On October 7, Elon Musk posted a note about Grok Bot, the agent product from SpaceXAI. “Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.”

Grok Bot went into beta in August, and according to Vellum’s breakdown it ran only on xAI models until then. A company that builds its own models just wrote down that it will send work to a competitor whenever the competitor does it better.

Satya Nadella said something similar from the distributor’s side. In an interview with Deirdre Bosa, he acknowledged that OpenAI still dominates Microsoft’s API usage in absolute terms. Then, as Karl Mehta quoted him, he said that because Microsoft’s platform is open, “we’re not locked into any one model.” A customer, he added, can deploy Copilot and bring their own model.

If you buy AI for a company, take the obvious conclusion. When a model maker and the biggest reseller of another one both refuse to marry a single model, you shouldn’t sign as if you had. That gets fixed in procurement before it gets fixed in engineering.

What Musk left out

No dates. No list of which tasks go to which provider. Nothing on whether users will see which model did the work, or whether outside models will burn through plan limits faster. Tbreak lists the same open questions, and the last one is the one that matters to you: what data policy applies when a task leaves for a third party.

That’s your side of the counter. You bought a product with a data promise attached, and the vendor now decides, task by task, who receives the work. If your contract says “our model” and the product routes elsewhere, your data terms may describe a system that no longer exists.

Four clauses that keep the exit open

Before renewing any AI contract, this is what I’d ask for.

1. A list of the models and subprocessors that touch your data, with written notice before one is added.
2. Data terms (no training on your data, retention, location) that hold for whichever model receives the task, not just the headline one.
3. Pricing per unit of work, such as a reconciled invoice or a resolved case, with a cap. Tie the price to one specific model and you have no defense when the vendor moves the rate. I covered how Microsoft charges for the layer above the model, and how GitHub Copilot doubled one model’s multiplier the same day.
4. The right to take your prompts, traces, memory and test cases with you when the contract ends.

None of the four depends on which model wins. I wrote not long ago that the runtime is a commodity now. If models turn interchangeable too, what stays with you is whatever you wrote into the contract.

Without evals, “best model” is an opinion

A router decides using a metric that isn’t yours. Musk’s “best outcome” is measured by Musk. Yours is measured by a set of real cases from your operation, each with a known correct answer. That’s your evals.

With them, switching models takes an afternoon. You run the current model and the candidate over the same cases and compare. Without them, you decide on a vendor demo or on price, and both are guesses.

My order of work is short. First, which processes matter. Then twenty real cases from each, with the correct answer attached. Only then, which contract lets you move each process to another model.

Where we come in

The first deliverable of AI Maestro is the Process Reality Map, a picture of how your company actually works. It’s the list of processes those cases come from, and also the list of data each process touches, which is what you need to read an AI contract with some judgment. The stage runs on the tools your company has already approved, so you buy nothing to start.

If you have an AI renewal coming up, ask this week for the list of models and subprocessors that currently touch your data. If the vendor can’t produce it quickly, you know how much control you have.

Map your processes before you renew the AI contract