What an AI receptionist actually does

A phone or chat agent answers common questions and captures information. Typical work includes stating hours and location, taking or requesting appointments, answering an order status question, and collecting a name, number, and reason for the call. It works from rules and content you approve.

It is not a general assistant. A useful agent has a narrow job, a clear script for the common cases, and a short path to a person. Example: a plumbing company uses one to take after-hours calls, record the address and the problem, and tell the caller when someone will respond. The agent does not diagnose the problem or promise a repair time.

  • List the calls it should answer
  • List the calls it should forward
  • Decide what it must never promise

Decide what it must never do

Write down the things the agent is never allowed to do. For most small businesses these include giving advice, quoting a firm price, promising a specific time, or handling emergencies. A blocked action should route to a person, not to a dead end.

Example: a dental office lets the agent offer open appointment slots but not give any clinical guidance. If a caller describes pain or an injury, the agent transfers the call or takes a message marked urgent. The rule is written before the build, so it is not decided in the moment.

  • Never give professional advice
  • Never promise a firm price or time
  • Never handle emergencies alone
  • Route blocked actions to a person

Handoff is the part that matters

The handoff is where most value is won or lost. Decide how a caller reaches a person, how quickly, and what the person receives. A transfer during business hours is different from an after-hours callback.

Example: a real estate office takes enquiries from ads. The agent collects the property, the caller's question, and a preferred contact time, then sends a summary. If the caller asks for a viewing, the agent records the request and does not confirm a time until an agent agrees. Nothing is promised on the spot.

  • Define how a caller reaches a person
  • Set what happens outside business hours
  • Pass a written summary, not just a number

Test it before you trust it

Test the agent with real questions, not just polite ones. Use background noise, interruptions, strong accents, and callers who change their mind. Check what happens when a caller says nothing, asks for a human, or uses a word the agent does not know.

Example: a restaurant agent is tested with a caller who asks about a dish, then a large group booking, then a refund. The goal is not to pass every test but to confirm the agent escalates early instead of guessing. Record the results so you can compare after changes.

  • Test messy and unclear requests
  • Test the human request path
  • Test a failed transfer
  • Keep the results on file

The costs that continue

An AI receptionist has three cost buckets. Setup covers call design, the build, and testing. Usage covers telephony minutes and any per-conversation fee the voice platform charges. Support covers tuning, changes, and fixes after launch.

Usage grows with volume, so call handling is not a fixed cost. Support is ongoing because your hours, language, and systems change. Ask which platform fees you pay directly and which flow through the vendor. Example: a seasonal business sees usage rise in its busy months, so it plans for a variable line item rather than a flat fee.

  • Separate setup from usage and support
  • Ask who pays each platform fee
  • Plan for usage to vary with call volume

Questions to ask a vendor

Ask what the agent does when it does not know, how a caller reaches a person, and what is recorded. Ask who owns the number, the scripts, and the call data if you leave. Ask what happens when the underlying platform changes.

Example: before signing, a clinic asks the vendor to walk through a failed transfer, a misheard name, and a caller who refuses to talk to a machine. Clear answers matter more than a polished demo, because these are the moments that shape what a caller remembers.

  • Ask how it fails and who is told
  • Ask who owns the number and data
  • Ask what support includes
  • Ask what is explicitly out of scope

Take this with you

Define the job, write the hard limits, design the handoff, and test the messy cases before launch. Plan setup, usage, and support as separate costs, and confirm which fees you pay directly. A narrow agent that escalates early beats a broad one that guesses.