AI quality assurance: a glance is not a review
Check AI output in three separate passes: format, source, expert judgment. The sender stays responsible. With a review table and a rule for model switches.
Checking AI output means checking format, source and substance separately, and whoever passes a result on owns it. In a lot of companies, quality control for AI means "human in the loop" instead: someone takes a quick look, ticks a box, and the result goes out. The assumption behind it is that the models are good enough now and checking just costs time. There are two ways to get AI wrong, trusting it too much or using it too little, and the checkbox is the first one. Pass things on unchecked and you become a pass-through. At SUMAX every AI call runs through a layer of our own that lets us switch models any time, and that's exactly why review can't be a one-time sign-off.
A clean format proves nothing.
A text can read smoothly and still contain a hallucination, a made-up statement that sounds plausible. A table can come back in the right format and still put things in the wrong rows. Formal checks catch missing fields, invalid formats and language patterns. They don't catch wrong content. If all you checked was the format, the real work is still ahead of you. Check form and content separately, and never treat a passed format check as approval.
Set the standard before you compare models.
Every model comparison and every pilot starts with the review standard. It defines what an output has to contain, which statements need a source and which errors are serious. Good test tasks include normal inputs as well as unclear cases and known edge cases from your own area.
The people who'll work with the tool help set that standard. They know which error has consequences on the job and which one nobody would notice. Without a standard, the comparison goes to whichever model sounds better. Write the standard down before the first model sees a task.
Three checks, and none replaces another.
An example: a draft claims a certain method works for companies of any size. Nothing is wrong with the sentence itself, and it passes the formal check. The source check establishes whether the cited source supports exactly that claim, only says something similar, or exists at all. Only then comes the expert question: does the claim hold in this specific case, and which conditions or exceptions are missing?
Another AI pass can point to issues. It isn't proof, because a model trained on similar data can repeat the same mistakes. AI detectors estimate whether a machine wrote a text, and they say nothing about whether it's correct. A statement is only cleared once it has passed all three checks.
A statement is only cleared once it has passed all three checks.
A review table for one statement.
| Check | Question | catches | misses |
|---|---|---|---|
| formal | Are fields complete, format and sentence OK? | missing fields, invalid formats, language patterns | wrong content |
| source | Does the source support exactly this claim, only something similar, or does it not exist? | made-up or stretched sources | whether the claim holds in this case |
| expert review | Does the claim hold in this specific case? Which conditions and exceptions are missing? | statements that are true but don't fit here | format errors and bad sources, that's what the first two are for |
| another AI pass | Where could something be wrong? | hints for the reviewer | proof, because a similar model can make the same mistakes |
| AI detector | Did a machine write this? | an estimate of where the text came from | whether the text is right |
A statement is cleared only after the first three rows. The last two help with the review, they don't replace it.
After every model switch, the old sign-off no longer counts.
A new model, a different prompt or changed input data can shift results without anyone noticing. As of this revision (October 2026), tens of thousands of AI calls a month run through our own layer. That layer makes switching models easy, and that's exactly why every change needs the same checks again, with the same test tasks and with feedback from actual use.
Every call is also booked with its cost. That lets you put a cheaper model side by side with the current one. The booking shows what it costs. Only the same checks on the same tasks show whether it delivers the same quality. Keep your test tasks and run them again after every change.
Review rules belong in one place.
Review rules are process knowledge. If they live only in a few heads or scattered spreadsheets, everyone checks by gut feeling, and the knowledge walks out with the next staff change. Process knowledge belongs in one place where everyone who reviews reads the same rules.
People learn to review best on real tasks from their own area. That's how we built our training too. The responsibility still stays with the person who passes the result on. Put the review rules where the tools are.
An AI result isn't done until a person has checked it and stands behind it. Take the AI output that leaves your company most often, put together a handful of real test tasks, unclear cases included, and write down for each what a correct answer has to contain. How rules and approvals fit into automated workflows is on the AI automation page.