A voice receptionist bounded by three approved sources — your published price list, your live calendar, and your service-area list — with a scripted deflection for everything outside them.
An AI agent for hallucination-free voice reception is a 24/7 digital assistant that answers callers only from approved sources — a published price list, a live calendar, and a written service-area list — books what those sources support, and says it will have someone confirm anything else. It stops a phone agent from inventing a price or promising a same-day slot when a caller pushes hard for an answer nobody has given it. This matters more on the phone than anywhere else. A spoken figure sounds like a commitment, the customer remembers it, and your technician is the one who has to argue with it on the doorstep. A quoted price that was never on your price list is a discount you did not agree to. A same-day slot that does not exist is a no-show with your name on it. The design here is narrow on purpose: three sources, one refusal phrase, and a written log of every question the agent could not answer.
Answers from approved sources, refuses the rest, and writes down what it refused.
Answers pricing questions only from your published price list, word for word
Offers only appointment slots that are genuinely open in the live calendar
Confirms coverage only against your written service-area list, never by inference
Uses one scripted line for everything else: someone will confirm that and call back
Holds that boundary when the caller pushes for a ballpark figure
Logs every deflection with the caller's own wording and a timestamp
The risk with a voice agent is not that it fails to answer. It is that it answers confidently and wrongly, and the caller believes it. Contractors carry the cost of that: an invented repair price becomes an argument at the door, a promised slot becomes a wasted morning, and a service area guessed from a postcode becomes a ninety-minute drive that was never billable. Constraining the agent to three sources removes almost all of it. The honest trade-off is that it will say it needs to check more often than a human receptionist would — and the deflection log tells you exactly which questions to add an approved answer for.
A simple, three-step flow.
You supply a published price list, a live calendar connection and a written service-area list. Anything not in those three is out of scope by definition, not by judgment.
On the call, the agent looks up the answer and reads it back. If there is no source for the question, it uses your approved wording and offers a callback instead of estimating.
Every deflection is written to a sheet with the timestamp, the caller's phrasing and the call recording link, so you can decide which questions deserve an approved answer.
A realistic use case with concrete timing and output.
Scenario: a six-technician home services company handles about 200 calls a week and publishes a diagnostic fee of eighty-nine pounds, plus a service-area list of fourteen towns. On Tuesday morning a caller asks what it costs to replace a condenser fan motor. That figure is not on the price list, so the agent says the diagnostic fee, states plainly that repair pricing is quoted after the technician has seen the unit, and offers Thursday between 8am and 11am. The caller pushes twice for a ballpark; the agent repeats the boundary and does not produce a number. The call ends booked. In the week's deflection log there are 31 rows, 19 of them asking about the same motor replacement — so the owner adds an approved price band, and next week that question is answered from a source.
Roles that gain practical value from this AI agent.
You are the one who has to honor whatever the phone told a customer last Tuesday.
Technicians stop arriving to arguments about a price nobody in the business quoted.
The schedule only contains slots that were real when they were offered.
The deflection log is the honest answer to "what will it get wrong", available in week one.
One consistent refusal phrase is easier to defend than five improvised ones.
Each branch's price list and coverage list bound the agent separately, so answers cannot cross over.
Key tools and what the AI agent does inside each.
Carries the call, supports barge-in when a caller presses for a figure, and hands over to a person on request.
Holds the approved price list and service-area list, and receives one row for every deflection.
Provides the live availability the agent reads from, so no slot is offered that is not open.
Supplies the job types and price book the agent quotes from, and receives the booked call.
Stores the recording and transcript alongside each deflection so you can hear what was actually said.
Six practical scenarios that this AI agent excels in.
Common questions about using the AI agent in workflows.
An AI agent for hallucination-free voice reception is a 24/7 digital assistant that answers callers only from approved sources — a published price list, a live calendar and a written service-area list — books what those sources support, and says it will have someone confirm anything else. It stops a phone agent from inventing a price or promising a same-day slot when a caller pushes for an answer nobody has given it. Unlike a general chat assistant on a phone line, it has no freedom to compose an answer from what sounds plausible.
The honest answer is that the risk is reduced by removing the opportunity, not by asking the model to be careful. Prices, availability and coverage are looked up, and if the lookup returns nothing the agent has exactly one thing it is allowed to say. Free-form small talk is still generated language, which is why the boundary is drawn around the three things that cost money when they are wrong.
It uses wording you approve, typically along the lines of: that is not something I can confirm, so I will have one of our team call you back today. It says it once, repeats it once if pressed, then offers a callback or a transfer. It does not soften into an estimate on the third attempt.
Every deflection is written to a sheet as a row: timestamp, caller number, the question in the caller's own words, and a link to the recording. You read the sheet. The agent does not summarize its own performance or grade itself — you look at the raw rows and decide what to approve.
On breadth, yes, and deliberately so. A human receptionist will improvise a helpful-sounding figure under pressure and you will find out later. This agent will say it needs to check. The trade you are making is fewer answered questions in exchange for no invented ones, and the log shrinks that gap week by week.
Yes, within the bounds above. If the job type is on your list, the address is in your service area and the calendar has a genuine opening, it books the call end to end. If any of the three checks fails it stops and routes to a person rather than booking something it cannot stand behind.
No to payments — the agent cannot take money, so deposits and card details are handled by your existing process. Safety advice comes from a fixed approved script only, covering things like a gas smell or a burning odour, and anything beyond that script is escalated to a person immediately and logged.
A voice receptionist bounded by three approved sources — your published price list, your live calendar, and your service-area list — with a scripted deflection for everything outside them.