Before you build your own CRM with AI, sort these out first

Right now I'm coaching a small business owner through building his own CRM. He's not a developer, and he doesn't want to be one. A couple of years ago I'd have talked him out of it. Now I think it's a perfectly reasonable thing to do.

The tools have got that good. Claude, Cursor, Lovable and the rest will turn a decent description into working software faster than most agencies can send a quote. That part isn't the problem anymore.

The problem is that the first few weeks fly, and then the decisions you didn't know you were making start to show up. By then they're expensive to change. So if you're about to build your own CRM with AI, here's what I'd want you to sort out before the first prompt.

Write down how your business actually works

This sounds too obvious to bother with. It's the bit that matters most.

AI builds what you describe. If your description is "a CRM for my business", you'll get a generic CRM with a colour scheme that vaguely matches your industry. If it's something more like "an enquiry comes in by email or phone, I quote within a day, about a third turn into jobs, a job can have several stages, and I need to chase the final payment two weeks before the date", you'll get something that fits.

Before you open any tool, sit down with a pen and walk through a normal week. Where does a new customer come from? What do you do next? What do you write down, and where? What gets forgotten? Which bit do you dread?

You know all of this already. That's the point. The business knowledge is the hard part, and it's the part only you have. The AI is very good at typing. It has no idea how your business works unless you tell it.

Decide what the "things" are

Every system is built around a few things it keeps track of. In a CRM that's usually customers, enquiries, bookings and payments. Sounds simple. It rarely is.

Is a customer a person, or a household, or a company? If a couple books together, is that one customer or two? If someone books three trips, is that one customer with three bookings, or does each trip have its own record? Can a booking change hands?

These questions feel like admin. They're actually the shape of the whole system, and getting one wrong is the classic way a build falls apart at week six. I wrote about a small version of this in When 'add sizes to a product' is secretly a data-model decision. A request that sounded like a dropdown would have quietly doubled the numbers a production plan runs on.

AI tools will happily build the obvious version. They won't stop and ask whether a couple counts as one customer. You have to.

Get the boring setup right on day one

There's a handful of things that cost almost nothing on day one and a lot later. None of them are exciting.

• Keep the code in version control. GitHub, usually. It means every change is saved and you can go back when the AI breaks something, which it will.
• Use a proper database from the start. Not a spreadsheet glued on the side. Moving real customer data later is a miserable job.
• Keep a test version separate from the live one. You want somewhere to try things that isn't where your actual customers are.
• Know where your backups are. And check you can restore one before you need to.

If you're not sure how to do any of these, that's fine. It's an hour's work with someone who's done it before. It's a lost month if you skip it.

Don't wing the customer data

This is the part I'd least want anyone to make up as they go.

A CRM is full of personal information: names, emails, phone numbers, sometimes passport details or payment info. In the UK that's covered by data protection law, and you're responsible for it as the business, whoever wrote the code.

You don't need to become an expert. You do need to know a few answers. Who can log in, and what can each person see? Where is the data actually stored? If a customer asks you to delete their details, can you? If someone leaves your business, can you shut off their access in one go?

The tricky thing with logins and permissions is that when they're wrong, nothing looks broken. Everything works fine right up until the day it really doesn't. That's exactly why it's worth a second pair of eyes before you put real customers in.

Know when to stop and ask

AI tools are confident. That's mostly useful, until it isn't.

The worst version is when something breaks and the AI goes round in circles fixing the wrong thing. I've done it to myself. I once spent far too long re-reading a model name that was right the whole time. The real problem was the account tier. I wrote that one up in The cheaper model was the one behind the paywall.

A few signs you're stuck in that loop:

• The same bug has been "fixed" three times.
• You can't explain what a part of your system does anymore.
• Each fix seems to break something else.
• You've stopped reading what the AI changes and started hoping.

None of that means you've failed. It means it's time to get someone to look at it with you, before you lose a week.

What good looks like

When this goes well, you end up with software that fits your business properly, that you own, and that you understand well enough to change. No subscription for something that half fits. No waiting on a developer every time you want a new field.

That's what's happening with the CRM I'm coaching. He's doing the building. I'm making sure the foundations are right and that what goes live won't fall over. If you're somewhere in the middle of a build like this, or about to start one, that's what my software coaching is for.

More to come as I keep doing this. If you've built something of your own with AI, I'd like to hear where it went sideways.