Voice intake

Two hosts cover why most voice AI projects miss the part that changes the business outcome, the two-surface design that makes a front desk trust an assistant, and the four situations where voice is genuinely the wrong answer.

A: For a dental clinic, or an HVAC company, the phone is the sales channel. It rings out at lunch, after hours, whenever the front desk is dealing with someone standing right in front of them.

B: And voicemail converts close to nobody. A caller who reaches a recording generally just calls the next listing.

A: So the obvious fix is a voice assistant that answers. And that framing is exactly where most of these projects go wrong — because answering is the easy part.

B: Three things have to happen in one call, and they get progressively harder. Hold a conversation that doesn't feel like a phone tree. Extract structured data from unstructured speech. And close the loop with something real.

A: And only the third one actually changes the outcome. An assistant that has a lovely chat and ends with "someone will be in touch" has converted a missed call into a slightly warmer missed call.

B: The caller still has an unsolved problem. They might still ring the next listing.

A: So the design decision that mattered most, in the build we're drawing on here, was making the call visible to staff while it's happening — not delivering a transcript afterward.

B: The caller gets a click-to-talk widget on the business's own site. Staff get a console where they watch the same conversation live — transcript streaming, qualification fields filling in as the caller talks, the calendar updating.

A: And that cross-surface handoff is basically the whole product, for a reason that isn't obvious until you actually watch someone use it.

B: A front-desk team will not trust an assistant they can't observe. Watching it converts the thing from "might be mishandling my customers" into a colleague you can glance at.

A: Trust is a design problem here. Not a model-quality problem.

B: What's genuinely hard, underneath that? Structured capture, for one. People don't answer questions in order — they volunteer their address mid-sentence describing the problem, they correct themselves. A system built for field-by-field answers falls apart on real speech.

A: Latency's the other big one — arguably the most damaging flaw of all, worse than an imperfect answer. A pause reads as "this is a machine and it's confused."

B: And the assistant needs a clear rule for handing to a human — frustration, anything out of scope, anything involving money. Getting that wrong in the confident direction is how these projects generate complaints.

A: When is voice just the wrong answer, though? Worth being honest about this, because the demos are persuasive.

B: Low call volume — below roughly a handful of missed calls a day, a shared inbox and a callback discipline is cheaper and does the job better.

A: Complex or regulated conversations — anything clinical, legal, financial needs a human, and the assistant should take a message, not advise.

B: No calendar to write to, and callers who genuinely won't talk to a machine — both real, both worth checking before you build anything.

A: One honest limit, too — a voice assistant doesn't reduce headcount. It absorbs the calls nobody was answering. The gain is captured revenue, not saved salary, and pricing it as the second one leads to disappointment.

B: One thing to check this week.

A: Look at your actual after-hours call volume for a week. If it's a handful a day and growing, that's the case for this. If it's one or two, it's probably not yet.

B: Written version's at atinnovators.in, under insights.