A support agent that resolved routine cases and explained every action
An evaluated, permission-bounded agent handled tier-one support while sensitive replies and account actions stayed with people.
The situation
Support specialists spent most of the day gathering account context and answering repeat questions across email, knowledge, and internal tools.
The constraint
The agent could not expose financial data, make irreversible account changes, or send sensitive replies without approval.
The approach
We designed the program around operational continuity and evidence. Each release had to prove it was at least as safe and understandable as the process it replaced.
Built evaluations from historical cases before connecting tools.
Ran draft-only and shadow modes against specialist decisions.
Expanded autonomy by intent only after quality and policy thresholds held.
What shipped
- An intent and evidence retrieval layer
- A tool-using support agent with scoped permissions
- Trace review, evaluation, and approval workspaces
The results
71%
resolved end-to-end
63%
faster first response
100%
actions traceable
“I trust the dashboard because I can open any answer and see exactly why it happened.”
Representative VP Customer Operations