Plain-English answers
A well-built agent that hits its limits has four good moves: ask again differently, admit it and take the details, transfer to a human, or capture and escalate. It has exactly one forbidden move — inventing an answer. Every AI agent, ours included, will eventually meet a caller it can't fully handle. The difference between a good agent and a liability isn't whether that moment comes; it's what the agent does in it.
Last updated: August 2026
This is the question behind every skeptical look we get: "And what happens when it doesn't know?" It's the right question — and the honest answer is that a good agent has a plan for it. (If you want the mechanics of what's happening underneath, we walk through how AI voice agents actually work.) An agent that hits its limits has exactly four good options, in roughly this order:
If the agent didn't catch something — noise, an accent, a mumbled address — it should confirm or rephrase, the way a person would: "I want to make sure I have this right — was that 42 or 402?"
For a question outside what it's been given ("do you guys also do dock work?"), the agent should say plainly that it doesn't have that answer and pivot to what it can do: take the details and have someone follow up with the answer.
For situations flagged as transfer-worthy — a caller who asks for a person, an emergency, a topic the owner wants humans handling — the agent hands the live call off. When no one picks up, it falls back to capturing everything and marking the message urgent.
The floor for every unhandled situation: full details taken, structured message sent, flagged by urgency, so the human callback starts with context instead of a mystery voicemail.
The agent must never invent an answer. Not a price it wasn't given, not a service area it doesn't know, not a promise about timing. In the trade this failure is called hallucination, and preventing it is a build discipline: the agent is given an explicit body of facts it may state, rules for what it must never commit to, and instructions that route everything outside that boundary into the four moves above.
This is a large part of what "custom-built" means — the guardrails are written to your business, not a template's. It's also the practical difference between an agent that's genuinely useful on the phone and one that's just a convincing voice; we cover the broader picture of what these things are in what is an AI voice agent.
Here's the counterintuitive part: an honest handoff is often better than an improvised answer, from anyone. An untrained human answering your phone — a temp service, a rushed family member, a generic answering service reading a card — will guess at prices and promise things you can't deliver. A properly built agent structurally can't. Callers forgive "let me have someone get you the exact answer" far more readily than a wrong answer confidently given.
That trade-off is the heart of the comparison we draw in AI receptionist vs. answering service — and it's worth judging with your ears, since how an admission sounds matters as much as the fact of it (see do AI phone agents sound robotic?).
You define the boundaries during the build: what the agent knows, what it quotes, what triggers a transfer, who gets texted and when. Afterward, every call produces a summary, so the calls the agent couldn't finish are visible instead of lost — you see what it was asked and how it responded.
When the same unhandled question keeps appearing, the fix is an update to the agent's knowledge, and that gap closes for every future call. The agent gets more capable over time, in exactly the places your callers push its limits.
Don't take our word for it — call our demo line at 1-855-952-4546 and deliberately go off-script. Ask it something it can't know. What you should hear is a graceful admit-and-capture, not a confident guess. If that's the behavior you want in front of your customers, we'll build your first agent free: no cost, no contract, and you decide afterward whether to keep it.
A well-built one admits it doesn't have that answer, captures the caller's details and question, and routes them to a human — by live transfer for urgent situations, or by a structured, flagged message otherwise. What it must never do is invent an answer.
Yes. Transfer rules are set during the build: callers who ask for a person, flagged emergencies, and any topic the owner wants humans handling. If the human doesn't pick up, the agent falls back to taking full details and marking the message urgent.
Hallucination is the model inventing a plausible-sounding answer it was never given — a price, a promise, a service it doesn't offer. It's prevented in the build: an explicit set of facts the agent may state, hard rules about what it can't commit to, and routing everything outside that boundary to confirmation, message-taking, or transfer.
Yes. Every call produces a summary, and unhandled or escalated calls are flagged, so nothing disappears the way a hangup on voicemail does. Recurring gaps become updates to the agent's knowledge, so the same question doesn't stay unhandled.
Call the demo line, ask it something it can't know, and listen to what it does next — then we'll build one around your business, free.
Ask Custom Voice Automations
Usually replies instantly