This question almost always comes from someone who's been left stuck with a tool before. It's not a technical question but a risk question: if I buy this, what am I locked into.

The answer below is more precise than "you're not locked into anything", because that promise can't be kept. There is a boundary, though, and it runs at a point you can know in advance.

01 / the answerWhich AI model do you use?

Claude, from Anthropic. Within that family we pick a different model per task: heavier models where real reasoning is needed, lighter and faster ones where it's about recognising and sorting. That choice sits under the hood and changes when something better comes along.

That's the factual state of things. The more interesting half of the question is what happens if that changes, and that's further down.

02 / the choiceWhy Claude and not something else?

Three reasons, in the order they weigh for us.

• Your data doesn't go into the model. For business use via the programming interface, the standard is that whatever goes in isn't used to train the model. For a system that works on your customer data, that's not a detail but a condition.
• It reasons best about work with multiple steps. That's our own finding from our own building practice, not a ranking. For a task like "read this conversation, pull out the agreements, put them in the system and watch for the exception", the difference between models is large.
• The vendor is accountable on safety. Anthropic publishes where the model's limits lie and what it shouldn't do. We'd rather build on a foundation whose builder writes down its own flaws.

What isn't said here: that Claude is the best model in the world. We don't know that and it changes by the month. It's the model that currently fits best with what we build.

03 / the switchWhat if there's a better model tomorrow?

Then we switch. That's not an exception but part of the maintenance you pay for monthly: model changes and adjustments to integrations are included. You don't get a bill for it and you don't have to do anything.

To be fair: switching is work, not a flick of a switch. Every model reacts differently to the same instruction. An instruction that does exactly the right thing with one model becomes too formal, too elaborate or just a bit too creative with another. That means retesting and adjusting, per role, on your material.

The model is interchangeable. Tuning it to fit isn't.

In practice, we therefore don't switch with every new model, but when it's measurably better for a task you actually use. Curiosity isn't a reason to open up a working system.

04 / the boundaryAm I locked into one vendor, then?

Partly, and here's where the boundary runs. Your system consists of three layers that don't all move house equally easily.

Anyone who claims there's zero dependency is selling the first layer and hiding the third. The layer that matters in a real break is the first, and that one is yours and transferable. That's what we can promise.

05 / the alternativeWhy not your own model, or an open one?

It's possible, and for some businesses it's the right choice. Putting a publicly available model on your own infrastructure gives the most control. It also costs the most: hardware or server costs, someone to maintain it, and less reasoning power per euro than what you get from a large provider.

For a business with a handful of people, that rarely pays off. If you ask about it explicitly because there's a reason things are different for you, that's a conversation we take seriously and don't wave away.

06 / usefulWhat should you ask any provider about this?

• Which model, and who decides when that changes: you or me?
• Is my input used to train the model? Ask for the answer in writing.
• What can I export if I want to leave, and in what format? No format means no answer.
• What does a model switch cost me? If the answer is "nothing", ask who does the work.

This text was produced with AI assistance and checked and approved by a human before publication.