Voice AI agents: when to talk, and when to type instead

"What if the agent picks up the phone and answers out loud?" It's one of the questions we get most these days, and it makes sense: AI voice demos sound almost human, and the temptation to add voice to everything is strong. But the useful question for your business isn't whether the machine can talk — it's very good at that now — it's whether your customer wants to talk in that particular moment. Sometimes voice solves something text can't; other times it's an obstacle with a robotic accent. Let's pull those two apart so you decide on merit, not on hype.

What "an agent that talks" actually means today

First, "voice" isn't one thing. When someone says "I want an agent that talks," they could mean three very different scenarios:

• Answering phone calls. The agent picks up, understands what you say and replies out loud, like a phone receptionist.
• Understanding voice notes. It doesn't speak; it listens. It takes those 40-second audio clips people send on WhatsApp, transcribes them and acts on them.
• Talking inside an app or device. Less common for a neighbourhood small business, more the domain of digital products.

Blurring the three leads to odd decisions. Most small businesses win or lose the game on the first two: the phone that rings today when no hand is free, and the voice notes that pile up faster than you can listen to them.

Where voice wins hands down

Some situations make talking clearly better than typing:

• Hands busy. In a workshop with greasy hands, in a restaurant kitchen mid-rush, on a building site or behind the wheel, typing isn't an option. Talking is.
• The customer who calls because they'd rather call. If your business takes calls — a clinic, a garage, a restaurant — answering the phone at 2:30pm in the thick of it is exactly what gets dropped today. An agent that picks up, takes the details and leaves nobody ringing into the void wins back bookings and customers who'd otherwise have hung up.
• Urgency and venting. "My boiler's broken and the house is freezing." That person wants to speak to someone, not fill in a seven-field form.
• Less digital customers. Older people, or anyone not comfortable typing on a phone. Voice drops the barrier instantly.

And there's one case that's nearly universal: the voice note. Plenty of customers send audio instead of typing. An agent that transcribes it, understands it and hands you the gist — or answers directly — saves you the toll of listening to each one while you're trying to do something else.

Where text still wins

Voice is convenient for saying things, but poor for keeping them. Text clearly wins when:

• There's data to re-read. An address, an account number, a price, a booking link, opening hours. Spoken, they slip away; written, they stay and you re-read them as often as you need.
• The conversation is asynchronous. Text doesn't require both sides at once. The customer writes at 11:40pm; you (or the agent) reply when it's time. A call forces you to coincide.
• The place is noisy or public. On the train, in a meeting, with the baby asleep next door — nobody's going to talk to their phone. Typing is discreet.
• You need a record. A written conversation stays put: you search it, forward it, prove what was said and when. "They told me on the phone" can't be forwarded.

Put another way: voice is good for the moment; text is good for whatever has to outlive the moment.

A simple rule: follow the channel the customer already chose

If you take one thing away, make it this: don't force the channel. The person who calls has already told you they'd rather talk; the one who messages on WhatsApp has already told you they'd rather type. The most common mistake is fighting that signal — turning a calm chat into an automated call, or trapping someone who just wanted to send two lines inside a voice menu.

Voice isn't there to replace text. It's there to cover the channel that's already voice — the phone — so it stops ringing out. Match each channel to what people expect from it and most decisions make themselves.

When NOT to add voice (yet)

Here's the honest limit:

• Noisy places, accents and people who interrupt. Transcription slips, the conversation breaks and the customer gets frustrated. A voice bot that can't understand you and won't let you reach a human is worse than voicemail — at least voicemail doesn't argue.
• The phone-menu maze. If your voice agent loops people with no clear way out to a person, you've made service worse, not better. The escape hatch to a human has to be one sentence away, always.
• Consent and recording. Recording calls means notifying people and complying with GDPR, and a synthetic voice that imitates a real person can mislead. Have the agent identify itself as an agent from "hello".
• Cost and fragility. Processing voice in real time is pricier and more brittle than text. If 90% of what reaches you comes in as written WhatsApp, standing up an "AI call centre" solves a problem you don't have.

The quick rule: if you don't have real call or audio volume, voice is a solution looking for a problem. Don't add it to look modern.

How to start without betting the farm

The good news is you don't have to go big. A sensible order:

1. Start by understanding, not speaking. Have the agent transcribe and summarise voice notes and voicemail. Zero risk, immediate saving, and no synthetic voice that could come off strange.
2. Reply on the same written channel. With what it understood, let it answer the predictable stuff on WhatsApp or email and escalate the rest to you.
3. Only if you genuinely get a lot of calls, try voice on a narrow case: pick up the after-hours call, take the details and pass it to a person the next day. With a clear route to a human and the agent introducing itself as what it is.

And at every step, the usual: you approve how it sounds and what it's allowed to say; you never let it loose all at once. It starts by proposing and earns autonomy one task at a time.

Whether your customers call or write, the underlying idea is the same: a single inbox that takes messages over WhatsApp, web, email and phone, handles the predictable ones and escalates whatever needs your judgement. That's exactly what our receptionist does — and deciding where it talks and where it types is part of getting it right. You approve; it handles the repetition without tiring.