Your Product Needs a Retirement Policy

Running AI infrastructure means handling models and everyone is always excited about the new models. There's a new name in the catalog, marketing can talk about new capabilities and typically there's a small burst of activity because customers want to try it. The real work (you know, the not exciting stuff) is what to do with the models that are reaching end of life.

Models reach end of life all the time. A provider releases a newer version. The older one gets too expensive to keep running. The terms change. Sometimes the provider gives you months of warning. Sometimes you find out because someone was looking for an entirely different thing and happened to notice a tiny “legacy” label.

Adding equals fun, retiring is the actual product work.

the catalog is not a list

I used to be a librarian, and libraries have a wonderfully unglamorous practice called weeding. Taking a book off the shelf is not just adding a book in reverse. You check whether anyone still uses it, whether another edition replaces it, and whether removing it leaves a hole in the collection.

An AI model catalog needs the same kind of unglamorous care.

A customer may have typed a model name into an application six months ago and never thought about it again. That name is now part of their product, whether they think of it that way or not. Deleting it from our catalog does not tidy up a list. It breaks something on the other side that we may not be able to see.

You need to stop asking, “How do we remove a model?” and start asking, “What have we promised to the customers still calling it?”

“This is going away” is not useful on its own. Customers need to know what to use instead. Every retirement needs a named replacement.

The old name should not suddenly stop working because we cleaned up our side. It can lead to the replacement while the customer moves on their own schedule. There is still a change, and model changes can affect tone, formatting, speed, and all sorts of things people have quietly come to depend on. But a deliberate change is very different from an unexplained failure.

Then there is the notice period. Not a vague line buried in a changelog after the fact. A real window before the change, on the surfaces people actually look at. If the provider gives us very little warning, we cannot manufacture more time, but we can at least say plainly that the deadline came from them. Making the customer discover it through a broken call is not communication.

There is also a choice hidden inside the model name itself. A general name (alias, if you will) like chat can mean “give me the current model.” A specific model name can mean “I chose this version because I need its output to stay consistent.” Both are reasonable. What is not reasonable is making customers guess which promise they bought.

None of this is particularly clever.

Name the replacement. Give people warning. Keep their calls working. Check the new path before removing the old one. The technology is the least interesting part.

What's harder to see coming is everything underneath those four rules. You cannot warn anyone about a retirement if you cannot tell who is still calling the old name, so the policy is only as good as your usage visibility. Keeping an old name alive costs money. Someone is paying to run a deprecated model so a customer's six-month-old integration doesn't break, and someone needs to own the decision about how long that subsidy lasts.

Some of this can be automated: the aliasing, the redirect, the usage alerts. Some of it can't. A provider gives you six weeks of notice and you have exactly one option, which is to move fast and communicate clearly.

Building the machinery is one thing. Productizing it is another.

Every AI platform is racing to add models right now. New things are fun, customers ask for them, and a growing catalog looks a lot better in a product update than “we wrote down what happens when the old thing goes away.”

But the second sentence is the product work.

The launch tells customers what they can start using. The retirement policy tells them whether they can safely depend on it.