Voice AI has become good enough that a phone agent can hold a natural conversation, check a calendar and book an appointment while the caller is still on the line. That makes it tempting to hand it the whole phone line.
The agents that work in production are usually more focused than that. They do one job well, and they hand over cleanly when a call doesn't fit.
Where voice agents work well
Voice agents are strongest when a call has a predictable structure:
- Capturing details. Name, number, reason for calling, postcode, preferred time.
- Booking and rescheduling. Checking real availability and confirming a slot.
- Answering the same questions repeatedly. Opening hours, service areas, what to bring, how long it takes.
- Out-of-hours cover. Taking a structured message instead of letting the call go to voicemail.
- Routing. Working out who the caller needs and passing them on with context.
What these have in common: a clear end state, a small number of variations, and a cheap path to a person when the call doesn't fit.
Where they struggle
- Open-ended, emotional or high-stakes conversations. Complaints, bad news and negotiations need a person.
- Calls that rely on judgement the business hasn't written down. If two experienced staff members would answer differently, so will the agent.
- Noisy lines and cross-talk. Speech recognition has improved a great deal, but a building site on speakerphone is still a building site.
- Long, multi-part instructions. People hold less in their heads on a call than on a screen. Confirmations need to be short.
None of these rule voice out. They mean those calls should reach a person quickly, and the scope should say so.
Pick one job first
The most reliable first voice agent does one job end to end. "Answer after-hours calls, capture the details and book a callback" is a good first scope. "Replace the front desk" is not.
One job means one set of scenarios to test, one set of escalation rules, and a clear way to tell whether it's working. Once it is, the next job can be added as a new version, scoped the same way.
What to decide before building
The call flow
Write the expected path as an outline: greeting, identification, the core task, confirmation, close. Then write the branches. What happens when the slot is taken? When the caller wants a person? When the calendar or CRM the agent depends on doesn't respond?
Escalation and handover
Decide when the agent transfers the call live, when it takes a message, and what the person receiving it sees. A transfer without context makes the caller repeat themselves, which is worse than having no agent at all.
Disclosure and consent
Callers should know they're speaking to an automated agent. If calls are recorded or transcribed, say so, and make sure your privacy notice covers it.
Outbound calling has its own rules in the UK, particularly for marketing calls and automated messages. Get your position on consent clear, with advice where you need it, before building anything that dials out.
Systems it touches
Calendar, CRM, booking system, ticketing. For each one, write down what the agent reads, what it writes, and what it does when that system is unavailable.
What success looks like
Choose measures you can read directly from the call logs:
- Calls completed without a transfer
- Bookings completed versus attempted
- Escalations, by reason
- Calls abandoned part-way through
Agree the targets as part of the scope, not after launch.
Choosing the platform
There are several capable voice AI platforms and telephony providers. They differ in response latency, voice quality, call control, integration options and running cost.
We compare them against the requirements in the project blueprint, record the recommendation and the trade-offs, and the client decides. The right platform for a high-volume booking line isn't necessarily the right one for a low-volume after-hours service.
Test with real calls
Run the scenarios in a test environment first. Then put the agent on a limited line with real callers before going fully live, and listen to the recordings.
The issues you find there, such as pauses that feel too long, confirmations that are too wordy, or an accent the recogniser struggles with, rarely show up in written tests.
The short version
- Start with one job that has a clear end state.
- Write the call flow, including every branch and handover.
- Tell callers they're speaking to an agent, and settle consent before any outbound calling.
- Decide the measures of success before launch.
- Choose the platform against written requirements.
- Test with real calls, and listen to them.