Voice agents provoke a stronger reaction than chatbots, and reasonably so. A phone call carries an expectation of a person in a way a chat widget does not.
The honest framing is narrower than most vendor material: the comparison is rarely agent versus receptionist. It is agent versus voicemail.
Size the problem first
Before evaluating anything, get two numbers from your phone system: how many inbound calls go unanswered, and when.
For many service businesses the answer is uncomfortable — a substantial share of calls arrive outside working hours or while everyone is already on a call, and a meaningful proportion of those callers do not leave a voicemail and do not call back.
That figure is the opportunity. If it is small, a voice agent is not your priority and any proposal claiming otherwise is not looking at your data.
Missed calls that leave no voicemail are invisible in every report except the phone system's own. Most businesses are surprised by the number the first time they look.
Where voice agents work
Three deployments are consistently defensible.
- Out-of-hours answering — the alternative is voicemail, so almost any structured capture is an improvement.
- Overflow during peak — the alternative is ringing out while the team is on other calls.
- Qualification and booking for high-volume, low-complexity enquiries where the questions are genuinely standard.
Where they fail
Complaints. A caller who is already frustrated and reaches an automated system escalates, and the recovery cost exceeds anything the automation saved.
Negotiation. Anything involving price flexibility, contract terms or exceptions requires authority the agent does not have and should not simulate.
Distress. In sectors where callers may be in urgent or vulnerable situations, qualification questions before a human is reached is the wrong design regardless of efficiency.
These should route to a person immediately, without qualification steps in front of them.
The disclosure question
Voice agents are now good enough that callers frequently do not realise. That is precisely why disclosure matters.
The agent should state it is automated within the first sentence. The argument against is that disclosure reduces engagement slightly — the arguments for are that callers who discover the deception react considerably worse, and that disclosure is a legal requirement in a growing number of jurisdictions.
Give it a clearly non-human name. A human name with a synthesised voice is the version that ages worst.
Transfer is the feature that matters
Design the transfer before the script.
A caller who asks for a person should reach one without repeating themselves. That means the agent passes context — who called, what they said, what was captured — so the human starts informed rather than from zero.
Making someone re-explain after they were already routed is the single behaviour that turns a tolerable experience into a memorably bad one.
A defensible voice agent flow
Transfer is available at every step, not only at the end.
- 1
Disclose
State it is an automated assistant in the first sentence.
- 2
Offer the exit
Make reaching a person available immediately, not buried.
- 3
Capture
Name, contact, reason for calling, urgency.
- 4
Qualify
Two or three defined questions. Not an interrogation.
- 5
Act
Book, route, or take a message with a stated response time.
- 6
Hand off
Transfer or summarise to a person with the full context.
Compliance details worth checking
Call recording consent rules differ by jurisdiction, and several require explicit notification at the start of the call. If you operate across regions, the strictest applicable rule is the safe default.
Anything captured is personal data and falls under your normal retention and access obligations. A voice agent that stores transcripts indefinitely because nobody configured retention is a compliance problem waiting to be found.
What to measure
- Calls answered that would previously have gone unanswered — the core case.
- Qualified appointments booked, not calls handled.
- Transfer rate and whether transferred callers had to repeat themselves.
- Caller drop-off point, which shows where the script is failing.
- Complaint escalations arising from the agent itself.