AI Call Note Summary
My Role
Senior Product Designer
Team
1 Product Owner, 1 Business Analysts, 1 Tech Lead, 1 Operations Manager, 1 Customer Service Representative, and 5 Developers
Timeline
3 months, kickoff to pilot launch
Impact
42 seconds saved per call · approved for rollout beyond the pilot
The Problem
When a customer calls, the customer service representative has to read through the case history before they can help — often while the caller waits on hold. On complex cases with long note histories, that read-and-connect step took two to three minutes. Prudential handles over 2 million calls a year, so cutting 30 seconds off the average call saves tens of thousands of hours.
Research → Direction
Through SME interviews and call listening, two things shaped the design:
Customer service representatives multitask constantly — piecing together long note histories while keeping the caller engaged — so a short, direct summary keeps their energy on the customer instead of the screen
Customer service representatives said wrong information over the phone is embarrassing and damages trust in the company
That last point set the hardest constraint: the tool had to stay useful even when the AI got something wrong. An AI summary that the customer service representatives couldn't verify would be worse than no summary at all, because they'd repeat it to a customer.
Designing for Output That Can Be Wrong
The summary pulls three things from the case history: the caller's name and relationship to the policyowner, case and claim summaries tied to the policy in the last 18 months, and any outstanding items.
I designed the summary as a starting point, not the final answer. Three safeguards let the customer service representatives check it, fix it, and report it:
Check it: links to the original case note. Every summary point links back to its source, so reps can fact-check in a click instead of trusting the output blind.
Fix it: a regenerate option. If a summary comes back [incomplete or hard to read], reps can regenerate it with one click instead of digging through every note themselves.
Report it: a feedback feature. Reps can flag bad output in the moment, giving us a path to correct it once the tool is live.
Scoping the Design
When to run it.
Auto-running the summary on page load seemed obvious — agents read case notes on nearly every call
Operations flagged that back-office processors use the page regularly and don't need the summary
Engineering flagged that running it for everyone would burn tokens and cost on traffic that never touches the feature
We moved it behind a button: customer service representatives trigger it when they need it, and we stopped paying to generate summaries nobody asked for
How far back to read.
Surfacing every case and claim in the full history pulled in outdated or resolved information and made summaries too long to scan quickly
After discussing the tradeoff with operations leaders and agents, we scoped the display to the last 18 months of cases and claims — the AI still reads the full history, but only recent activity surfaces in the summary
Validating Before Launch
Before the pilot went live, I used AI to quickly explore layout directions, then tested three with operations leaders and three customer service representatives in a roundtable format.
The group confirmed the core direction: keep the summary high-level, and let agents verify anything they're unsure of by clicking through to the source note. They picked the layout that best supported that balance, and that's the version that shipped to the pilot.
Results
To de-risk the launch, we piloted with a single call center — life insurance claims — for one month in late March 2026, compared to the same period in 2025.
42 seconds saved per call, averaged across every call in the pilot
Leadership approved rollout to the broader claim team, then the full call center
"I know AI can hallucinate and make things up convincingly. I was concerned at first, but the links next to the results let me fact-check against the original call note easily, which gave me more confidence."
— Customer service representative, pilot group
What I'd Do Differently
The pilot validated the concept for one business unit. Different units have different workflows, and I'd want to keep listening to calls and iterating against those differences before a broad rollout.
The 18-month cutoff for what surfaces in the summary came from stakeholder discussion, not usage data. With more time, I'd validate it against real call patterns with customer service representatives and customers — some cases may need a longer window, others less.
I didn't get visibility into whether customer service representatives flagged inaccurate summaries after the pilot launched, since my role on the project ended shortly after rollout. That's the one piece of evidence I'd want to validate that I created a trustworthy AI-enabled experience.
