If you are a manager or a public-administration official who suddenly has to decide which artificial intelligence to buy, how do you actually make that call?
A lot of decision-makers right now - in the private sector and in the public one - are stuck in a kind of schizophrenia. On one side: “Let’s just take the commercial product.” But then the commercial product ships all our data off to American servers. On the other side: “Let’s have something built custom for us.” But then how do we integrate it with our existing systems? What are the risks? And who teaches us to use it?
Here is the core idea that dissolves most of that anxiety.
AI is not a product you buy in one block
You don’t buy “the artificial intelligence” as a single thing. AI is not a product. Saying “I’ll buy the AI” is like saying “I’ll buy the web.” What are you actually purchasing? You’re purchasing several different pieces - and once you see them as separate, the whole decision gets clearer.
The harness and the model are two different things
What people now call the harness - think Claude Code, Copilot, or Cheshire Cat 2 (already on main, go look at it on GitHub) - is the software that actually runs AI agents. The harness is one piece. The language model it connects to is another piece entirely.
These really are separable. With Cheshire Cat, for example, you can use Claude today, OpenAI tomorrow, and the day after a model running in your own house on your own server. Same harness, different engines underneath.
The first question: am I free to change models?
So when you buy software that plugs into AI, the first thing to evaluate is:
Am I free to swap the model whenever I want? Can I run this AI on a pay-per-use basis, whether the model comes from outside or locally from machines I own? Am I free to change?
That is your first constraint to check. Today maybe you use a cloud model like Claude; at some point, when it costs less, you put a little something in-house, you have your people experimenting with small local models. You need to be able to change the model as you please.
The second question: can these pieces be composed?
The second constraint is about the composability of these agents - their ability to be integrated with pre-existing systems, according to interoperable standards, so nobody puts you in a cage.
What do I mean? Say you want your AI agent to work with your email. If a consultant from some Big Four shows up and tells you, “Don’t worry, we’ll give you one fully integrated thing, you won’t have to think about anything” - that’s exactly when you should worry. If they hand you an all-in-one box where you never have to think, you’ve already been locked in.
Instead you say: “No - I want everything agnostic.”
- I want to change the language model as I please.
- I want every software integration - email, the invoicing system, holiday requests, payroll, customer relationships - all the software I already run - connected to this AI through open standards.
MCP is the standard that keeps you out of the cage
That standard is MCP (Model Context Protocol). MCP is going strong precisely because this is what it’s for: it’s how the data and context an agent needs to do its job - the emails, the invoices, the calendar events - flow in from pieces you compose yourself.
So the rule is simple:
- Don’t buy the fully-integrated everything.
- Buy software that is agnostic both toward the language model and toward the components of the agent.
- Explicitly ask that those components respect the protocol (MCP).
How to talk to a consultant
Don’t tell the consultant, “Build me the whole stack, sell me the complete solution.” That’s how you get taken for a ride.
What you actually go and buy is one of two things:
- Software that supports open standards, or
- The specific missing piece - for that, go to a software house that does system integration and tell them: “We use Claude Code, we use ChatGPT, we use Claude, we use this stuff. We need the connectors so we can handle our own data and our own internal services directly from ChatGPT or Claude.”
That’s the level you should be operating at. Don’t buy the entire stack.
And who teaches your team?
As for who teaches you all this - how to choose, how these standards work, how agents work, how to get your technical people up to speed quickly, with critical sense and some hope for the future - that’s what my corporate course Reskill is for.
It’s companies only. Write to me for more information. The program covers both the properly technical part - how agents are actually built, so you understand what we’re even talking about - and a hands-on part about working with agents day to day.
Stay agnostic, stay composable, and don’t let anyone sell you the whole box.