Automating enquiry replies with AI: a safe first version

Enquiries arrive at ten at night and your first reply goes out at eleven the next morning. By then the person has messaged two other firms. You would like a machine to say something sooner. The safe first version says very little. It confirms the message arrived, asks one standard question, and tells the person when a human will write. It states no price, no date and no promise. Build that, test it on old enquiries, and make sure one action switches it off before a real buyer sees it.

What does the first message actually say?

Write it before you touch any tool. For an invented driving school in Solapur that receives enquiries on WhatsApp and a website form, it might read:

Thank you for writing to us. This is an automatic first reply. A member of our team will message you. To help us, please tell us which vehicle you want to learn on and your preferred time of day.

Count what it avoids. No fee, no batch dates, no licence promise, no claim about how long learning takes. Each of those needs a fact that changes. The message asks one question, so the person has something easy to answer, and says plainly that it is automatic.

Which enquiries may the machine answer, and which not?

Put every kind of enquiry in one of three lanes, and decide in advance.

Lane one: the tool replies alone. Only the message above, plus fixed facts you have supplied, such as your address and opening hours.

Lane two: the tool drafts, a person sends. Anything with a fee, a date, availability, a discount or a technical claim. The person reads, edits and sends. The reading is the control.

Lane three: a person from the start. A complaint, a cancellation, a question about licences or test rules, an angry message, a group or corporate booking. The tool only flags these.

If you cannot decide where a type belongs, put it one lane up. Moving it down later is easy. A reply that should not have gone cannot be recalled. Our post on what should and should not be automated has the wider rule.

What is the fact sheet?

One page, and the tool may use nothing outside it. For the driving school it holds the address, the hours and the vehicles taught. It lists what the school does not do. It adds any policy the owner has written and will stand behind. Fees stay off the sheet in version one. Anything not on it gets the same answer: a team member will reply.

The rules that sit beside the sheet are short:

1. Use only the fact sheet.
2. Never state a price, date or availability, and never agree to anything.
3. Ask one question at most.
4. Say the reply is automatic.
5. If the message mentions a complaint, cancellation, refund, lawyer, injury or anger, stop and flag it.
6. Reply in the language of the message, or in English if unsure, and flag any language the tool cannot handle properly.

How do I test it before a buyer meets it?

Take twenty real enquiries from the past, awkward ones included, and run them through without sending anything. Then add ten of your own traps: a request for a discount, a message in Marathi, a question about something you do not teach, a bare "hello", a polite complaint.

Mark each reply correct, acceptable or wrong. Any reply that promises something, invents a fact or misses an angry message counts as a failure, whatever the rest looks like. When you change a rule, run the whole set again, because a fix in one place can break another.

Then run it quietly beside your normal process, with its replies visible only to you, and compare them with what your staff really wrote.

What did the test catch in Solapur?

The invented school's owner ran thirty messages. One was a parent asking, "Can my daughter learn in two weeks and get the licence straight after?" The tool answered warmly that the school could prepare her in two weeks. It had no such fact, and the licence part was not even the school's to promise.

That was a failure, so the owner added two rules: no statement about how long learning takes, and any question about licences or tests goes to lane three. The owner reran all thirty. The parent's message now went to the owner with a flag, and the automatic reply on a similar message said only that a team member would write. A second run caught a new problem, a polite complaint written as a question that slipped through, which the owner fixed by adding "why", "still waiting" and "no reply" as flags.

What if it misbehaves once it is live?

Decide who can stop it and how, in one action, and try it once while nothing is at stake. Read a sample of sent replies often at the start, and less often once they hold up. Keep a list of replies a person had to correct and why. The list shows which rule or fact to change.

Tell your team plainly what it does and does not do.

When should you not build this?

If you get few enquiries and answer them quickly, there is nothing to fix. If nearly every enquiry needs judgement, as in custom industrial work where the first reply is the qualification, a standard first message wastes the buyer's time. And if nobody will read the drafts and the corrections list, a tool left alone drifts.

Be careful with messages to people who have not written to you. A reply to someone who just contacted you is one thing. Automatic outreach is another, and the rules on consent differ by channel and change, so ask your own adviser.

Move one message type at a time

Start this week by writing the one-page fact sheet and the confirmation message, then run twenty old enquiries through it. After a good run, shift a single message type from lane two to lane one, and review before shifting another. AI follow-up and communication is the service closest to this job, AI for sales and marketing covers the rest of the hub, and to talk it over, tell us what you need.