Let’s talk about the work you made with a machine that stops existing on Friday.
On the twenty-third of October, OpenAI switches off seven models at once. gpt-4o, gpt-4, o1, o1-pro, o3-mini, o4-mini, and the thing that started all of this, gpt-3.5-turbo. Google pulls Gemini 2.5 off Vertex three days earlier. Anthropic’s rounds were already done in June and August. Twenty-three and twenty, two dates, and a stack of behaviour that a lot of real businesses are built on top of.
You’ve heard the developer version of this. Dependency risk. Pin your versions, build an abstraction layer, keep an eval set. All true, all useful, and by now it’s the most repeated advice in the industry.
Here’s the version nobody is saying out loud.
Picture a small consultancy in March. They price a tender using a model. Not illegally, not lazily — they use it the way everyone uses it now. Feed it the cost schedule, the assumptions, the risk register, ask it to stress the numbers, take the output, argue with it for two hours, and submit. The tender wins. Eight months later there’s a dispute about scope and the client’s lawyers ask a simple question. Show us how you arrived at that figure.
The consultant opens the folder. The spreadsheet’s there. The emails are there. The output is there. And the machine that produced the reasoning is gone. Switched off on a Friday, notice posted on a documentation page aimed at engineers, no archive, no way to re-run it, no way to show what it actually said when it said it.
That is not a maintenance event. That’s the disappearance of the evidence trail.
Think about what professional accountability has always rested on. Not the answer. The route to the answer. An auditor doesn’t accept your number, they accept your working. A court doesn’t want your conclusion, it wants the chain. Medicine, engineering, accounting, law — every profession that carries liability carries it because someone can go back and reconstruct the judgement. Show me how you got there.
Reconstruction assumes the instrument still exists. A slide rule, a spreadsheet, a statistical package from 1998 that someone keeps on an old machine precisely because a regulator might ask. We understood this. We built an entire discipline of record-keeping around the fact that tools must survive their own use.
Then we adopted tools that get deleted on a vendor’s schedule, and we forgot to notice.
The retirement notices are written in the language of housekeeping. Deprecated, sunset, end of life. It sounds like cleaning a garage. What’s actually happening is that behaviour is being destroyed. A model isn’t a file you keep. It’s a set of weights and a serving configuration and a dozen undocumented decisions about how it handles an edge case, and when the endpoint goes dark, all of it goes. Nobody can reproduce it. Not you. Not your auditor. Not a court.
And notice the asymmetry, because this is the part that should make you angry.
The vendor keeps the logs. Your prompts, your outputs, your corrections, the whole session history — that sits in their system, subject to their retention policy, available to them under their terms. You get a PDF and a screenshot. They get the transcript.
So in any dispute, one side has the record and the other side has a memory of the record. Guess which side that favours.
Now the practical question, because I’m not interested in a lament.
If a machine helped you decide something that matters, you have to capture the decision at the moment it’s made. Not afterwards, because afterwards you’re reconstructing from whatever survived.
At the time. The exact model and version string. The inputs, verbatim. The output, verbatim. What you changed and why. Who approved it, by name. And a hash of the whole package, timestamped, so the record can’t be quietly edited later.
That’s not paranoia. That’s the same discipline a bookkeeper applies to a receipt.
And then the harder part. Put a human’s name on the decision. Not as a rubber stamp on the machine’s recommendation — as the author of the judgement, with the reasoning written in their own words, in the record, next to the machine’s contribution. The person is the one who carries it. That’s what makes it a decision rather than an event.
Because here’s where this lands in about eighteen months. Regulators are already asking deployers of high-risk systems to keep logs and demonstrate oversight. Auditors are starting to ask for the working, not just the number. And the first serious case where a company says “the model recommended it” and cannot produce the model, cannot produce the reasoning, and cannot produce a person who owned the call — that company is going to discover that “the algorithm did it” was never a defence. It was a delay.
You cannot outsource judgement and then outsource the receipt.
The thing I keep coming back to is smaller than all of this. Somewhere this week a person is finishing a piece of work they’re proud of, and the tool that helped them get there is being retired in five days, and nothing in their process will preserve how it happened. The work will stand. The proof of the work will evaporate.
That’s a strange kind of loss. Not the work disappearing. The account of it.
So before Friday, do the boring thing. Open the files. Write down which models your business actually depends on, and what decisions they touch. Build the record that survives the switch-off. Put a name on the judgement.
The machine can help you think. It cannot be your alibi, and it will not be your witness.
You’ve got this.