What an AI chatbot should and should not handle
The technology got good quickly. The judgement about where to point it did not. An honest map of both sides.
Chat assistants are genuinely useful now. They hold context, they sound natural, and for a large share of inbound questions they are faster than a human would be. That capability arrived alongside a lot of vendors overselling it, so it is worth being precise about the line.
Where it works well
- The same twenty questions. Opening hours, delivery times, sizing, compatibility, returns. Answered instantly and consistently.
- Qualification. Asking six questions in the same order every time and routing accordingly. Machines are better at consistency than a tired human on a Friday.
- Booking. A well-defined task against live calendar availability. This is the strongest use case I deploy.
- Out of hours. The alternative is a contact form nobody reads until Monday.
Where it should not be used
- Complaints. An upset customer wants a person. Automation reads as a brush-off and escalates the anger.
- Vulnerable or distressed people. This needs judgement the model does not have.
- Anything with real consequence. Medical, legal, financial or safety-critical answers.
- Negotiation. If price and terms are genuinely being agreed, a human should do it.
Say that it is a bot
Every assistant I build identifies itself in its opening line. Partly that is a compliance question. Mostly it is a commercial one: a customer who works it out halfway through a conversation trusts you materially less than one who was told upfront. Disclosure costs almost nothing and protects the thing that is hardest to rebuild.
Set escalation generously
The cost of an unnecessary handover to a human is trivial. The cost of a mishandled call or chat is not. Build the triggers wide and tighten them later once you have read a month of real transcripts.
