Who changed your model?

In a regulated firm, changing a model is a project. There's testing, there's sign-off, there's a paper trail. On November 3, 2026, a major AI vendor announced that its assistant will choose which model handles each part of a task.

I'm sure it's efficient. I'm less sure anyone's compliance team was in the room.

What changed

OpenAI's update (opens in a new tab) says its Dots assistant "can draw on different models for different parts of its work," now including a model called GPT-6.1 Terra. Fortune (opens in a new tab) reported that Terra "is smaller and uses less computing power than GPT-6 Astra," and that the router "will allow OpenAI to reduce how compute intensive Dots is."

To be fair to OpenAI, it describes Terra as faster, with "a nice personality upgrade." For most people using it, the router may well be an improvement. They get answers sooner and never think about which model produced them.

My point isn't about quality. It's about visibility.

The vendors say it themselves

This isn't one company's quirk. Anthropic's SEC filing (opens in a new tab) for its public listing says it plainly: "We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities."

I appreciate the candor. Compute is expensive and scarce, and vendors will manage it. Routing work to a cheaper model, retiring an old one, adjusting what a product can do: these are normal business decisions for an AI company.

They're just not decisions a regulated firm is used to having made for it.

Why it matters in regulated work

Think about what an approval actually covers.

Say a team tests an AI assistant that summarizes policy documents for client service staff. They check its summaries against a sample set, look at where it gets things wrong, write down the limits, and a named person signs off. That approval describes a specific system, behaving a specific way, on a specific date.

Now say that, a few months later, some of those summaries are produced by a different, smaller model. The results might be just as good. They might be better. But the approval no longer describes what's actually running. Nobody retested. Nobody signed. Nobody even knew.

So what, exactly, was approved?

Anyone who has worked inside a large financial institution knows the routine for a production change. What's changing, and why? Who tested it, against what? Who signed off? How do we roll it back if something goes wrong? I spent more than 12 years in-house at large organizations, and that routine can feel slow from the inside. It exists because "it seemed fine" is not an explanation anyone accepts after the fact.

An AI service that changes its own internals without notice quietly steps around all of that. Not out of bad intent. It's simply built for a world where the user cares about the answer, not about which machine produced it.

What to ask for

None of this means regulated firms can't use AI services that route work across models. It means the contract and the setup should give people back the visibility they'd expect from any other production system. Here's what I'd ask for.

1. Disclosure of which model handled each piece of work, in terms your own staff can read.
2. A log of it that your firm keeps, not just one the vendor holds.
3. Notice before a model your approved workflows depend on is changed, swapped or retired.
4. The right to pin a specific version for approved workflows, so the thing you tested is the thing that runs.
5. Or, where pinning isn't offered, the option to run an approved model somewhere you control, whether that's your own environment or a dedicated, private deployment.

If a vendor can offer the first four, many firms will be comfortable with routing for everything else. If it can't, that's useful information. It tells you which work belongs on that service and which doesn't.

Approval should mean something

Model routing will keep getting smarter, and most of the time nobody will notice. That's fine for drafting a birthday message. It's not fine for work your firm has to stand behind.

Approval is a human decision. It should still mean something the day after it's given.

If you can't tell which model did the work, you can't really say you approved it.