AI Systems
What an AI receptionist should never do
The value of a voice agent is in what it refuses to attempt. Boundaries are the product, not a limitation of it.
Escalation is a feature
The most common failure in voice automation is not a wrong answer. It is a system that keeps going when it should have handed over. A caller who is confused, upset, or asking something outside the agent's scope needs a person, and every second spent proving otherwise costs goodwill you will not get back.
Escalation therefore needs to be designed, not left as a fallback. That means defined triggers, a warm handover with the context already gathered, and a route that works outside office hours.
Five things to rule out explicitly
Never guess at identity. If the agent cannot verify who it is speaking to, it must not read back account details, appointment history, or anything else specific.
Never invent availability. A booking is only useful if it was checked against the real calendar at the moment of the call. An agent that promises a slot it cannot hold creates two problems instead of one.
Never give clinical, legal, or financial advice. Answering what a service covers is information. Telling someone what they should do about their symptoms is not, and no disclaimer makes that safe.
Never take an irreversible action alone. Cancellations, refunds, and anything that moves money belong behind a human approval step, with the threshold agreed with you.
Never pretend to be a person. Callers work it out, and the discovery costs more trust than the disclosure would have.
What good looks like instead
Answer on the first ring. Handle the questions that genuinely repeat: hours, location, pricing structure, what to bring, how to prepare. Book against live availability. Take a message with enough detail that the callback does not start from zero. Log everything where your team already looks.
That is a narrow scope, and narrow is why it works. A voice agent that does six things reliably is worth more than one that attempts thirty and needs supervising.
How to test it before launch
Write the awkward calls down and run them deliberately: the caller who mumbles, the one who switches language mid-sentence, the one asking something the business does not do, the one who is angry. The correct behaviour in several of those cases is to stop and fetch a person.
If nobody has written that test set, the system has not been evaluated. It has been demonstrated.
Next
Retrieval quality beats model choice
Teams spend weeks comparing models and an afternoon on how documents are prepared. The second decision usually determines whether the answers are any good.