Your first agent, working in two weeks

We write down the assumptions behind the payback calculation so you can check the arithmetic. The go/no-go is on day three; if the answer is no, we receive no payment, including no reduced fee for those days. That is why the calculation happens at the start rather than being sold as a separate service. The walkthrough measurement uses your cases. You can also buy the analysis from an independent adviser and bring us the conclusion; we will build against their brief.

The fixed price covers one task, your existing data, and draft-and-approve rather than autonomy. We write those boundaries down before the work starts. Measurement remains part of the deliverable and can show that the agent does not meet the agreed bar.

Real enough to do the work on your data and be measured doing it. It is one task, and it runs where a person still approves the output. What is not in the fortnight: production hardening, wider autonomy, and integrations that have to be written from scratch against a system we have not seen. Those are a separate decision, and we say which of them your case needs before the start rather than at the walkthrough.

Then we say so in the first days, with the math, and you pay nothing at all, including no reduced fee for the days we used. The write-up is still yours: what would have to change for the answer to become yes, the data you would need to start keeping, the volume at which it starts to pay, and the step that would have to be written down first. Reaching that answer takes three hours of your team's time.

Yes, and for a first engagement we prefer it. A slice costs €1,500 and takes 20 hours, and it is not a report: what you get is a working thing on one narrow part of the task, which you run yourself. You put your files in, you get the result out, on your own machine without us. With it comes the number it scores on your real cases, the errors broken down by kind so you can see what a rule would fix and what runs into data you do not have, a set of forty to fifty cases with agreed correct answers that stays yours, and what it would cost to widen this to the whole task. What a slice cannot include in 20 hours is a connection to your mail, CRM or accounting system: it works on an export, and the integration is quoted separately. If no working slice comes out of those hours, you do not pay for them. The fortnight remains available on its own when the task is obviously well defined, but we will not propose it first: for €1,500 you get something running and a number on it, and you decide about the larger figure having seen the work rather than a description of it.

Then we say so on day three and offer a different shape of work, not a discount. A fixed-price outcome requires your own people to agree on what a correct answer looks like. In the first three days, two of your staff label the same cases independently and we compare them. Where they agree, what remains is engineering, and engineering can be priced. Where they half agree, the missing piece is the rule itself, and a rule is written by the hour: the same block of 20 hours, except what it buys is our time and a written finding rather than a promised number. Where they do not agree at all, we do not take the work in any form.

About three hours of your team's time in the first days, a look at your data and systems, and — this is the part that decides the quality — someone who can say what the correct answer is for the cases in the test set. We do not need production access for the fortnight.

The same PhD-trained researchers and engineers who ship our projects. You can see who they are on the team page.

Whatever the number says it should. Usually one of three: widening the agent's autonomy where the measurement supports it, putting your documents behind it so it answers from what the company knows, or keeping the test set running against every model and prompt change so a drop shows up before your users find it. The first two are bought in the same unit: a block of 20 hours at €1,500, with a walkthrough at the end of it. The third is not a block, because a one-off cannot catch a model that changes six weeks from now: it runs monthly at €300 a month, which buys the set re-run on every model and prompt change, the number sent to you each month, and the repair at our expense if it falls more than three points below the threshold. It is also optional, and we say so before invoicing rather than after: the test set and the script that runs it are yours either way, and if someone on your side will press the button once a month, you do not need us for that. How many blocks a given step takes depends on the number the first agent scored, and quoting that total before the number exists is guessing. What we can say is that blocks after the first go further, because the test set is already built.