The best answer a support bot can give is "I don't know"

There's a particular way AI support bots fail that's worse than being unhelpful: they answer confidently, and they're wrong. A customer follows the instructions, something breaks, and now you've spent trust you didn't have to. A bot that had simply said "I'm not sure — let me get a human" would have done less harm and more good.

We learned this building an AI support assistant for our own products. The hard part was never making it sound knowledgeable. The hard part was making it honest.

Confidence is not competence

A general-purpose model will happily generate a plausible answer to almost anything, because that's what it's optimised to do. Point it at customer support and, unchecked, it will invent a setting that doesn't exist, cite a step from a different product, or reassure someone about a risk it hasn't actually assessed. It's not lying — it has no idea it's wrong. That's precisely the problem.

Ground it, then let it admit the gaps

The fix isn't a smarter model; it's a disciplined one. The pattern that works:

• Answer only from your own material. The bot should draw from your documentation, manuals and known-good articles — not its general training — so answers reflect how your product actually behaves. (This is what "retrieval-augmented generation" means in practice: fetch the relevant source first, then answer from it.)
• Cite where the answer came from. If it can't point to a source, that's a signal it's guessing. Make it show its working.
• Make "I don't know" a first-class outcome. The bot must be allowed — encouraged — to say when a question is outside what it can safely answer, and to hand off cleanly to a person. A graceful escalation is a feature, not a failure.
• Draw hard lines it won't cross. It never asks for passwords or OTPs, never invents account-specific facts, and stays on the topic it's there for.

Why this is the commercially smart choice, not just the safe one

It's tempting to chase the impressive demo — the bot that answers everything. But in support, the metric that matters is resolved without a mess, and an honest bot beats a confident one on that every time. It deflects the genuinely answerable questions instantly, and it routes the rest to a human before damage is done, with the context already gathered. Customers forgive "let me get someone who knows." They don't forgive being confidently misled.

An AI assistant that knows the edge of its own knowledge is slower to wow in a demo and far better to live with. That's the trade we make on purpose — and the one worth insisting on from anyone selling you a support bot.

If you're weighing an AI assistant for your product or helpdesk, that judgement — where automation helps and where it must step aside — is exactly the conversation to have first. Talk to us.