The request was simple: stop the inbox from eating the first hour of every day. A government relations firm with active clients, cold outreach running, compliance obligations, and a steady stream of newsletters, calendar responses, and notification spam was drowning in email. The principal was sorting manually every morning just to figure out what actually needed attention.
The inbox was the bottleneck. Not because there were too many emails, but because every email looked the same until someone read it.
The architecture: four layers of filtering
The system processes every inbound email through four layers. Each layer is designed to handle a specific class of email and pass everything else downstream. Nothing gets dropped. Nothing auto-sends.
Layer 1: Hard rules (zero AI cost)
Before any AI runs, the system checks the sender against a set of hard-coded rules. Industry newsletters get routed to their own label. Internal team emails are flagged as no-draft so the system never tries to reply to a colleague. Compliance emails from regulatory bodies get escalated immediately. And a self-loop detector catches the system's own outputs and kills them before they re-enter the pipeline.
This layer handles about 15% of all inbound email at zero cost. No API calls. No latency.
Layer 2: Noise detection
Calendar accepts, out-of-office replies, DMARC reports, LinkedIn notifications, Google Drive share notifications, Substack digests, GitHub digests, Slack notifications, Zoom recording alerts. All labeled and archived automatically. Never seen by a human.
The noise patterns are fully configurable. If the client decides Slack notifications are actually important, they flip one setting. If a new noise source appears, it takes one line to add it.
Layer 3: Critical Check (AI Agent 1)
Everything that survives layers 1 and 2 goes to the first AI agent. Claude reads the sender, subject, and body and decides within seconds whether this email is an emergency.
The definition of "critical" is tight and specific: vendor service shutoffs, compliance deadlines within 14 days, legal notices, active security incidents, client system outages, or urgent client escalations with explicit time pressure. The agent knows every active client by name. Billing receipts are not critical. Routine login alerts are not critical. Only genuine emergencies pass this gate.
If an email is flagged critical, the system labels it, sends an immediate alert, and moves on. Response time: seconds.
Layer 4: Classification and drafting (AI Agents 2 and 3)
Non-critical emails hit the classifier agent. Claude categorizes each one: briefing request, positive cold response, client intel update, client concern, scheduling, new stakeholder, government official, or unknown.
A confidence threshold is enforced. If the classifier is less than 60% confident, it forces the email to unknown and flags it for manual triage. No guessing.
Only two categories trigger the draft writer: client concerns and scheduling. The draft agent writes in the principal's voice, follows 16 mandatory style rules, references real federal programs with real dollar figures, and produces a complete reply. Every draft lands in Gmail as a draft for human review. Nothing sends without approval.
The style rules are not generic. No bullet points. No headers. No contractions except one permitted phrase. No filler language. Every claim backed by a specific data point. Under 200 words. The system writes exactly how the principal writes because it was trained on their rules, their voice, their instructions.
The safety rails
Production-grade means things cannot go wrong silently. Internal emails never get drafted. Government officials never get an auto-reply. The system never re-ingests its own output. Every JSON parse has a three-tier fallback: direct parse, brace-walking, and regex extraction. If all three fail, the email escalates to a human instead of getting dropped.
The result
The team opens a clean, pre-sorted inbox every morning. Noise is gone. Critical items are already flagged. Everything else is classified and, where appropriate, pre-drafted. The first hour of the day went from sorting email to doing actual work.
The system runs on autopilot. The team trusts it because it never sends without permission and it never drops an email. It just organizes, classifies, and prepares.
Drowning in inbox hours?
Book a free bottleneck audit. We'll tell you exactly what we'd automate, what it would cost, and how fast we can ship it.
Book a Free AuditGet every teardown in your inbox.
One email per build. The full node-by-node breakdown, plus the JSON to fork it. No fluff, no spam.