This is the question that exposes a seller of promises fastest. Not because the right answer would be to refuse a lot, but because a professional who has run real automations necessarily has limits, and can explain them.
Here are ours. They do not come from distrust of the technology: they come from the ratio between what an error costs and what supervision costs.
Decisions about people
Rejecting an application, appraising an employee, deciding on a sanction, refusing a private individual's file. An AI can prepare, summarise, compare, flag. The decision stays human, and it must remain explainable by the person who carries it. This is not only an ethical stance: it is also what the European regulatory framework expects from systems that weigh on people's rights.
Irreversible gestures towards the outside
A mass send, a permanent deletion, a transfer of funds, a post on the company's accounts, a report to a third party. The technology allows it perfectly well. We do not do it without validation, because there is no way back and the time saved is trivial next to the cost of one unfortunate send.
We do not automate what cannot be undone.
Difficult conversations
A dispute, a serious complaint, announcing a major delay, a termination, an angry customer. A generated reply will be correct and polite, and that is precisely the problem: it will read as an evasion. These moments are where the relationship is decided. AI serves there to prepare the material, never to speak in your place.
Statements that commit the company
A negotiated price, a warranty, regulatory advice, a delivery commitment, a tender response. A model produces plausible text; plausible text about a price commits the company just as surely as accurate text. Anything with the value of a commitment goes through a source of truth and a validation.
Processes the company cannot describe
This is our most frequent refusal, and the least expected. When two people handle the same file in two different ways and nobody can say which is right, there is nothing to automate: there is a rule to write. Automating at that stage freezes an inconsistency and repeats it at scale.
We prefer to say so at the first meeting, even if it pushes the project back by a month. That is more honest than invoicing development that will have to be redone.
What we automate without hesitation
The opposite list is long, and that is where the value sits: classify, extract, summarise, reconcile, draft, check a file for compliance, follow up, update, monitor, alert, produce a report. Anything repetitive, reversible and verifiable belongs out of human hands. Our AI agents work in exactly that perimeter.
Refusing a few uses is not timidity. It is what makes the rest defensible.
