Powsoo is a drop-in widget that adds visual intake cards to AI voice platforms. When a user is on a voice call with your agent, Powsoo renders synced cards in the browser so users can tap-to-confirm emails, phone numbers, dates, and other intake fields instead of spelling them out.

How do I add Powsoo to my voice platform?

Create a widget in the Powsoo dashboard and connect your Retell account, then drop one web component and one versioned script tag into your site. You add powsoo_show_card as a custom function on your agent. Setup takes about 5 minutes.

Which voice platforms does Powsoo support?

Retell today. Vapi and ElevenLabs support are coming soon. SMS fallback works for traditional PSTN calls.

Can I customize the look of the cards?

Yes. Powsoo supports full theming via JS configuration: colors, fonts, border radius, spacing. For deeper control, use Powsoo Atoms (React components) to build your own UI.

Where does the data go after a user confirms?

Powsoo fires webhooks the moment a card is confirmed, with the field name, value, and call session ID. Drop the payload straight into your CRM, database, or backend.

Does Powsoo work for phone calls?

Pure phone calls (PSTN) don't have a screen, so Powsoo's web widget can't render directly. Instead, Powsoo offers an SMS fallback: when the agent needs to confirm an email or phone number, the caller gets a one-tap SMS link.

Is Powsoo HIPAA / GDPR compliant?

Powsoo is built with healthcare and EU customers in mind. We're SOC 2 Type II in progress and offer DPAs and BAAs on the Pro plan and above.

Conversational AI voice agents: how they work and where they fail

Updated 2026-07-20

A conversational AI voice agent handles a full phone conversation end to end: speech-to- text turns the caller’s words into text, a language model decides what to say, and text-to-speech voices the reply, all over a telephone or WebRTC connection. They are good at conversation and weak at one specific thing: capturing structured data. Spoken emails come through right about 53% of the time, and 96.6% word accuracy can still mean only 77% exact match on a full identifier, because a single wrong character breaks it. The fix is a visual confirmation layer: show the transcribed value on screen and let the caller tap to confirm, so intake data is accurate without slowing the conversation.

How a conversational AI voice agent works

Three components run in a loop. Speech-to-text (ASR) transcribes the caller in real time. A language model interprets that transcript and decides the next turn. Text-to-speech voices the response. The whole loop runs over telephony or a WebRTC session, and platforms like Vapi, Retell, and ElevenLabs package it so you do not have to wire each piece yourself.

The weak spot: structured data

The loop is good at conversation and bad at transcribing rare, high-stakes tokens. Because ASR is optimised to minimise word error rate, it protects common words and leaves names, emails, and numbers exposed. The measured cost:

Spoken email addresses: about 53% captured correctly by standard voice AI.
Ten-digit phone numbers: roughly 82% fully correct even at 98% per-digit accuracy.
Any full identifier: 96.6% word accuracy can still be only 77% exact match.

The fix: confirm by sight

A visual layer renders the captured value on screen mid-call. The caller reads it, taps Confirm or edits one character, and the confirmed value flows to your backend. It does not slow the conversation. It removes the one step where a voice-only agent quietly loses the data you called to collect.

Frequently asked questions

What is a conversational AI voice agent?

An AI system that holds a spoken conversation over the phone or web: it transcribes what the caller says, uses a language model to respond in context, and speaks the reply back. It is used for receptionists, intake, and support.

Where do conversational AI voice agents fail?

On structured data capture. The conversation flows naturally, but emails, phone numbers, and dates get mis-transcribed because speech-to-text is weakest on rare tokens. One wrong character breaks the whole value.

How do you fix the data-capture weakness?

Confirm by sight. A visual layer renders the captured value on an on-screen card so the caller reads and taps to confirm, instead of verifying a character string by ear. It sidesteps the transcription bottleneck for the fields that matter.