Datumplate

What a fair fixed price looks like for AI work

Hourly billing moves all the risk onto the buyer, and a vague fixed price is worse. The four marks of a price that is fair to both sides.

There are two reliable ways to overpay for AI work. The first is hourly billing, where the meter runs on someone else's learning curve. The second is a fixed price with no edges, where the number on the contract is only the entry fee and the real total arrives one change order at a time. Most buyers who have been burned once have met at least one of them.

A fair fixed price is a third thing. It exists, it is checkable from the outside, and this is the checklist we would want a buyer to hold us to.

Why hourly is the wrong instrument now

Hourly billing was always a proxy. Nobody ever wanted hours; they wanted the working system the hours were supposed to produce. The proxy held up while effort and output moved together. In AI work they no longer do. Modern tooling lets a small experienced team compress what used to be weeks into days, which means an hourly contract now punishes a vendor for being good at the job and rewards them for being slow. Neither incentive is one you want to fund.

Hourly also moves every risk onto you. If the ugly ten percent of documents turns out to be uglier, the estimate grows and the invoice follows. The vendor is insulated from the one thing you hired them to manage. A vendor who insists on hourly for scoped AI work is telling you, politely, that they would prefer you to carry the uncertainty.

What a fixed price actually prices

A fixed price done honestly is not a guess at build cost with a margin on top. It prices the outcome: the hours of manual review that stop happening every week, the cycle time that drops from days to minutes, the risk that a filing goes out wrong. Those numbers exist in your business today and can be written down before anyone builds anything.

This has an implication buyers sometimes find uncomfortable, so it is better said plainly: a fast vendor keeps their speed as margin. That is the trade. You get total cost certainty and a counterparty who absorbs the delivery risk; they get rewarded for efficiency instead of billed out of it. It is the same bargain you accept every time you buy a product instead of commissioning one by the hour.

The four marks of a fair fixed price

Fairness can be tested from outside the contract, before signature.

A success metric in writing. One number both sides will use to judge the work: extraction accuracy against a hand-checked sample, hours removed per week, turnaround per document. A current value and a target. A fixed price without a metric fixes the cost of something undefined, which is not the protection it appears to be.

Deliverables counted, exclusions written. "One pipeline over one document set, one integration, two revision rounds" can be finished. "Document automation" cannot. The exclusions matter more than the inclusions: what is out of scope, written down before signing, is the only thing that keeps a fixed price fixed.

A defect defined, and separated from a change. Anything delivered that does not meet the written scope is a defect, and fixing it costs nothing inside the warranty window. Anything new is a change, priced in writing before work starts. When this line is missing, every bug becomes a negotiation and every wish becomes a dispute.

Milestones, paid before work starts. Staged payment protects both sides: you never fund the whole build ahead of proof, and the vendor never works ahead of payment. A vendor with cash flow discipline is also, not coincidentally, a vendor who will still exist when the warranty is called on.

Where the number should sit

The honest anchor for the price is what the outcome is worth and what the market charges for it, and the market is not shy. Large agencies commonly quote fifteen to forty thousand dollars for a pilot on one narrow use case. Whether that is fair depends entirely on what the pilot buys: a real test has a metric it can fail, and an honest one ends with a recommendation to stop when the results say so.

So put two figures side by side before you sign anything: the price, and the fully loaded yearly cost of doing the work by hand. Priced against the manual process it replaces, a well-scoped pilot in the low thousands and a production build in the low tens of thousands are usually small numbers. Priced against nothing, every number is large. If a vendor cannot explain their number in terms of yours, the price is decoration.

One question that settles most of it

Ask the vendor to state, in one written sentence, what happens if the agreed metric is not met. A fair fixed price has an answer: the defect gets fixed at their cost, or the recommendation to stop was already delivered at the pilot stage, before the bigger cheque. An unfair one has a paragraph.

This is how we work: fixed scope, fixed price, one build at a time, with the metric, the exclusions, and the main risk in writing before anything is signed. The first step costs nothing. We send a free written diagnostic of one workflow, two useful pages, no call required. If the framing is wrong, you have lost a few minutes of reading. If it is right, you will know what a fair price looks like before we ever quote one.

All articles