The questions that come before the price.

It keeps running. The system runs on your accounts and in your environment, the code is in your repository, and the copyright in it transfers to you as each increment is paid for. There is no licence key, no service of ours in the request path, and nothing that phones home. Nothing we could switch off, in other words, because there is nothing of ours left switched on.

It is also ordinary software in ordinary languages rather than a platform only we understand, so another firm can pick it up and read it. What you would lose is the monthly line, which is monitoring, updates, re-testing and a named person to call. You would not lose the thing you paid to have built.

Because the part that usually fails is not the model. A demo is built to succeed on examples somebody chose. A system has to fail safely on the ones nobody chose, and that is a different piece of engineering that most projects never budget for.

So the boundary is written down before anything is built, the thing is tested against your own past cases before it goes near live work, and there is an evaluation set that states what correct means for you. Without that last one, quality is an argument about taste and a drop cannot be measured, only felt.

It gets re-tested. On every model change, on every prompt change, on any measured drop, and at minimum once a quarter whether or not anything changed. You get the result either way, including when it is bad news.

The cost per action is modelled before launch, with caps and alerts, and the provider account is in your name so the bill you see is the bill we see. Which provider to use is a decision with a switching cost, and both are written down when it is made rather than discovered in the year it matters.

You do. Copyright transfers to you as each increment is paid for, and the repository, the hosting and the model provider account are in your name from the first day rather than migrated to you at the end.

No. Not our agents, not a shared model, not an internal dataset, and the model provider is used on terms that exclude it as well.

No. What gets delivered is software with a model inside it, and it is operated the way you operate the rest of your software.

What you do need is somebody who can say what a correct outcome looks like in your business, in enough detail to disagree with a wrong one. That is the single thing nobody outside your company can write down for you, and it is the thing the evaluation set is made of.

Three weeks to the operations map and a costed shortlist, which is the same work whatever gets built afterwards. After that the build arrives in increments you can evaluate one at a time, rather than as one delivery at the end of a programme.

Yes, and it costs less at the start than at the end. Data residency, retention and who may access what are architecture decisions. Retrofitted six months in for a review the pilot was never designed to pass, they are a rebuild, and that is the most common way a working pilot dies.

What we hold, who else processes it, where it sits and what we are not certified for are published on this site rather than answered one questionnaire at a time.

The sector is rarely the hard part. Where it is, it shows up in the boundary, which is written per engagement anyway.

Ten sectors are written up with the specific place the work breaks in each, and the specific things an agent is not allowed to do there. If your week is described accurately on one of those pages, the sector is understood well enough to start.

A fixed fee for the first three weeks, a range for the build quoted against a boundary agreed in writing, and a monthly charge to keep it running with no minimum term and no exit fee.

The bands are published, along with the nine decisions that move a quote inside them and the six things worth insisting on in any quote, including ours.