Domain
Agentic sales
Duration
1.5 Months
Sales teams were sitting on thousands of hours of conversations but had no scalable way to extract insights. We designed a feature that turns every call into structured intelligence.
Team: Product Designer (Me), Product Manager, Head of Product, Sales Team, and Eng. Team
Led end-to-end product Design, from discovery with SDRs/BDRs to defining the prompt model, designing widgets, and owning error states. Collaborated with PM on scoping and engineers on JSON-to-widget mapping.
Sales teams capture valuable information in calls, but analysing them is difficult.
Insights buried in conversations
Signals like objections, competitors, and product feedback are hidden inside long call recordings.
Manual review is time-consuming
Sales teams must listen to entire calls to extract useful information.
AI summaries lack flexibility
Most tools generate generic summaries that don’t adapt to what different teams want to track. (generic summaries in big 2026? nah).
I Ran discovery with 6 SDRs/BDRs, observing post-call workflows; most skimmed recordings or relied on generic summaries due to time constraints.
SDR
Pain:
Important conversation details are scattered across calls and difficult to use later.
Need:
Structured insights and visual metrics from calls that can be used for presentations, reporting, and improving future conversations.
BDR
Pain:
Hard to quickly understand what resonated with the lead and what moved the conversation forward.
Need:
Clear insights that reveal customer intent, interests, and signals to help bring leads closer to closing.
⌛️
To quickly retrieve important conversation details from past calls without searching through recordings.
📊
To get structured insights and visual metrics they can directly use in presentations and reports.
👨🏻💻
To use call data to identify patterns and continuously improve their future conversations.
Working with the PM, we identified that the real problem wasn't the lack of data, it was that fixed templates couldn't reflect what each team actually needed from a call. Our design response was to shift control to the user.
DECISION 1
Define insights using prompts
Users specify what they want to extract from calls in simple natural language.
DECISION 2
AI extracts and structures data
Conversations are analysed and converted into consistent, structured outputs.
DECISION 3
Build a personalised dashboard
Insights are saved as widgets, creating a dashboard that updates automatically with every new call.
After multiple iterations, the workflow we finalised.
The hardest problem wasn't the interface, it was the unpredictability of AI output. A prompt like "What objections came up?" could return a list, a string, a null, or a nested object. I designed a widget system with different output types that kept the UI consistent regardless of what the model returned.
We also tried a tagging approach early on, dropped it when users couldn't define categories before seeing data. The prompt-first model was the insight that changed everything.
I kept the experience simple and structured with upfront transparency and error handling so users set things once and it works across.
A demo call ID kickstarts the experience, users build a personalized dashboard from their very first interaction, with zero setup friction.
Personalized prompts surface from call recordings. Widgets are generated with a live preview, an AI rationale, and quick-edit options, so users always know what they're adding.
Before applying, users choose - add a widget for this call only, or automatically for all future calls, giving them full control over how the dashboard evolves.
When no response is found, I guide users with close matches and alternative prompt suggestions they can explore.
⭐️ Final Dashboard Preview
How a smart, custom dashboard looks for each call - built directly from call recordings.
✅
90% faster call analysis
Insights generated automatically, instantly and consistently, instead of manually reviewing recordings.
✅
10+ types of conversation signals
Enabled teams to track 10+ types of conversation signals through customisable AI insight widgets.
✅
Eliminated repeated manual effort
Dashboards auto-generated for every call, removing repeated setup and saving hours of manual effort
This project pushed me to think beyond screens, understanding prompts, outputs, and system logic made me a far better collaborator with engineers & PMs.
Designing for AI taught me that uncertainty needs to be designed for, not hidden. Users trust the interface before they trust the output.
