How to set up an AI receptionist in an afternoon

Vendors advertise five-minute setup, and the configuration genuinely does take about that long. What takes an afternoon is getting the thing to answer well, and most of that work is preparation you can do before you have chosen a provider, whichever one you land on after comparing your options.

Here is the order that works.

Before you sign up: gather the source material

These systems are built from what you give them. Vague inputs produce vague answers, so this hour matters more than any other.

Write down your twenty questions

Ask whoever answers your phone to list the questions they field most often. You will get to twenty quickly and it will be the most valuable document in the project. Include the awkward ones: pricing, availability, whether you take a particular insurance, why the last invoice was what it was.

Write the answers, precisely

This is where most setups go wrong. 'Pricing varies' is useless. 'A standard consultation is £80 for thirty minutes; complex matters are quoted after an initial call' is something a system can actually say. Where an answer genuinely depends on circumstances, write the rule rather than the outcome.

Decide what it must never do

Just as important. Common entries: do not quote for anything over a threshold, do not give clinical or legal advice, do not discuss an existing complaint, do not confirm whether a named person is a client. Write these down before you start, not after something goes wrong.

Decide your escalation rules

Who gets the call when it needs a human, at which hours, and what happens if they do not pick up? A fallback that dead-ends is worse than not offering a transfer at all.

How do you configure an AI receptionist?

1. Point it at your website. Most systems ingest your public pages as a starting knowledge base. It will get the basics right and some details wrong, which is what the next step is for.
2. Paste in your twenty answers. These should override anything scraped. Where the website and your document disagree, your document wins.
3. Add the prohibitions and escalation rules.
4. Pick a voice and write the greeting. Say it out loud before you save it. Greetings that read well often sound stilted spoken.
5. Set the recording announcement. See our guide to call recording consent for what it needs to cover.
6. Connect your calendar and POS. Check what Hilda integrates with, and confirm appointment types and durations map correctly.

How do you test an AI receptionist before going live?

Almost everyone tests by asking the questions they configured, hearing correct answers, and going live. That tells you nothing. Test the edges:

• Ask something genuinely outside its knowledge and check it says so rather than inventing an answer.
• Interrupt it mid-sentence. Real callers do this constantly.
• Ask two things in one breath: 'what time do you close and do you take walk-ins?'
• Ask for a person immediately, before it finishes its greeting.
• Give a deliberately unclear answer when it asks for a date. 'Sometime next week, maybe Thursday?'
• Ring from a mobile in a noisy place, not from your desk on a good line.
• Ask an emergency question if you are in a field where those occur, and confirm it escalates instead of booking.

Go live gradually

Resist switching everything over at once. The sensible sequence:

1. Out of hours only, for a week. Nothing changes during your working day. You find out how much after-hours demand you have.
2. Add overflow, for a week. It picks up only when your line is engaged or unanswered after several rings. Your team keeps first refusal.
3. Then decide about daytime. Many businesses stop here deliberately, and that is a perfectly good outcome.

The week-one habit that matters

Read every transcript for the first week. It is an hour of work and it will be the most informative hour you spend on the project, not because the system will be badly wrong, but because you will discover the questions your customers actually ask, as opposed to the ones you assumed.

Most businesses find two or three questions in the top ten that never appeared on their prepared list. Add them, and answer quality jumps.

After that, a monthly skim is enough. Look for calls that ended in a transfer, and calls where the caller repeated themselves; both are reliable signals of a gap worth filling.