Four things to settle before opening an AI agent to your clients
What decides an agent's quality: what it says when it doesn't know, what it relies on, when it hands over, and where the questions asked of it go.
In this article4 sections
A wrong answer given confidently costs more than no answer
A language model doesn't distinguish between knowing and guessing: when it lacks information, it still produces something plausible, well written, and wrong. It even tends to be more assertive when it's wrong.
The client doesn't make the distinction: they remember being given wrong information, and they blame the company, not the machine.
The safeguard isn't a setting, it's a building rule: the agent answers from your documents and nothing else, and every answer comes with the passage it came from. What isn't in your texts doesn't come out of its mouth.
An agent is never better than the documents it's given
In most companies, information exists in three versions: the one in the shared file, the one on the old intranet, and the one in someone's head.
As long as those three contradict each other, no agent can answer correctly: it will choose, and nobody will know why.
The clean-up work everyone keeps putting off therefore becomes the precondition for the project. That's the bad news. The good news is that it serves more than the agent.
The right reflex isn't knowing everything, it's knowing when to hand over
Agents are often judged on what they handle alone, as if every question passed to a human were a failure. It's the opposite.
Users are perfectly fine being told "I don't know, let me put you through to someone". What they can't stand is going round in circles for ten minutes with something that refuses to let go.
A corollary on deployment order: start internally, on the questions your own teams ask ten times a week. It puts the documents in order and shows how the agent behaves when it's wrong, while the stakes are low.
The question you type into an agent is data that goes somewhere
Your employees are already pasting extracts from documents into consumer tools, with no bad intent and without anyone having decided it.
In Switzerland, the law doesn't require hosting your data here: it requires knowing what you entrust, to whom, and being able to prove it. Many companies discover the question the day a client asks it in writing.
What you need to be able to produce that day fits in four lines:
- Which model, and hosted where
- What is sent to it
- What is kept, and for how long
- Who has access to it
None of these four points concerns the model chosen, and that's what surprises people most. What makes a good agent is what you give it, what you forbid it, and the way it stops: three decisions taken before the first line of code.
Ce sujet touche à notre expertise AI Agent & Chat.
We start with a conversation, not a quote.
Voir l'expertiseRead next
