Shutting down an AI system: keeping it is a decision too
Shut down an AI system when usage, quality, upkeep or a better alternative say so. Four checks for every live system, and a shutdown order that strands nothing.
An AI system that nobody ever reviews for shutdown keeps running out of habit. In most companies a tool only goes away when it breaks or the license runs out, and until then running is the default that needs no reason. I see it the other way around: keeping a system running is a decision too, and it needs the same review as the launch. Shut it down when usage, output quality or the effort of keeping it running no longer justify it, or when an alternative does the job better. At SUMAX every AI call goes through our own layer between our applications and the models and gets booked there with its cost. That's why we can see, tool by tool, whether something is being used.
Falling usage is a prompt, not a verdict.
Falling usage is usually the first sign. On its own it explains nothing. The task may come up rarely, or access may be awkward. Or the output needs so much rework that nobody bothers with it anymore.
As of this revision (October 2026), our layer handles tens of thousands of AI calls every month, and most of our people use our own tools. A total still says little about a single tool. When one drops off, that's a reason to talk to the people who were supposed to use it. Read usage per tool and talk to the people before you judge.
When review eats the savings, the system is done.
The effort behind an AI system sits mostly in checking, correcting and maintaining it. When checking the output takes more time than the system saves, its value no longer justifies running it. Nothing gets passed on until someone has checked it, so you can't just skip the review to make the numbers work.
Then the options sit side by side: fix the workflow, take an off-the-shelf product, build it yourself, or go back to doing it by hand. Build vs. buy cuts both ways here. At SUMAX we've replaced several purchased tools with our own, and a tool we built can just as well give way to a standard product if that fits better. The effort already sunk into it is part of the weighing, but it doesn't get to decide. Compare review effort with savings, and treat your own tool the way you'd treat one you bought.
The tool goes, the task stays.
Settling the alternative means the task stays and only the tool goes. Write down which task the system handled, why it's ending and what replaces it. Process knowledge belongs in one place, not in individual heads. When it's scattered, the knowledge of what a tool did can disappear with the tool, and when the need comes back, somebody starts from zero.
Technically, switching off is rarely hard when access and the connection to the models sit in one place. The hard part is everything that depends on the system. Write the task down before you switch the tool off.
Shutting down has an order.
When an AI system goes off, dependent workflows and interfaces lose their source unless they were moved over first. So the order matters:
- Map the dependencies: which workflows, interfaces and reports rely on the system?
- Have the replacement path ready, so the task keeps running the next day.
- Back up the data. Deletion and retention duties for personal data still apply after the end, so settle them beforehand.
- Tell users when the system goes and what takes its place.
- Remove access and interfaces.
- Document the decision and the reasons.
Follow that order and you can switch a system off without a workflow standing still the next morning.
Keeping a system running is a decision too, and it needs the same review.
Four checks for every running system.
This table applies to systems that run well, too. Those are exactly the ones nobody else checks.
| Check | Question | Points to shutting down |
|---|---|---|
| Usage | Do the people it was built for use it? | Usage drops, and talking to them turns up no fixable reason. |
| Output quality | Does the output get used without heavy rework? | Output needs so much rework that nobody uses it anymore. |
| Upkeep | Do review, correction and maintenance take less time than the system saves? | Review eats the savings. |
| Alternative | Would a better workflow, an off-the-shelf product, your own build or doing it by hand handle the task better? | An alternative is better, and the replacement path is ready. |
If none of the four points to shutting down, the system keeps running, and now that's a decision with reasons behind it. Set a date for every system when these four questions get asked.
A good system can take the question.
Shutting a system down doesn't mean launching it was wrong. Tasks change, and better models and tools come along. Replacing tools is part of normal operations. If you never ask the shutdown question, you pay for tools nobody needs anymore and lose track of which system carries which task.
My position: every AI system has to be able to take the shutdown question, including the most popular one. Pick the AI tool you haven't looked at for the longest time and run the four checks on it. How to make cost per tool visible is in AI API costs: know them per call.