AI product documentation for UK businesses

What an AI documentation agent is

An AI documentation agent reads what your team actually changed and updates the documentation that describes it. It goes through the places the change is recorded — the code that shipped, the tickets behind it, the meeting where it was decided, the chat where it was argued about — works out which parts of your documentation are now wrong, and rewrites those parts. The rest of the document it leaves alone.

That is the opposite of a documentation generator. A generator writes a document once, from whatever was true that day, and the document starts going stale the moment someone ships the next change. The work that keeps documentation true is not the writing, it is noticing: knowing that Tuesday's release made the third paragraph of the pricing page a lie. An agent can do that noticing every night, which is the part nobody on a busy team has time for.

We build that agent for you rather than selling you a product to configure yourself. You tell us where your documentation lives and what feeds it; we build the agent, you watch it work on your real product, and we keep it running from then on. It is one of the agents we build, and it is built the same way as the rest of our work as an AI automation agency.

What our documentation agent has done

We run one of these on a live software product, Wilow. Every night it reads what shipped that day and updates the product knowledge base, section by section, and writes down what it changed. The case study below shows what it was given, the safeguards it works within, and what it produces.

We have deliberately put no figure on that page. The thing worth showing is not a percentage, it is the record: a document that still describes the product long after the people who wrote it stopped editing it.

Case studies

Wilow's product documentation is rewritten each night by an agent. Wilow's product knowledge base is rewritten each night by an AI agent that reads what shipped, edits only the sections that change made wrong, and marks each one with how the claim was verified.

What the agent does each night

The shape of the work is the same in every business we build this for, even though the sources and the document are yours.

1. It reads the document first. All of it, before it writes anything, so it knows what the document already claims and when each claim was last checked.
2. It reads what changed. What shipped, from the system of record rather than from someone's summary of it, plus the tickets, meetings and messages that explain why.
3. It works out what that makes wrong. A change matters to the document only where the document says something about it, so the agent matches each change to the sections it touches.
4. It rewrites only those sections, in the words your document already uses. A section nobody changed comes out of the run byte for byte as it went in, so you can see at a glance what the night actually did.
5. It marks each section with where the claim came from. Which source it was read off, and the date it was last checked, so a reader can tell a verified fact from an old one.
6. It flags what it cannot settle. Two sources that disagree, or a change whose effect on the document is a judgement call, are written down as open questions rather than resolved by a guess.
7. It publishes, and writes down what it changed. The document goes where your team already reads it, and the run leaves a dated record of every edit, so you can audit any line back to the night it was written.

The sources you give it, and what it hands back to a person

An agent is only as good as the sources it is given, which is why we agree them with you before it touches anything real. They are the sources you would point a new technical writer at: the repository or release log that says what shipped, the tracker that says what it was meant to do, the recordings and threads that say why, and the document itself. We also write down which source wins when two of them disagree, because that question decides every hard call the agent will make.

Then we agree what it must never do on its own. Publish a claim no source supports, state as settled something the sources leave open, quietly delete a section, or write about a feature your customers cannot reach yet: each of those is a stop, and each becomes an open question with the reason it stopped. An agent that hands back ten questions a month is doing its job. One that never hands anything back is guessing.

How we build it, and how long it takes

We build the agent around your real product, and you see it working on your real changes before it does the job on its own. You pick one person on your team to work with us day to day, usually whoever gets asked to update the docs and never has time. Here is how it runs.

1. A 30-minute consultation. You tell us where your documentation lives, who it is for, what feeds it and where it is already wrong. We tell you what we would build and give you the price.
2. Scope. We watch how your documentation is kept today and write down your sources and the rules for reading them with you, so we agree what the agent will do.
3. Build. We build it inside the tools you already use, connected to your repository, your tracker and wherever the document is published.
4. Test. It runs over changes you have already documented by hand, so you can compare what it wrote with what your team wrote, then on your real changes each night, checked by us, until it gets them right.
5. Live. It does the job on its own, and we keep it working.

How much of the work it does on its own depends mostly on how your documentation is published: where your documentation system lets another program write to it, the agent publishes each edit itself, and where it does not, it prepares the edit for someone to paste. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are.

What it costs

Ours is one flat monthly fee for unlimited agents: this agent, and any other you want after it, for the same fee, month to month. We give you the price on the call.

We set it on the consultation once we know how many sources it has to read and where it has to publish. What it will not do is charge you by the document or the run: the usage is in the fee, so a busy release month costs the same as a quiet one, and you can stop any month. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are. What it would cost you to run an agent like this is the subject of what an AI agent costs.

An agent, a documentation tool, or someone's Friday afternoon

There are three ways to keep documentation true, and the right one depends on how fast your product changes and how much it costs you when the document is wrong.

A documentation tool that generates pages from your code is the cheapest to start with, and where your documentation is a faithful description of an interface it may be all you need. What it cannot do is read the meeting where you decided to drop a limit, or notice that the sentence about pricing is now out of date.

Asking a person to keep it current buys judgement, and a writer who notices the thing that reads wrong. It is also the task that loses every argument with a deadline, which is why most documentation is a snapshot of the week it was written.

An agent sits between the two: it does the noticing every night, inside rules you wrote, and it hands you the questions it could not answer. Testing it on your real product is how you find out which of the three you actually need.

Your data, and the tools the agent can reach

We handle your data under UK GDPR, and we walk you through exactly where it goes on the consultation. The agent can only reach the repositories, trackers, recordings and documentation systems you approve, and you can withdraw that access at any time.

The document stays yours, in your own system, in the place your team already reads it. Nothing about the way the agent works depends on your documentation living with us.

Tell us where your documentation lives today

Tell us what your documentation is for, who reads it, and which part of it you already know is wrong. We will tell you on the call whether an agent can take that work, what we would build, and what it would cost.

One flat monthly fee for unlimited agents, month to month. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are.