The Three-Cent Government Hour

A small number in a payments-data paper has been bothering me. Among firms most exposed to generative AI, every $1 fall in spending on online labour marketplaces by the third quarter of 2025 was matched by roughly three cents of added AI model spending. Three cents where a dollar used to sit. That is not a productivity nudge. It is a repricing of a whole category of work.

What the ratio actually says

The number comes from a Ramp analysis picked up in commentary on US federal AI policy. The exposed firms were not abandoning the work. They were buying the same first-pass output — drafts, summaries, light code, document review — at a fraction of the prior price. A slice of knowledge work has moved from a labour line item to a software line item, with the ratio between the two collapsing by more than an order of magnitude.

For private firms, this plays out through hiring and margins. For governments, it plays out through almost everything: workforce composition, training pipelines, procurement rules, and the implicit deal that lets young officers grow into senior ones.

Why tax administrations sit in the bullseye

Walk through any direct-tax office and ask what the work actually is. Reading. Drafting. Summarising. Comparing one provision against another. Translating dense statute into plainer language. Spotting the inconsistencies between a return, a third-party report and a bank statement. Almost the entire stack is language-heavy. That is exactly where the three-cent ratio bites hardest.

I have watched this from the inside. Mapping an old direct-tax statute against a new one — tracking which old section migrates where, what is dropped, what is reorganised — is genuinely difficult professional work. A few years ago that effort would consume dozens of officers for months. A current-generation model, given the right corpus and a careful prompt, will now produce a competent first draft of much of it in an afternoon. Not the final word. But a credible first pass.

The same is true of taxpayer-facing communication. Building an assistant that can field lakhs of routine questions on a new law — what the slabs are, how to elect a regime, what to fill where — was, until recently, a serious capital and talent project. Today the floor for that capability has fallen sharply. The hard part is no longer building the bot. The hard part is governance: what it is allowed to say, how its mistakes are caught, how a taxpayer appeals an answer that turned out to be wrong.

The junior officer problem

A Stanford working paper this year found a 16 per cent relative employment decline for workers aged 22 to 25 in the occupations most exposed to generative AI. Read that and a managerial instinct should fire. In any large department, the junior cadre is not just there to do work. It is there to learn. The years spent reading scrutiny files, drafting orders, sitting through hearings — those years are how a tax officer becomes a tax officer.

If the model does the first draft, what does the junior do? The wrong answer is: nothing, until they are senior enough to “supervise” the model. A supervisor who has never written the draft cannot meaningfully review one. Within a decade we would be running an administration whose middle ranks know how to prompt AI for a note but cannot tell when the note is quietly wrong on a point of law.

What the junior role should become

I think the redesign is closer to this. The junior officer is no longer the drafter. They are the evaluator, the contester, the case-builder. From day one they are taught to interrogate a machine-generated draft: where is the citation, is the authority current, does the inference survive cross-examination by a contrary view. They are taught to construct the hard cases the model gets wrong, and to document them. They become the institution’s quality control function rather than its typing pool. That is more demanding work, not less. It suits the calibre of people the service actually recruits.

A proposal worth piloting

If a tax administration wants a concrete way in, here is one. Pick a single, well-bounded workflow — say, drafting routine rectification orders, or first-level responses to grievance petitions. Build a model-assisted pipeline with three deliberate seams: a machine first draft, a structured human evaluation against a checklist, and a logged audit trail of every override. Track three numbers — time saved per case, override rate, and downstream litigation outcomes for the cases that went out. Run it for a year. Publish the numbers.

This is unfashionable advice in a moment that prefers headline pilots and grand strategies. The three-cent ratio is not going to wait for a strategy document. It is already changing what a dollar of knowledge work buys in the market. A serious department asks what that means for its own internal economics, builds the governance to capture the gain safely, and quietly redesigns its junior roles before the redesign happens to it.

The departments that get this right will not be the ones that bought the most expensive tools. They will be the ones whose officers learned, early, to argue with a machine and win.