The calls an AI receptionist should hand back to a person

Three calls, three different answers

Consider a hypothetical villa-rental company in Goa with eight properties and a small front office. An illustrative example of three calls arriving on the same Saturday evening:

Call one: a routine booking. A caller asks whether a three-bedroom villa is free for two nights next month, the rate for six adults, and whether breakfast is included. The answers sit in the booking calendar and the rate sheet. The AI checks availability, quotes the published rate, explains the inclusions, takes the guest’s details and either holds the dates for a set period or sends a payment link, depending on the company’s rules. This call should stay with the AI end to end.

Call two: a disputed charge. A guest who checked out yesterday says the security deposit refund was short by ₹4,000 and nobody explained why. The AI can look up the booking and see that a deduction was recorded. It should not explain, defend or reverse the deduction. Damage assessments involve photos, staff notes and judgement. The right move is to acknowledge the complaint, confirm the booking details, state plainly that a person will review it, and create a case with a named owner and a promised callback window.

Call three: an urgent request. A current guest calls at 9:40 pm: the power has gone out in the villa, there is a child with them and the backup inverter is not working. This is not a booking question. The AI should recognise urgency from the content of the call, skip any menu, and transfer immediately to the on-duty caretaker. If the caretaker does not answer, it must try the next person on the list, not leave a voicemail and move on.

What a good transfer carries

When the AI hands a call to a person, the person should not start from zero. A useful transfer note includes:

• Who is calling, and their booking reference if found.

• What they asked, in one or two sentences.

• What the AI already told them, including any rate or time it quoted.

• Why it is being transferred: dispute, exception, urgency, or the caller asked for a person.

• What the caller was promised next (“a manager will call within two hours”).

This note should reach the person’s screen or phone before they say hello. For a callback, it becomes the case record.

When nobody answers

A design tested only for the case where a person picks up leaves the fallback unproven. The harder question is what happens at 11 pm on a holiday weekend when the duty manager’s phone is on silent.

Write a fallback chain for each type of handoff:

1. Primary owner for the shift (the duty manager, or the caretaker for that cluster of villas).

2. Secondary owner if the primary does not answer within a set number of rings or minutes.

3. Escalation to the owner or operations head for anything involving safety.

4. Honest message to the caller if all attempts fail: what has been logged, who will call back and by when. Never a promise the business cannot keep.

For urgent safety issues, the AI should also give the caller any emergency numbers the business has approved, rather than waiting for a staff member.

The same structure applies to a boutique hotel in Dubai, where the night desk might be one person covering reception and guest requests at once. The chain is shorter, but it still needs a second name.

A pre-launch test you can run

Before any AI receptionist goes live, script at least these calls and run them against the system. Record each result as pass or fail.

1. A straightforward availability and rate question. Does it quote the right rate for the right dates?

2. A booking for dates that are already taken. Does it offer alternatives without inventing availability?

3. A guest disputing a charge. Does it refuse to rule on the dispute and create a case with an owner?

4. A guest reporting a safety issue. Does it transfer immediately, and does the fallback fire if the first person does not answer?

5. A caller who asks for a person in the first sentence. Is that respected without argument?

6. A caller speaking a mix of languages, such as English and Hindi. Does it understand, or does it fail politely and hand over?

7. A question the rules do not cover, like bringing a pet. Does it say it will check rather than guess?

8. A transfer at a time when the primary owner is off shift. Does it route to the right person?

Then read the transfer notes the test calls produced. If a staff member could not act on the note alone, the handoff is not ready.

Where SolvTree fits

This maps to our receptionist AI and voice agents: handling routine booking calls, transferring disputed or urgent ones with context, and running the fallback chain when staff are unavailable. We usually start with one call type and one property, test it with scripted calls, and expand only once the handoffs hold up.

Next step: Bring one inbound call flow to a free 30-minute operational audit and map it with us, including its human fallback. The report is yours to keep.