The situation
A high-volume email queue was consuming service-team capacity, stretching reply times, and carrying real cost — the kind of workload AI should absorb, if it can be trusted in production.
The bar was operational, not experimental: real customer emails, measured accuracy, and humans in the loop.
What we did
- 01Built the email AI agent on the company's Salesforce platform, inside the systems the service team already worked in.
- 02Ran a tight six-week pilot-to-production path: scoped the email categories, built, evaluated, and hardened for production.
- 03Put human-in-the-loop agents in the workflow — both as quality gate and as the ongoing accuracy measure.
The outcome
- In production after six weeks, managing 150+ emails daily.
- Reply times and handling costs reduced significantly.
- 90%+ accuracy, as rated continuously by the human-in-the-loop agents who review its work.