Four services, and the last one decides the others.

Somebody maps the process. It has seven steps, three decision points and a clear happy path. The automation is built against it, tested, released — and afterwards the team spends roughly the same amount of time on the work as before.

The process document described the standard case, because that is what a process document is. It is written by asking somebody what they do, and what they describe is the version that runs cleanly. The exceptions were where the hours went, and they are still being handled by hand.

So we map the exceptions before the standard case, count them before scoping, and say plainly when the numbers mean the project is not worth doing. That conversation costs less before the budget is committed than after.

Workspaces and knowledge systems fail the same way every time. They are built correctly, launched properly, and then nobody uses them — because using them is slower than the thing they replaced, for the first fortnight.

Adoption is a design property, not a change programme. If the new way is faster on day one for the work people actually do, it gets used. If it is faster on day thirty, it does not, because nobody is measuring on day thirty.

So we build for the first day. Which means knowing what somebody does on a Tuesday afternoon, rather than what the process document says they do.

A retrieval agent finds the documents. A summarising agent condenses them. A reasoning agent draws a conclusion. The output is wrong, somebody asks which one caused it, and the honest answer is that nobody knows.

The logs exist. What they contain is the input to each agent and its output, and both look reasonable in isolation. The intermediate states are the evidence, and they are gone by the time anybody asks.

So attribution is designed in from the start. Every agent records its reasoning rather than only its output, every artefact carries what produced it, and the whole chain can be replayed six months later without the running system.

Deployments rarely fail on capability. They fail where an output reaches a desk and nobody has been given the authority to act on it — a question that could have been settled in a room, in an afternoon, months before anybody opened an editor.

So we start there. What is actually wrong, what the evidence says about it, what should be done, and whether building is the answer. A fair number of engagements end at that fourth question, which is the cheapest outcome either of us will get.

Three free instruments sit under this, and so does everything we publish. Neither asks for an email before it gives you something.