A chief executive I work with had just come back from his ERP vendor’s annual conference. The big announcement, naturally, was AI: soon you will be able to ask the platform questions and it will answer from your data. His CRM vendor is doing the same. And he asked me a genuinely good question. How do we build a relationship between the way we want to use AI in our business and the way our platform providers are using it, so that we are not doing part of the work here and part of it over there, with bits everywhere?

I gave him a longer answer than he was expecting, because I have been chewing on this for a while, and I think it matters for every business that pays for software. Here it is, tidied up.

The vendors are in trouble, and they will not tell you

The software-as-a-service business model was built on one fact: building software was expensive, so you rented it. Per seat, per month, forever. That fact is going away. A capable person with an AI coding tool can now build in a fortnight what used to take a vendor’s team a year, and a general-purpose AI assistant can do many of the things a specialist product used to charge extra for.

So the vendors are struggling, and their response is the only one available to them: embed AI in the product, sell it as a new tier, and hope you do not notice that the assistant you already pay thirty dollars a month for can do more. They are not going to say this out loud. Nobody wants you cutting a subscription next quarter.

I am not gloating. I have built software for a living and I know how hard it is. But you should read the conference announcements with that context.

What the vendors still have, and why you should keep it

Here is the part where I disagree with the people who say rip it all out.

There is a concept in systems design called the system of record: for each part of the business, where does the data live, and which copy is the truth? Your ERP is the system of record for stock and orders. Your accounting package is the system of record for money. Your CRM for customers. Those systems earn their keep. They have years of edge cases baked in, they are stable, and the business runs on them. Ripping one out because AI exists would be hugely risky and largely pointless.

Keep them. Keep them as exactly what they are good at: the trustworthy home of the data.

What you should not do is use their AI

If you rely on the AI inside a vendor’s product to do your thinking, you are fighting with one hand behind your back.

Their AI sees only their data. Your question almost always spans systems: which customers who bought this also have an overdue invoice and an open support ticket. Their AI runs on whichever model they licensed, upgraded when they get to it. Your own assistant runs on the best model available this month. Their AI is priced as a seat. Yours is already on your desk. And their AI is designed, reasonably enough, to keep you inside their product.

So the stance I recommend is simple to say and takes some work to do.

Keep the vendor as the system of record. Get your data out through the door they already have. Point your own AI at it. Put governance around what it may do.

The door they already have

Nearly every serious business system has an API, which is the technical name for a door that other software can use. Many also have scheduled exports. Getting data out through that door is not hacking and it is not disloyal; it is what the door is for. The gap is usually commercial rather than technical: the API may sit on a higher tier, or need a conversation with an account manager. Ask early, because the answer changes the shape of everything you build.

Once the data comes out, it lands in a thin layer you own, in a plain shape, in your own accounts. Your reports run on that layer. Your automations run on it. Your team’s assistants query it. The vendor never knows or cares which model you used.

The test I gave him

Imagine your ERP vendor comes to you next year and says: a hundred dollars a month per seat, and your staff can use AI inside the platform. How do you decide?

If your data already flows out to a layer you own, the question is easy. Does their AI do anything mine cannot already do with the same data? Usually no, and you say no thank you. Sometimes yes, for one specialised thing, and you buy that one thing.

If your data is trapped, you do not have a decision. You have a bill.

And the same test works the other way. If, in two years, a different AI company has the best model, how quickly can you switch? If your logic is written against your own layer, it is a small change. If it is written inside your vendors’ AI features, you cannot switch at all.

What this asks of a business

Not much, done early. An inventory of the systems and which ones have a door. One neutral layer, built when the first automation needs it. A rule about which report is the official number, so cheap analysis does not produce three answers. And a habit: when a vendor offers AI, ask whether you can have the data instead.

The businesses that will be in a strong position in a few years are not the ones with the most software or the least. They are the ones whose information lives somewhere their own AI can reach, under an account they control, with the vendors kept exactly where they are useful.

I have written up the mechanics of owning your data if you want the practical version. And if a vendor has just offered you AI seats and you are not sure whether to say yes, send me the offer. It is usually a shorter conversation than the vendor hopes.