Case study
Relay Agent Console
Designing trust for an AI agent that runs back-office tasks on a team’s behalf.
Client
Relay Labs
Role
Design Lead
Duration
8 months
Year
2025
The problem
01
Relay’s agent could reconcile invoices and update records by itself, but operations teams did not let it. They could not see what it planned to do, and they had no simple way to step in.
Pilots stalled at the approval stage, and most customers kept the agent in read-only mode.
Approach
02
We designed a plan-first console. The agent proposes its steps in plain language, shows the data it will touch, and waits for approval on anything above a risk threshold the team sets.
I worked with the ML team to expose confidence and reasoning without turning the interface into a log file.
Outcome
03
Customers moved the agent from read-only to autonomous mode for routine work, while keeping approvals for edge cases.
The approval pattern is now part of Relay’s public design guidelines.
Results
04
73%
Tasks run without manual edits
-58%
Time spent on reconciliation
12 hrs
Saved per operator each week
Process
05
