How to choose an AI voice provider in Singapore

Every provider will show you a polished demo, a feature list and an impressive-sounding latency figure. None of that predicts whether the agent survives a real call with a distracted person on a mobile in a hawker centre. This page gives you a weighted framework, one live call that settles most of it, the two limitations nobody volunteers, and why the script matters more than the voice.

The central point is simple. AI voice providers are not interchangeable, and the differences that matter in Singapore are not the ones on the comparison table in a sales deck.

The gate: does the call connect at all?

Check this before you evaluate anything else, because a provider can fail here while scoring well on every other criterion.

Since December 2022, Singapore telcos block incoming international calls displaying +65 3, +65 6, +65 8 and +65 9 prefixes, and robocalls have been blocked by pattern recognition since 2020. Many AI voice platforms run overseas. If yours routes calls internationally while showing a Singapore caller ID, those calls can be blocked before they ring, and you will experience it as a mysteriously poor connect rate.

The script decides whether any of this works

The agent is only as good as what it was told to say. A weak script does not fail cheaply: you pay for connected minutes either way, so every bad conversation costs you the same as a good one and returns nothing. This is the most common reason a technically capable deployment produces no appointments.

Buyers evaluate voices and latency because those are easy to hear. Script quality is harder to assess in a demo and matters more over twelve months, which is exactly why it deserves weighting rather than being lumped in with onboarding.

Writing for an AI is not writing for a person

A human caller can absorb a rambling answer and steer back. An AI voice agent cannot, reliably. So the script has to do work that a person would do instinctively:

• Ask leading questions, not open ones. "Are you still looking at coverage for your parents?" produces an answer the agent can act on. "Tell me about your insurance needs" produces thirty seconds of speech it will mishandle.
• One question at a time. Two questions in a sentence and you get an answer to one of them, usually the wrong one.
• Short turns. Long monologues invite interruption, and interruption is where most agents fall apart.
• Ask permission early. A quick "is now a good time?" costs three seconds and prevents the caller feeling ambushed, which is where complaints come from.
• Branch the objections that actually happen. Not thirty of them. The five or six you hear every week, written out properly, with a graceful exit for everything else.
• Recognise an opt-out in plain speech. People do not say "please remove me from your marketing list". They say "not interested, don't call again", or say it in Mandarin.
• Do not ask what you cannot act on. Every question needs to change what happens next, or it is just lengthening a call you are paying for by the minute.

Five questions about the script itself

1. Have you written for my industry? Ask for an example and read it properly. If they have none, you are paying to be their first attempt.
2. Who writes it, and how many revisions are included? First drafts are rarely right. A provider offering one revision is telling you something.
3. Can I change it myself, and how quickly does a change go live? You will want to adjust the opening line in week two. If that needs a support ticket and five days, iteration stops.
4. How do we test a change? Running a new opening on a hundred records before committing the list is worth far more than debating it in a meeting.
5. What happens to conversations the script does not cover? There should be a defined graceful exit, not improvisation.

Treat the first script as a draft you will rewrite twice. Budget the time for it, and pick a provider who expects that rather than one who hands you a finished document and moves on.

Two things most agents cannot do, whatever the deck says

These are not edge cases. They are the two assumptions that most often break a deployment after it has been signed. Decide how you will work around them before you buy, not after.

Most cannot switch language mid-call

Moving between English and Mandarin partway through a sentence is ordinary in Singapore conversation, and it is one of the hardest things for a voice agent to do. Most cannot, and several that claim to will do it badly enough that the prospect notices.

Plan around it. Pick the language before the call rather than during it, and run a separate agent and a separate campaign for each. That means segmenting your list by likely language preference, which is worth doing anyway. If a provider says their agent code-switches comfortably, that is exactly the claim to test on your live call.

Most cannot transfer a live call to a person

Buyers frequently assume a warm prospect can be handed straight to a human. In practice most agents offered in this market either cannot transfer at all, or cannot do it reliably enough to build a process on.

Design for callback instead. The agent captures interest, confirms a good time, ends the call politely, and a person rings back. That is a perfectly good workflow, and it is more honest than promising a transfer that drops the call half the time. What matters then is how fast you find out: ask how a warm lead reaches you, whether that is an instant notification, a CRM record or an email at the end of the day. Speed to callback is the number to optimise.

Latency, in language buyers can use

Latency is the delay between the person finishing their sentence and the agent starting its reply. It is built from several stages: recognising the speech, deciding what to say, synthesising the voice, and carrying it over the network. A provider quoting one number is usually quoting one stage.

Do not plan around published figures. They are measured under favourable conditions and rarely reflect a mobile network in Singapore at 6pm. Judge it on a live call instead: if you find yourself wondering whether the line has dropped, it is too slow.

Watch for the second-order problem too. Agents often respond quickly until they have to look something up in a CRM or knowledge base, then stall. Test the paths that involve a lookup, not just the greeting.

Scoring two providers side by side

Score each criterion out of 10, multiply by the weight, and total. The arithmetic is less important than being forced to give a number to something you were going to judge on impression.

If the totals land within a few points of each other, stop optimising and pick the one whose people answered your questions most directly. Over a twelve month deployment that matters more than a marginal score.