BUILT, IN TESTING / PHONES

An AI receptionist, put through 17 difficult test calls

IN SHORT

An AI that answers a business's phone: it answers questions from the business's own website, takes a message and reads every detail back, and hands a Spanish speaker to a Spanish agent. Before it takes a real call it has to pass 17 scripted calls written to trip it up. On the latest run it passed 16. It is not taking real calls yet.

  • In testing
  • Squared Away
  • Tested Sept 2026

THE PROBLEM

The risk is what it says when nobody is listening

A missed call is often a missed job, and an AI receptionist sounds like the answer. The danger is in the calls nobody planned for: an opening time it made up, a price the owner never agreed to, medical advice from a body shop, or a caller who talks it into ignoring its instructions.

A receptionist is only as good as the calls it was tested on, so the test calls came first.

WHAT IS BUILT

An agent, its rules, and the calls that test it

WHERE ITS FACTS COME FROM

  • The business's own website and a one-paragraph brief, so every fact it uses is the business's own.
  • Reviewed line by line: in the instructions it was given, every fact came from the site and none was invented.

WHAT IT DOES ON A CALL

  • Answers questions from the site, and never quotes a price, an opening time or a repair time.
  • Takes a message one detail at a time, then reads all five back once at the end.
  • Hands a Spanish speaker to a Spanish agent on their first Spanish sentence.

RULES THE TESTS ADDED

  • A sales call ends politely, with no message taken.
  • An emergency gets "call 911" and no medical instructions.
  • Insurance and legal questions go to the insurer or an attorney.

HOW IT IS TESTED

  • Each test call is a script with its own pass condition.
  • Another AI plays the caller, and a judge grades every call against its condition.
  • Every miss becomes a rule, and the whole set runs again.

THE TEST CALLS

17 calls written to trip it up

The third run, September 2026. Each square is one scripted call.

  1. Pushes on insurance
  2. Pushes back on price
  3. Asks about Saturday hours
  4. Presses for a repair deadline
  5. Corrects their phone number
  6. Says everything at once
  7. Asks two questions at once
  8. Asks if it is a robot
  9. Reports an injury
  10. A sales call
  11. Switches to Spanish
  12. Asks to book
  13. Only has a question
  14. Asks for advice
  15. Asks for a person by name
  16. Tries to rewrite its instructions
  17. Asks about warranty and competitors

Passed: 16Missed: 1

The one miss: after the caller corrected their number, it confirmed their name early, then left the name and number out of the final read-back because it had said them already. The fix is now a rule: a correction changes nothing else, and the final read-back always has all five details.

HOW A CALL FLOWS

What a caller gets

  1. AnsweredAs the business, every time the phone would otherwise ring out.
  2. UnderstoodA question is answered from the site; a Spanish speaker is handed over.
  3. Message takenFive details, one at a time, all read back at the end.
  4. Called backThe owner has the message and calls back.

The AIA person

Left before a real customer calls it: sending each message to the owner's inbox in the same form as a website enquiry, a test on a real phone line, Spanish on a live call, and the same 17 calls on a second AI model to choose the better of the two.

WHAT I DID

What I did

Compared the vendors

Read eight vendors' terms, prices and uptime before choosing one to test.

Wrote the brief

What it may say, what it must never say, and when it hands over.

Wrote 17 test calls

Each one a way a real call goes wrong, with a pass condition.

Ran and graded them

Three runs, every failure read in the transcript before deciding the fix.

Turned misses into rules

Sales calls, emergencies, advice and read-backs, each fixed with one rule.

Designed it to move

The phone number stays ours and the setup is built from the business's site, so changing vendor is a setting, not a rebuild.

BUILT WITH

Built with

  • Retell
  • GPT-5.6 Terra
  • Claude Sonnet (simulated callers)
  • Cartesia (voice)
  • The business's own website