An AI receptionist only knows what your office wrote down

Between June 1 and September 10, 2026, the AI receptionist we build and run on a Toronto property management line took 2,147 graded calls, and in 100 of them the caller asked something it had no rule for. Fifty-eight were the same question: "When will someone call me back?" The agent had no time to give, because nobody had given it one.

Most advice on AI receptionist setup is about the script and the voice. Our gaps came from policy the office had never written down, because the people who answer the phone never needed it written.

A receptionist knows when maintenance usually calls back. They know which buildings the company stopped managing, who looks after which owner, and that tenants check their balance somewhere new since the payment system changed. An agent knows what it was given. Asked anything else, it can stall or improvise, and both are worse than a person who knows. Ours mostly stalled.

We built that agent, so each gap below is a question we should have asked before it went live. The worksheet at the end is the list we now work through first.

The biggest gap was a callback time nobody had set

At 9:30 on a Friday morning in July, one caller asked again and again when someone would call back, and got the same line each time. The other notes read alike. Callers asked whether it would be today, how long it would take, whether anyone would call before a holiday. An owner asked for the typical response time on a request. Where a note says what the agent did offer, it was "as soon as possible" or its closing line, repeated.

We sorted 56 of the 58 as low severity, meaning the office still got what it needed to call back. They cost something anyway. A caller who hangs up without a time has no way to tell a slow afternoon from a lost message.

The fix is a decision the office makes once: a callback time for each kind of request, set at a level staff already meet nine times in ten. Then the agent says it, and the office is measured against it. A promise the office does not keep is worse than "as soon as possible", because now the caller has a time to be angry about.

Most of the rest sat in the front desk's head

The other questions were small, and most had an answer somebody in the office knew:

• Whether the company still manages a particular property, and who handles an owner's account. Two of them hung up without leaving any details when the agent could not say.
• Weekend hours, the right address or email for a request, what to do about keys returned late.
• Which portal takes maintenance requests, how a tenant can pay, and where to check a balance now.
• Whether they were talking to a person. Five callers asked, and each time the question was logged as unanswered.

None of this needs judgment. Each item is one line in a document the agent reads, plus someone whose job is to change that line when the answer changes. When the payment system changed, the office knew and the agent did not, and it kept taking tenants' calls about it.

Emergency callbacks and payment emails need a rule before go-live

Six callers rang back about an emergency they had already reported, to say nobody had responded or to ask when help would arrive. Four times the agent repeated its dispatch line, the one thing it had been told to do with an emergency. Twice it had nothing to say. We sorted all six as high severity, the top of our scale.

An emergency path written only for the first call has nothing to say when the first attempt failed. It needs a second branch: another number to ring and a named person on call, plus a sentence that tells the caller what happens next. What counts as an emergency in the first place is in our after-hours triage rules.

Four callers wanted to know whether an email they had received was genuine. Three of the emails asked for payment, and on one of those calls the caller also wanted to know whether a person named in the email worked there. Three got no answer. On the fourth call the agent said the payment address looked legitimate, which it had no way of knowing. An agent should never vouch for a payment instruction. It should name the one place real payment details live and pass the email to a person who can check it.

Who gets put through is policy too

On the same line, 243 callers asked for a live person and did not reach one, which we counted in what our voice agent got wrong. Who gets put through, and when, is a decision of the same kind, and it belongs in the same document.

On the same operation's brokerage line, not one of 1,211 graded calls carried a question the agent had no rule for. Brokerage callers mostly want a named person or a showing, and the agent's instructions covered the routing. On these two lines, the gaps gathered where callers wanted answers from the company itself, which on a property management line means tenants, owners and vendors.

The worksheet

Before an AI receptionist answers its first call, sit with the person who answers your phone today and fill this in. Budget an afternoon. Each "start with" is where we would begin; change every number to one your office already hits.

1. A callback time for each kind of request. Callers asked: "Will someone call me today?" (58 calls). Start with: emergencies within 30 minutes at any hour; anything else reported before 3pm, the same business day; after that, by noon the next business day.
2. What happens when an emergency caller rings back. Callers said: "Nobody came" (6 calls, all high severity). Start with: a second number, a named person on call, and one sentence on what happens next.
3. Payment instructions. Callers asked: "Is this payment email real?" (4). Start with: never confirm one; name the single channel for payment details and pass the email to a person.
4. What you manage, and who handles it. Callers asked: "Do you still manage this building?" "Who is my account manager?" (11). Start with: a current property list with an account lead for each owner.
5. Hours, addresses, keys. Callers asked about weekend hours, which address or email to use, late key returns (6). Start with: one line per answer, including what to do outside hours.
6. Systems and payment methods. Callers asked: "Which portal?" "Where do I check my balance now?" (7). Start with: every change goes into the agent's notes the day staff hear about it.
7. "Am I talking to a real person?" Asked and logged as unanswered (5). Start with: one honest sentence, then the offer of a message or a person.
8. Who reaches a person. Asked for one and did not reach one live (243). Start with: who is put through, during which hours, and what the agent says when nobody is free.

Put a name beside every line: the person who changes that answer when it changes. Then read the first two weeks of calls against the sheet, and add a line for every question the agent could not answer.

How we counted

A separate AI grader reads every call on the line after it ends and writes a report, including any question the caller asked that the agent left unanswered. For our count of the agent's mistakes we sorted every report on both lines from June 1 to September 10, 2026 by problem type and severity against a written rubric. The types and severities are our judgments, not logged fields. This post takes the property management calls typed "no rule for the situation" and sorts them by the question asked, with a keyword script whose every assignment we then read. The grader's notes are short and generic, and nothing in them identifies a caller. The 9:30 Friday call is one of them, paraphrased.