What do you do when your AI model is retired?

Rerun your own past work on the new model, compare every answer with the old one, fix what differs, and switch weeks before the date. It's a Sunday job if you've prepared 3 things: one door in the code, stored inputs, and a second option you already know.

The short, polite email

The provider wrote to say the model my document product used would be retired in a few months, with a newer one recommended. After the date, calls to the old one would fail. I wrote the date on the calendar, and beside it the date 4 weeks earlier, by which the move had to be done.

Nothing about a product with a model inside it is entirely yours. Somebody else decides what it costs, how fast it answers and how long it exists. None of that's malice; old models are expensive to keep. But every one of those decisions arrives with a notice period somebody else chose.

What made it easy

• One door. Every call to a model goes through one small part of the code. I changed one setting in one place. (There are 2 kinds of lazy. This is the good one.) Products with the model's name typed in place after place take a week to move, and you miss one.
• Stored inputs. Because the product runs on a queue, I had 3 months of real documents and the old model's answers in a table. I ran the new model over all of them.
• A second option. The cheap model I already routed to, tested on my own work, meant I was never choosing in a hurry.

What the comparison found

One difference that mattered. The new model wrote dates in a different format. My check had been written against the old format, so some correct answers were being rejected, and a few wrong ones would have slipped through. I fixed the check, reran it, found nothing else, and switched on a Sunday evening, weeks early. No customer saw any of it.

Who Pays for the AI? has the chapter, including the price rise the year after.