How we use AI
We use AI in some of our work, not all of it, and how much depends on the project and on what you agree to. That raises a fair question about what happens to your data. This is the answer. Most of what we build is written by hand. Where AI is used, it speeds up parts of the job. It does not do the job, and it never decides anything.
Last updated 12 September 2026.
You decide, before anything starts
No AI tool goes near your project until it is agreed in writing. That happens at the start, not once the work is underway.
No is a complete answer. If you would rather no AI touched your work at all, we build it without, and it changes nothing about what you get or what it costs.
Where you have your own environment and your own tooling, we work inside it and your rules govern. That is the cleanest arrangement and we are happy with it.
Most clients do not have that, and that is normal. In that case we use our own tooling, configured so nothing you give us is retained or used to train anything, and we tell you what we are using.
Your data does not go into a public model
Client data is never entered into a free or consumer AI tool. Those services can train on what is put into them, and that is not a risk we take with someone else’s business.
That covers personal data, credentials, keys, financial records, client documents, and anything specific to how your business makes its decisions.
A person is accountable for every line
AI-generated code is read line by line, understood and tested before it goes anywhere near your system. It is treated exactly as if it had been typed by hand, because as far as responsibility goes, it was.
Everyone here can build what they build without it. That is the condition for using it at all. A tool that speeds up someone who already knows the work is worth having. A tool standing in for knowing the work is a liability, and it shows the first time something breaks.
We take full ownership of maintaining it. If something breaks in six months, the answer is never that a model wrote it.
The same applies to judgement. Where a tool helps us read or process something, the reasoning stays visible and a person checks it. Automation nobody can audit gets switched off the first time it is wrong, so we do not build it that way.
Plenty of firms paste output and ship it. If that is what you have had before, this is the difference, and it is the reason we can still answer a question about the code two years later.
What we watch for
Libraries a model suggests are checked for licences that would cause you a problem commercially.
Generated code goes through the same review, testing and security checks as anything else. It gets no easier ride for being fast.
Anything a model asserts as fact is verified before it reaches you. They are confident when they are wrong, which is exactly why a person stays in the loop.
Where we will not use it
Anywhere you have said not to.
On anything where a wrong output would reach a customer or a regulator without a person seeing it first.
As a substitute for understanding your business. The diagnosis is the part we do ourselves, and a model cannot do it because it has not sat in the room.
When the work ends
Whatever you gave us to do the job is destroyed or returned when the job is done, on whatever terms the agreement sets. Ask for it sooner and it goes sooner.
What stays is your code, your accounts and your infrastructure, in your hands. That is the whole point of how we build.
Questions
Where your own policy sets a higher standard, send it and we will work to that instead.