Main idea

"Does the AI make mistakes?" is a badly framed question: it does, just as the person handling the same files would. The question that decides lies elsewhere — what does an error cost here, and what do you get in exchange.

An agent structuring five hundred building-management emails occasionally attaches a quote to the wrong contractor, or merges two entries that weren't the same person. The error is repaired by reopening an email, and it is paid for with hours of thread reconstruction you no longer do. The same agent on banking documents has no such margin: the error isn't repaired by reopening a file, and no time saving offsets it. The tolerance threshold is therefore not a property of the model, which is the same in both cases — it is a property of what an error destroys there and of what it leaves recoverable.

The threshold also moves by design. Keeping a copy of every raw email in the repository, alongside the aggregated information, changes the nature of the error: it stops being a loss and becomes a detour, since the source of truth remains directly queryable. The structured system is then an accelerator, never a replacement — and it is precisely that way out that makes an error rate acceptable in a domain where it wouldn't be without it.

Why it matters

This gets you out of the argument of principle about model reliability, which never concludes because it looks for a universal threshold where there is none. The question to ask before delegating any information processing becomes concrete and local: what does an error destroy in this material, and by what route do you get back?

It also gives a lever rather than an observation. You don't merely measure an error rate: you make the domain more tolerant by keeping the raw source accessible, which authorizes a delegation the same rate would forbid elsewhere.

Nuances and limits

A domain isn't homogeneous. A building-management email can carry an appeal deadline or a formal notice, and the domain's average threshold then masks a subset where the error is expensive. The criterion has to descend to the type of information, not stop at the file.

And the way out only serves if you know you have to take it. An error no signal reveals leaves the raw source intact and useless, since nobody will go and consult it.

Open questions

• Who sets the threshold when the person delegating the processing isn't the one who will bear the error?