This is published, not speculation #
You do not have to take anyone's word for it, because the providers write it down. Anthropic publishes a model lifecycle of active, legacy, deprecated and retired, states that requests to retired models will fail, and lists every retirement it has made with dates. OpenAI publishes the same kind of page, with a dated table of what has already been shut down.
Both also publish the minimum notice they will give before a model is retired.
60 days
Anthropic, at least this before a publicly released model is retired
Anthropic, Model deprecations
6 months
OpenAI, at least this for generally available models
OpenAI, Deprecations
3 months
OpenAI, at least this for some specialised variants
OpenAI, Deprecations
about 2 weeks
OpenAI, as little as this for preview models
OpenAI, Deprecations
Read either page for two minutes and the position is clear. This is not a risk somebody is warning you about. It is a scheduled property of the technology, run as a published policy, with a list of things that have already stopped working.
Four different things change, and they behave differently #
| What changes | How you find out | What it costs to deal with |
|---|---|---|
| A model is retired | Published in advance with a date, and a notification if somebody is watching the account | Predictable, and cheap if handled before the date rather than after |
| A model is replaced and behaves differently | Nothing breaks. The output is just not quite the same shape any more | The awkward one, because it needs somebody to notice |
| A product around it changes | A new screen, a moved setting, a changed export, a permission that now needs granting | Small and frequent, and it adds up |
| A system you connect to changes | Your accounting package, your inbox, your CRM, your file storage, on their own schedule | Usually the most common source of real breakage |
The fourth row is worth dwelling on, because businesses expect the AI part to be the fragile one. In practice the connections to ordinary software are what break most often, for the ordinary reason that every one of those suppliers is shipping changes too.
The failure that is hard to see #
A hard failure is the good case. Something stops, somebody notices on the day, it gets fixed.
The version that hurts is drift: the summaries get slightly blander, the extraction starts missing a field that used to come through, the tone shifts. Nothing errors. The work carries on. And by then nobody is reading closely, because the whole point of the automation was that they stopped having to.
The counter is unglamorous. Somebody looks at a sample on a schedule, against an example of what good looked like when it was signed off. A business that keeps a handful of reference outputs from week one can answer the question at a glance. One that does not is reduced to arguing about whether it used to be better.
What this means for the money #
It means the running cost never reaches zero, and a proposal with a build price and no arrangement for keeping it working is incomplete. That is the fourth part of the cost set out in how much AI automation costs, and it is the one most often left off.
None of this is unusual. A business already accepts that a van needs servicing, that the accounting software will change its screens, and that the phone system will need attention. The only reason it lands badly here is that AI is sold as a thing you buy once.
What to expect of a supplier #
Four questions, and they are fair to ask before anything is built.
-
Who is watching for this?
Somebody has to be reading the deprecation notices and the release notes. If the answer is nobody, the answer is you.
-
How would I find out?
A message when something needs doing, rather than a phone call from you when the invoices stopped going out.
-
What is the arrangement for fixing it?
Named, with a cost attached, agreed before it is needed rather than negotiated in the middle of a problem.
-
What happens if it breaks and nobody is available?
The honest answer is always that the work goes back to being done by hand for a while. The question is whether anybody still remembers how, which is a real argument for not deleting the manual process the week it stops being used.
The full list of things worth asking before signing is in questions to ask an AI supplier.
What reduces the exposure #
- Do not build the business on a preview. The published notice periods are shortest for exactly the newest and most impressive things.
- Keep the accounts in your own name. When something has to be changed quickly, the worst position is needing somebody else to log in for you. That is one of the forms of lock in that only becomes visible under pressure.
- Keep a person in front of the irreversible step. Drift that reaches a draft is an annoyance. Drift that reaches a customer is an incident, and the line that prevents it is in what should always wait for a person.
- Prefer boring over clever where you can. The more moving parts a piece of work depends on, the more often something underneath it will move.
The honest conclusion #
Anyone who tells you their automation will keep working indefinitely with no attention has either not run one for a year or is not telling you the truth. The reasonable position is that this is maintained software, that maintenance is a line in the budget rather than a surprise, and that the work is still worth doing when the job underneath it is big enough. Whether yours is, is what the cost of admin time calculator is for.