An AI email team for a nationwide industrial-services brand
Seven role inboxes, zero staff assigned to them, replies out in under 2 minutes at any hour — with a human still approving everything a prospect sees.
A nationwide industrial-services brand routes all inbound email through 7 role inboxes and had no one dedicated to answering them. We built a team of named AI agent personas that draft and send replies in under 2 minutes around the clock — with escalation alerts, a complete audit log, and one-click human approval on every sales reply.
The problem
The company runs a consumer-facing brand and routes every inbound email through role inboxes — hello@, sales@, dispatch@, billing@, vendors@, and more, 7 in total. That structure is good discipline, but nobody owned the answering. Messages sat until the owner got to them, which meant hours or days depending on the week. Prospects cooled off, vendors chased, and billing questions stacked up.
Hiring someone to watch 7 inboxes for a message volume that arrives in bursts made no sense. What the business needed was coverage, not headcount.
What we built
We built a set of named AI agent personas — one per role. Each agent has its own role definition, personality, and system prompt stored in a database, so tuning how the billing agent talks never touches how the sales agent talks.
A router polls the inbox every minute. It identifies which role inbox a message was addressed to, drops bounces and self-mail, fetches the matching persona from the database, drafts a reply with an LLM, and sends it from that agent's own address. To the person on the other end, it reads like a small, responsive team — because functionally, it is one.
The guardrails
An email system with no human in the loop is a liability, so most of the build is the safety layer:
- Escalation scoring. Every inbound message runs through keyword and heuristic checks for anger, legal threats, urgency, and excessive caps. Anything that scores high alerts the owner by Telegram and email instead of getting an automatic reply.
- A full audit log. Every interaction — what came in, what the agent sent, when — is written to a log. 100% coverage, nothing invisible.
- A hard-coded fraud refusal. The billing agent will not act on bank-detail-change requests, period. That request is a common fraud vector, so it is refused in code, not left to the model's judgment.
- Human approval on sales. Sales replies are never auto-sent. They queue into an email showing the prospect's message and the draft side by side, and the owner approves with one click.
What changed
Replies now go out in under 2 minutes, at 2pm or 2am. All 7 inboxes are covered without a single person assigned to them. The owner's email involvement shrank to two things: approving sales drafts and handling the rare escalation. Everything else answers itself — and every one of those answers is in the audit log if anyone ever wants to check the machine's work.
Total elapsed build time: about a week.
Drowning in inboxes nobody owns?
A 30-minute call is enough to tell you whether an agent team fits your email volume. No deck, no pressure.
Book a discovery call →