How we use AI, and where the humans come in.

We were building machine learning into client systems years before anyone said the words "generative AI", training vision models like DeepLabV3 to find damage, count objects and do quality control. So when the current generation of tools launched, we were using them the same day. We'd been waiting for them. Every tool we use was chosen, tested and argued about here before it went anywhere near client work.

AI is brilliant at the work that used to eat your budget without you ever seeing it. First drafts. Boilerplate. Reading a thousand lines of legacy code and mapping what it does. That time used to be billed as typing. Now it goes on the things that actually move your business: understanding your workflows, designing the right feature, catching the edge case that would have cost you money in March.

Same senior people, more of their brain on your problem, and the hours you buy go further than they ever have.

Worth being straight about the limits too, because they're the reason the rest of this post matters.

AI has a working memory, called a context window, and it's smaller than people think. A real business system is far bigger than what fits in its head at once, so it only ever sees a slice of your codebase at a time. Push it past that and it starts guessing: it forgets decisions made ten minutes ago, contradicts itself, or rebuilds something that already exists elsewhere in the system.

And it's a confident guesser. It will invent a function that doesn't exist, make up an API, or write code that looks perfect and falls over the first time it meets real data. It never says "I'm not sure". It hands you something that looks right, and looking right and being right are different things. It's also trained to agree with whoever's typing, so left unchecked it will happily confirm a bad idea.

None of that makes it a bad tool. A nail gun doesn't check what's behind the wall either. It just means the person holding it needs to know what they're doing.

AI doesn't know your business. It doesn't know your biggest customer needs invoices formatted a certain way because of a system they bought in 2011. It can't take responsibility for a deployment, and it can't pick up the phone when it matters. That's what we're for.

So everything AI produces here is reviewed by a senior developer who knows your business, then tested, staged and monitored once it's live. AI does the heavy lifting, humans do the judging, and you get the speed with the quality you hired us for.

Worth saying plainly: nothing about AI changes who owns what. Your intellectual property is yours, it says so in our agreements, and no tool changes a word of that. The tools we use run on business terms that keep your code and your information out of anyone's training data. Credentials and secrets are managed properly and never pasted into a chat window. If a tool can't meet those terms, we don't use it, however clever it is.

Toolbelt AI, our quoting and invoicing product for trades, has AI running through it, drafting quotes and reading documents, with a human check designed into every feature. We wouldn't ask a customer to trust something we wouldn't trust ourselves.

It's even opened up a new line of work. Some software built entirely with AI hits a wall sooner or later, and more and more of it lands on our desk to be put right. We've turned that into a service. No judgement if that's you, it usually just takes the right pair of hands to get things moving again.