Most scoping conversations about a voice agent start with what it should say: the pitch, the offer, the tone. Few of them finish with a written answer to a narrower question — what it should never be asked to handle at all. That gap does not stay theoretical. It surfaces live, on a real call, the first time someone asks for something the team never decided how to refuse.
The fix is not a longer script. It is a shorter list, decided once, before the agent goes live, rather than worked out call by call against a caller who is already frustrated.
A boundary decided at launch is a rule
A scope decision made calmly, before any call happens, becomes a line in the constraints brief the agent is built around — the same document that already holds what the agent is allowed to promise. Written down in advance, it is a rule the model can be held to. Discovered mid-call, it is whatever the model happens to do under pressure, and the caller is the one who finds out which.
Four categories that belong to a human by design
None of these are edge cases in the sense of being rare enough to ignore. They are edge cases in the sense of needing a person, every time, regardless of how often they come up.
The job is to recognise the shape of the call and get a person or an emergency service on the line immediately — not to attempt a diagnosis, a fix, or a calming script of its own.
A price, a refund, a contract exception — anything that needs discretion rather than a lookup. The agent can quote what is published; granting an exception belongs to whoever holds the authority to grant it.
A complaint about a real failure is a request to hear from someone who can own it, however well the agent apologises on the way to a transfer.
A caller who names a lawyer, a regulator or a formal dispute needs a documented human response, not a scripted one.
The scope decision is not a limit on the agent. It is what makes the rest of the script worth trusting.
Finding these calls before a caller finds the gap
None of this should be a guess made from a meeting room, for the same reason the script itself should not be. Locator listens to 100% of a team's existing calls and scores them on 30+ parameters, which is enough coverage to show how often these four situations actually reach the line, in the caller's own words rather than an assumption about them. Listening to those calls before writing a script is the same argument one step earlier: read what is really being asked before deciding what the agent is allowed to answer.
What this does not mean
Not a case for a smaller or more timid agent. In practice, the four categories above are a small share of total call volume — which is exactly why they can be carved out by name instead of used as a reason to route everything to a human. The rest of the call stays the agent's job, including the objection, the price question and the booking, because none of those need an authority the agent does not have. What changes is that the excluded slice is decided, named, and built into the brief on day one — not discovered the first time it happens live.
The order that holds up
Decide the four exceptions before launch. Build their exit into the brief. Then let the agent take everything else.
The exit path itself — what the human on the other end needs to receive, and how a caller is told the call is moving — is its own decision, made once. This one comes first: deciding what triggers it.