Every customer reply used to be a trade-off between speed and quality. Relay removed it.
By putting an AI drafting-and-review layer in front of the team's inbox, the client cut first-response time by more than 3× while raising message quality and consistency — turning a stretched communications function into a competitive advantage, without adding headcount.
The challenge
A distributed team was fielding a growing volume of customer messages across email and chat. Responses were slow, tone drifted from one team member to the next, and onboarding a new rep meant weeks of shadowing just to learn “how we say things.” Speed and quality were in constant tension — every gain on one came at the expense of the other.
What we built
Relay is an AI-mediated communication layer built on Claude. It sits between the team and the customer: when a message comes in, Relay drafts a context-aware reply grounded in the company's knowledge base and past conversations, matches the house voice, and flags anything that needs a human decision. The rep reviews, adjusts if needed, and sends — every message still gets a person's judgment, but none of them start from a blank page.
It grew out of Aurora, our research line on LLM-mediated communication — the same questions about trust, tone, and reliability, applied to a real inbox.
How it works
- Ingests the incoming message plus relevant context — account history, knowledge base, prior threads.
- Drafts an on-brand reply with Claude, tuned to the company's documented voice.
- Runs tone and factual-accuracy evals before the draft ever reaches a rep.
- The rep reviews, edits, and sends — their corrections feed back to sharpen the system.
The stack
▪ Key takeaways
- Human-in-the-loop by design — AI drafts, people decide.
- Evals on tone and factual accuracy run before a draft is ever shown.
- Voice consistency is a system, not a style guide nobody reads.
- New reps ramp in days, not weeks, because the system encodes the house voice.