We made a whole CRM work from a chat. Here is what it took

We made Plan@Job, our CRM for contractors, work from a chat. What you used to do by clicking, you can now do by asking an AI assistant. That's more than five hundred actions. This is how it went, including the parts that went wrong.

Where did we start?

With seventeen. A small set of things the AI could do, mostly looking things up.

It made a nice demo. It was also useless on a real day, because the moment you asked for something outside those seventeen, it said it couldn't. People don't keep using a tool that says no.

So I set a rule. Anything that can be done by hand in the platform has to be doable from a chat. Not most things. Anything.

What was the first thing we built?

Not more actions. The wall.

Before the AI could do more, it had to be impossible for it to do harm. Two lines we would not cross: it can never reach into another company's account, and it can never approve its own work.

So every action runs as the person who is logged in, through the same checks the buttons use. If you can't do it by clicking, the AI can't do it for you.

How did we get from seventeen to five hundred?

In batches, by part of the business. Messaging first. Then jobs, quotes, invoices and reports. Then site visits, documents and contracts. Then the company settings.

The trick was a list. We wrote down every action in the platform and ticked each one off as it got its chat version. Anything left off needed a written reason.

Then we turned that list into an automatic check. Add a new feature without its chat action and the check fails. That's the only reason the number stays complete as the product grows.

What needs a confirmation?

Anything that spends, sends, approves or deletes.

The AI prepares it and shows a preview: who gets the message, what it says, what it costs. You confirm that exact thing, and only then does it happen.

It's slower than letting it run. It's also why I'd trust it with a real company.

What broke?

The limits. We put a cap on how many messages could go out in an hour, and we set it far too low.

It got hit straight away. Worse, the error was vague, so the AI reported the send as unconfirmed. Nobody knew if the message had gone.

We raised the cap, and we fixed the message so it says plainly: you've hit the limit and here's when it resets. We also made sure a retry can never send the same thing twice.

The lesson: an AI repeats whatever your software tells it. If your errors are vague, your AI is vague.

What surprised me?

How much of the work was writing. Each action needs a description the AI reads to decide when to use it. Write it badly and it picks the wrong one.

We ended up writing them like a briefing for a new member of staff. That's what it is.

The other surprise was who it's for. I thought it was for me, the owner. But subbies, operatives and property managers each get their own view. A subbie who asks about his jobs gets his jobs, and nobody else's.

What does a day look like now?

You open the chat and ask what needs you today. It tells you what's waiting.

You say chase the invoices and book the plumber. It shows you what it's about to send. You say yes.

That's the morning's admin, done before the kettle has boiled.

What would I tell someone doing this?

1. 01Build the wall before the features.
2. 02Aim for everything, not the easy tenth. A partial tool gets abandoned.
3. 03Put a confirm step on anything you can't take back.
4. 04Make your errors honest.
5. 05Add an automatic check so it stays complete.

This is what we now build for other people: software an AI can run, with the wall included. The thinking is in the future is not software and the checklist is in what makes software AI-ready. If you want yours done, start a conversation.