Most of the questions a business receives are the same twenty questions. Opening hours, do you do this, how much roughly, where are you, can I bring my own, how long does it take.
Answering them is not skilled work, but it has to happen quickly and it happens at nine in the evening. That is what an AI agent is genuinely good at.
Where it works
- Questions with a settled answer. Hours, location, what you do and do not offer, how the process works.
- Questions asked in a hundred different ways. This is the real advantage over a menu of buttons: somebody can ask in their own words, or in another language, and be understood.
- Collecting what you need before a human takes over. What they want, when, their number. Handed to you complete rather than as "hi, are you there?"
Where it does not
- Anything where being wrong is expensive. A quote, a delivery date, medical or legal or financial advice, whether a specific case is covered. If a confidently wrong answer would cost you money or trust, that question needs a person.
- Anything it cannot look up. An agent that does not have access to your calendar cannot say whether Tuesday is free. It can say it will find out — which is a good answer — or it can guess, which is the failure everybody has now experienced.
- Anything a customer is upset about. A complaint met by a machine is a complaint that becomes a review.
The design decision that matters
Not "what can it answer" but "what does it do when it does not know?"
There are two possible behaviours and only one is acceptable. It can say it does not know and pass the conversation to a person, with what it has already collected. Or it can produce something plausible.
Every agent worth having is built around the first. Ask any supplier to show you what happens when a customer asks something outside its knowledge. If they demo only the happy path, you have not seen the product.
The things people forget
It needs a way out. A visible route to a human, at every point, without arguing.
It has to say what it is. Letting somebody believe they are talking to a person is a bad idea both ethically and practically — they find out eventually and feel deceived.
It handles personal data. Whatever it collects is subject to the same rules as anything else you hold, and conversations often contain more than people intend. Ask what is stored, for how long, and where.
It goes stale. It answers from what it was given. Change your prices and tell nobody, and it will keep quoting the old ones — confidently. Whoever maintains it needs to be named before it launches, not after the first wrong answer.
Where to start
Take a week of your actual messages. Sort them into "same answer every time" and "needed thought". The first pile is what to automate, and its size tells you whether this is worth doing at all.