For founders
Case study: a voice AI agent for calls
A voice agent that takes calls is only as good as the screen a human sees after the call. What was said, what was promised, what happens next, and one click to correct the agent. I designed and built that screen. Here is how.
Prashant Kumar Singh · · 4 min read
In short
- The hard part of a voice agent is not the voice. It is the handoff: what a person sees after the call and what they have to do about it.
- The post-call screen has four parts: transcript, summary with what was promised, next actions, and one place to correct the agent.
- Corrections feed back into the agent's instructions, so the screen is also how the agent improves.
Voice agents are impressive in a demo. The voice is natural, the latency is low, the agent handles an interruption. What the demo never shows is the next morning, when a person has to find out what the agent said to forty callers and whether any of it needs fixing. That morning is the product. This case study is about designing for it.
The brief
A voice agent that answers inbound calls for a service business, handles routine requests, books appointments and takes messages, with a human team reviewing what it did. The client wanted the agent to take calls on its own. The team wanted to be able to trust it. Those two wants meet on one screen.
What the agent produces
For every call: a recording, a transcript, a structured summary (caller, reason, outcome), a list of commitments the agent made (an appointment time, a callback, a quote), and the actions it took (booked, messaged, escalated). That is a lot per call. The design problem was turning it into something a person processes in seconds.
Decision 1: lead with commitments
The thing a human most needs to know is what the agent promised on the company's behalf. So the post-call card leads with commitments: "Booked Tuesday 10am. Promised a callback about pricing." Each commitment links to the exact moment in the recording. The summary and transcript are one click below. A reviewer scanning forty calls reads forty commitment lines, not forty transcripts.
Decision 2: next actions as buttons
Every commitment implies an action: confirm the booking, make the callback, send the quote. The card offers those as buttons with sensible defaults, so the reviewer's job is to confirm rather than compose. Where the agent already completed the action, the button shows done and links to the evidence. The reviewer's time goes to the exceptions.
Decision 3: one place to correct the agent
When the agent gets something wrong, misunderstands a request, promises something it should not, the reviewer needs to fix the outcome and prevent the repeat. Both happen on the card. "Fix" edits the commitment or action. "Teach" opens a one-line correction that is added to the agent's instructions, reviewed by an admin, and applied to future calls. The screen that catches mistakes is the screen that reduces them.
| Card section | Contents | Reviewer action |
|---|---|---|
| Commitments | What the agent promised, linked to the recording | Confirm or fix |
| Actions | What the agent did, with evidence | Done, or do it |
| Summary | Caller, reason, outcome | Read if needed |
| Transcript | Full text with timestamps | Read if needed |
| Teach | One-line correction to the agent's instructions | Submit for admin review |
What was delivered
Six screens: the call queue, the post-call card, the commitment and action detail, the teach flow, the admin review of corrections, and a daily summary. Designed and built, with the agent integration, the transcript pipeline and the correction loop. The client is not named; the call scenarios shown in the designs are realistic service-business calls, not real recordings.
Questions people ask
- What is the hardest part of building a voice AI agent?
- Not the voice. It is the handoff to a human after the call: showing what was said, what was promised and what happens next in a form a person can review in seconds, with a way to correct the agent.
- What should a post-call review screen show?
- Commitments the agent made (linked to the recording) first, then actions taken with evidence, then summary and transcript one click away, and a single place to fix an outcome and teach the agent.
- How does a voice agent improve from human review?
- Corrections made on the post-call screen become one-line additions to the agent's instructions, reviewed by an admin and applied to future calls, so the review screen doubles as the training loop.
If you're a founder
Building an agent product? Send the screen your users see after the agent acts. The free teardown will show whether it leads with commitments or with output.
If you run an AI lab or data team
Evaluating voice or conversational agents has the same structure: grade the commitments and actions the agent produced, not the fluency of the transcript.
Sources