A voice agent can answer routine calls once certified, but a phone line punishes small errors. It must say it is an AI at the start, read back every number and name it acts on, keep card details out of the conversation, and transfer to a person with a note the moment a call goes beyond its bounds.
What does answering calls with a voice agent involve, and who owns it?
The phone remains the channel customers choose when something matters or is urgent. A voice agent picks up, works out what the caller needs, authenticates them, answers or completes routine requests, and passes the rest to a person. It differs from a chat agent in three ways: speech recognition makes mistakes chat does not, callers cannot scroll back to check what was said, and silence or delay feels like failure.
The contact centre head owns the service. Telephony and security teams own the call path, recording and authentication. Compliance approves disclosures and scripts. The agent is one participant in that system, and the case for keeping its role narrow is made in our guide to the AI support agent that knows when to say no.
What does the agent prepare, and what does a person approve?
- Agent, depth 3: discloses. The first sentence says the caller is speaking to an AI system, and that they can ask for a person. Recording notices follow your existing rules.
- Agent, depth 1: identifies the reason for the call and checks whether it is in scope. Out-of-scope reasons transfer immediately.
- Agent, depth 3: authenticates the caller through your approved method, such as a one-time code sent to the registered device. It does not invent extra security questions.
- Agent, depth 3: answers routine questions from approved sources: balances, opening hours, order or claim status, how to do something.
- Agent, depth 3: reads back critical details before acting: amounts, dates, account endings, addresses, names spelt out.
- Agent, depth 1: routes card capture to a secure payment flow. Card numbers are never spoken to the agent or stored in its transcript.
- Agent, depth 2: writes the handover note for any transfer: who called, verified or not, what they want, what was already said.
- Person, adviser: takes transferred calls and everything that needs judgement, money movement beyond the agent's bounds, or care.
- Person, quality team: reviews sampled calls, including transcripts and the agent's verdicts.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Disclose | States it is an AI | 3 | Approving wording |
| Reason | Identifies intent and scope | 1 | Nothing |
| Authenticate | Runs approved method | 3 | Failed or unusual cases |
| Answer | Responds from approved sources | 3 | Novel questions |
| Read back | Confirms details aloud | 3 | Nothing |
| Card data | Routes to secure flow | 1 | Nothing |
| Handover | Writes the note | 2 | Adviser takes the call |
| Review | Nothing | none | Quality team samples calls |
Which controls matter, and what evidence does each leave?
Disclosure at the start of every call. Under the EU AI Act, Article 50 requires people to be told when they are interacting with an AI system; Article 50 disclosure in chat, voice and email agents covers how that reads on the phone. Evidence: the disclosure segment in each recording.
Read-back before any action. Speech recognition confuses "fifteen" and "fifty". Evidence: the confirmation turn in the transcript before each action.
Card data never enters the agent. PCI DSS v4.0 applies to card data wherever it travels, so keeping it out of the agent's audio and transcript shrinks what must be protected. Evidence: payment flow logs and transcript scans.
Transfer on request, every time. "Can I speak to someone?" is honoured, not negotiated. Evidence: time from request to transfer. The mechanics are in how to let an AI agent transfer calls to a person.
Each call's verdicts are filed with the transcript. Evidence: the append-only record.
What should you measure?
- Calls resolved by the agent, by reason, against calls transferred.
- Time to transfer after a request for a person.
- Read-back corrections: how often callers correct a detail. High rates point to recognition problems.
- Repeat calls within a short period for the same reason.
- Conduct issues found in sampled calls; can an AI agent review customer calls for conduct risk? covers how that review can scale.
Where does it usually go wrong?
The agent fills silence with guesses. When it does not know, it should say so and transfer; what happens when a support agent does not know the answer sets the pattern.
Interruptions break the flow. Callers talk over the agent. It must stop, listen, and not repeat its whole script.
Accents and noise defeat recognition. Test on recordings that sound like your customers, not studio audio.
Transfers lose context. The adviser asks the caller to start again. The handover note must arrive with the call.
Vulnerable callers are kept too long. Distress, confusion or mention of a crisis should trigger a transfer early.
The agent sounds too human. A natural voice is good; one that pretends to type, sighs or claims to "check with a colleague" misleads callers about what they are speaking to. Keep the manner plain and the disclosure honest throughout the call, not just in its first sentence.
Where Pryme fits
Attest certifies a voice agent on your own call recordings, replayed through the same speech recognition it will use live, so recognition errors count against the score. The refusal and escalation gates test whether it transfers when it should.
Questions people ask
Can the voice agent take payments?
It can start the payment and hand the card capture to a secure flow. Moving money is depth 4, so most firms hold any payment the agent initiates for confirmation by the caller through that flow.
Should callers be able to skip the agent?
Yes. Offer a route to a person from the start. Forcing callers through the agent turns a convenience into a barrier.
How should the agent answer questions about the recording itself?
From an approved script describing why calls are recorded and how the data is used, as in answering customer questions about how their data is used.