What an AI receptionist should answer after hours — and what it should not

The safest place to put an AI agent is the place where the current experience is nothing. Nobody is answering the phone at 9pm. The caller gets a ring-out, or a voicemail box that someone empties on Tuesday. Anything coherent beats that, which is why after-hours is where most sensible deployments start.

It is also where people overreach, because "it's only after hours" sounds like the stakes are low. The stakes are your reputation with someone who was motivated enough to ring you outside business hours. That is not a low-intent caller.

What it should answer

Are you open, and when

The single most common out-of-hours question, and the one voicemail handles worst.

Where are you, and do you cover my area

Facts that do not change and cannot be got wrong if you have written them down.

What does it cost, roughly

Only if you have a published price. If your pricing is "it depends", say that it depends and take a callback — do not let an agent improvise a number.

Book me in

The one that earns the deployment. It requires real calendar availability, not a promise someone reconciles in the morning.

Take a message properly

Name, number, what it is about, and how urgent. A structured message on the contact record beats a voicemail nobody transcribes.

What it should refuse

Write these as explicit refusals in the prompt, because the failure mode of a helpful model is being helpful about things it should not touch.

• Anything medical, legal or financial that is specific to the caller. "Should I come in tonight?" is not a question for a language model.
• Emergencies. Detect and redirect to the right number, immediately, in the first sentence. Then stop.
• Complaints and refunds. Take the details, promise a human, do not negotiate.
• Anything about an existing order it cannot actually look up. If the function is not wired, the honest answer is that a person will check in the morning.
• Discounts. An agent that can be talked into a price is a liability.

An agent that says "I can't help with that, but I've taken your number and someone will call you at nine" has done its job. An agent that guesses has cost you the caller.

Say that it is an agent

Disclose it in the opening line. It costs you nothing, it is required in some jurisdictions, and callers behave better when they know — they speak more plainly, they ask smaller questions, and they are far less annoyed when it hands over. Pretending is a short-term win and a long-term embarrassment, and it is the fastest way to end up in a screenshot.

The handover is the whole design

After hours there is nobody to hand to, which is the awkward part. You have three honest options, and you should pick one deliberately rather than discovering which one you built:

1. A callback task timed for when someone is actually in, with the transcript attached.
2. An on-call number, for the small set of situations that genuinely cannot wait.
3. A booked slot, which converts the problem into a calendar entry and is the best outcome when it fits.

In Convarza AI the callback becomes a real task with a reminder, on the contact record, assigned to whoever opens in the morning — not a note in a voicemail box. How that is wired.

Measuring whether it worked

Not "calls answered". Three numbers:

• Resolved without a human — the caller got what they rang for.
• Converted to something — a booking, a callback, a qualified lead.
• Repeat calls the next morning — the honest failure signal. If they rang back, the agent did not finish the job.

That third one is the one nobody reports and everybody should.

Start narrower than feels necessary

Hours, location, take a message, book a slot. Four jobs. Read the transcripts for two weeks, find the questions it kept fumbling, and add those deliberately. A narrow agent that never embarrasses you earns the right to do more; a broad one that improvises does not get a second month.

Point a test number at one and ring it yourself before you point anything real at it.