A demo is easy. The week after it is not.

The demo was never the hard part.

A demonstration is built to succeed on examples somebody chose. A system has to fail safely on the ones nobody chose, in front of people whose work depends on it, on a Tuesday, when the person who understands it is on leave.

Everything that makes that difference is unglamorous and gets cut first: the written boundary, the approval gates, the audit trail, the evaluation set, the behaviour when the thing cannot proceed, the handover to people who did not build it. Cut them and you have a demo that reached production, which is a different and worse thing than a system.

We build apps and the agents inside them for organisations where a wrong automated action is expensive: money moves, a patient record changes, a container sits. In that setting the question is never whether a model can do it. It is whether anyone can tell what it did, and undo it.

Six positions, and they show up in the quote.

None of these is a value statement. Each one costs something, either us or the client, which is the only way to tell a position from a slogan.

The boundary is the product

A model that can be prompted into anything is not a system. What makes it one is the written authority around it: what it may do, what it must ask about, and what happens when it cannot proceed. That is the engineering, and it is the part most projects never budget for.

Nothing irreversible without a person

The default is to propose, not to execute. An agent that can spend money, send a message or change a record on its own is not more useful than one that asks. It is the same usefulness with the liability moved onto you.

If it cannot be measured it is not finished

Quality without an evaluation set is an argument about taste. Before anything goes near live work it is run against cases that already happened in your business, and the result is shown to you whether or not it flatters us.

The system belongs to the client

It runs on your accounts, in your environment, and the copyright transfers as each increment is paid for. There is no licence key and nothing of ours in the request path. Stopping work with us stops the work, not the software.

Small enough to evaluate

Work arrives in increments you can judge one at a time and stop after any of them. A twelve-month programme with a reveal at the end is not ambition, it is a year of nobody being able to tell whether it is going wrong.

Say the inconvenient thing early

A drop in quality, a cost that has moved, a boundary that turned out to be in the wrong place. These are cheap to hear in week three and expensive in month nine, and the only reason not to say them is that they are awkward.

The work we say no to.

A supplier who has never turned anything down has not told you anything about what they are good at.

• We do not resell one client’s system to the next one. An operation that works like everybody else’s was not worth automating in the first place.
• We do not train anything on your data. Not our agents, not a shared model, not an internal dataset, and the model provider is used on terms that exclude it too.
• We do not take work that cannot be measured. If nobody can say what a correct outcome looks like, nobody can say whether it was delivered.
• We do not fit a general model over a business and call the mismatch change management.
• We do not claim experience we do not have. There are no client logos on this site and no numbers on it you cannot check.

What every engagement includes regardless is on the home page, what it costs is on pricing, and what we do with data is on security.