Quoting AI-native delivery
We keep meeting two versions of the same meeting.
In one, a capable team has quoted an AI-native delivery, and the buyer is looking at a line they have no way to evaluate. They ask what it costs. Nobody asks what it contains. The deal stalls on a number rather than on the work.
In the other, the same capable team has broken the same delivery into parts a non-technical buyer can follow. The conversation moves to scope within twenty minutes.
The difference between the two is not competence in the work. It is whether the method can be accounted for.
That is a teachable skill, and it is currently rare enough to be a competitive position.
Four buckets, not one figure
Most AI-native quotes we see collapse four unrelated costs into one flat monthly number:
- Tools -- the team's own licences and subscriptions
- One-off work -- migration, initial localisation, verification and reruns
- Phase-bound effort -- runs that exist only during a specific project phase
- Genuinely ongoing operation -- what keeps costing after launch
Only the last justifies a fixed monthly fee, and it is almost always the smallest. The real curve is two humps with a thin line between them. Flattening it produces a number that is too high most months and too low exactly when it matters, which is a promise you will have to renegotiate later.
Quote the four separately, with a project total and the number of months per phase. This costs you nothing and removes the single most common reason a buyer stalls.
Count what actually needs judgement
The strongest thing you can put in a quote is a count.
Not "content will be migrated using AI", but "twelve thousand objects, of which roughly four hundred need a model because they are unstructured free text that has to be reinterpreted as components; the remaining eleven thousand six hundred are a deterministic mapping."
That sentence does three things at once. It shows the work has been examined, it prices per object instead of per night, and it tells the buyer you know the difference between where a model earns its cost and where a script does the job.
The same discipline applies to the build chain. Gates that must return the same answer every run belong in deterministic tooling, and saying so in a quote reads as judgement rather than as a smaller invoice.
Show where the efficiency lands
If the method makes you faster, the quote should say where that shows: in hours, in price, or in delivery time.
When it shows nowhere, the buyer is paying twice -- once for the tooling meant to produce the gain, once for the hours the gain was meant to remove. Most buyers cannot yet articulate that, but they feel it, and it surfaces later as pressure on the rate.
The version that wins is a comparison: what this delivery would have cost without the AI-native setup, and what it costs with it. Very few suppliers put that on paper. It is disarming precisely because it is checkable.
Name the gate
Overnight agent runs are legitimate practice, and the licensing reason for running them through an API rather than personal subscriptions is sound. What buyers increasingly ask is who reads the output in the morning.
Answer it as a mechanism, not as a reassurance. What counts as finished, who approves, whether anything can reach production without a named person saying yes, what the test coverage is, and how many review hours per day are in the estimate. In a migration, add the natural point of comparison: old against new, verified how.
"We review everything, quality matters to us" is the answer that loses. It is indistinguishable from the answer given by a team that does not.
Where this sits in our work
This is the commercial face of what a Genomlysning produces. The same examination that tells a team which steps need judgement and which need a rule is what lets them write a quote a buyer can read.
It is arithmetic and description work rather than technical work. It takes roughly an afternoon per project.
Knowing your craft and being able to account for it are two different abilities. Right now the second one is rarer, and therefore worth more.
Related: Governing AI cost at scale and The accountability test. The buyer's side of this material runs as a five-part series on tomasandre.se, beginning with An even number over uneven work.