we wanted to do with the difference.

The better view, I think, is to treat this moment as a budgeting exercise, not a tooling exercise. Ask, for every freed officer-hour, what higher-value question we now want that officer to be asking. Then build the model around the answer. The cheapest first pass in history will be wasted on the same old second pass.