Every software vendor now has an AI story. Your CRM has one, your ERP has one, and so does the tool someone in marketing signed up for last month. On top sit the model providers themselves. In most organisations the result is a dozen overlapping AI products, bought separately, each with its own contract and its own data terms.
Choosing well isn't about picking the best model. Models change every few months, and today's leader may not lead next quarter. It's about making decisions that still hold when they do.
Disclose the bias first
I'll start with mine. I'm a Claude Community Ambassador for Anthropic in Copenhagen. I know their tools well and use them every day. I also recommend other tools when they fit better, and I say so when I do.
Anyone advising you on AI vendors has some bias. The useful ones tell you what it is. Ask the same of everyone advising you, including your existing IT partners. Reseller margins and partner programmes shape recommendations more than most people admit.
Start from the work, not the product
Vendor conversations go best when you arrive with specific use cases, the data they need and what success looks like. Arrive without them and the vendor will supply their own, which will fit their product.
Five things to pin down
Where your data goes. Whether it's used for training, where it's stored, how long it's kept, and what the enterprise terms actually say. Read the terms, not the sales deck.
How you leave. What it takes to switch. Prompts, workflows and integrations built around one vendor's features are the real lock-in, more than the contract.
What it costs in use. Per-seat prices are easy to compare. Usage-based costs aren't, and they grow with success. Model the cost at the volume you hope for, not the pilot volume.
How it fits what you have. Whether it works inside the systems people already use, through standard connections, or needs a detour. Tools that need a detour are one of the reasons pilots stall.
Who's accountable when it's wrong. What the vendor commits to, what they don't, and what that means for your own obligations under the EU AI Act.
Keep the architecture open
The safest assumption is that you'll change models at least once. A few habits keep that cheap.
Keep prompts, instructions and test sets in your own systems, not only inside a vendor's product.
Prefer tools that support more than one model, and open standards for connecting AI to your data. The Model Context Protocol is becoming one of those.
Test against your own tasks. Public benchmarks say little about your documents and your processes. A small set of real examples with known good answers says a lot.
Buy less, decide once
Most organisations need fewer AI products than they have. One general-purpose assistant for everyone, chosen carefully. A few specialised tools where the use case justifies them. And AI features in existing systems, used where they're good and ignored where they're not.
That decision should be made once, written down and revisited on a schedule, not remade by every team. Which tools and models to buy is one of the five decisions in what a Chief AI Officer actually owns.
Sometimes the right choice is to buy nothing yet. That's a decision too, and often a good one.