Case Study — Production AI Agent

Sales Inbox Agent: 80+ inbound leads answered a day, unattended.

An AI agent that reads the sales inbox of a global project-management SaaS (8M+ downloads — anonymous at the client's request), classifies every inbound email, and sends the right personalized response within minutes. It runs on a five-minute schedule in production and reports to the owner every evening.

82Replies on the day captured

7,693Replies since launch

5 minRun interval, 24/7

0Human touches per reply

The agent's end-of-day report for July 2, 2026: 201 agent runs, 15 trial responses and 67 registration responses sent that day, zero errors, and cumulative totals of 7,693 replies since launch. Customer email addresses are blurred and client branding removed. — Scroll · Production data The agent's real end-of-day report — 82 responses sent that day, 7,693 since launch, zero errors. Customer emails blurred; client branding removed at their request.

The problem

Hot leads were sitting unanswered.

The company's sales inbox receives a steady stream of trial requests and new-user registrations — dozens a day, around the clock, from every timezone — mixed with a constant drizzle of spam and bot submissions. Answering fast matters: a trial request answered in five minutes converts very differently from one answered tomorrow morning.

Doing this by hand meant the founder burning hours a week triaging email. Doing it with a dumb autoresponder meant replying to spam and sending trial instructions to people who asked something else entirely.

What we built

An agent that owns the inbox end-to-end.

Two-stage classification. A deterministic pre-filter catches obvious form submissions and spam patterns without an AI call; Claude handles everything ambiguous. Most messages never cost a token.

Personalized template replies. Trial-request and registration responses rendered with the sender's extracted name and details, sent through the Gmail API with the client's branding.

Dedupe & labeling. A sent-log guarantees no one is ever double-replied; processed mail is labeled and marked read so the inbox stays clean.

Daily accountability. An end-of-day report emails the owner every send, skip, and error, plus cumulative totals since launch. Autonomous, but never opaque.

Operational safety. A dry-run mode that classifies without sending, a test-recipient mode that redirects all outbound mail, and a one-command kill switch.

How it works

Watch, classify, respond — every five minutes.

The production log, live

Below: the agent classifying a trial request and a registration deterministically (no AI call needed), sending both replies seconds later, and the pre-filter catching bot signups — followed by a live dry run demonstrating the safety tooling. Customer emails and the client's name are masked on-screen.

Terminal showing the production agent log: SENT events for trial and registration replies, DETERMINISTIC classification hits, PREFILTER spam catches, followed by a DRY_RUN=1 execution and its run summary. Email addresses and client name are masked. — Production log Real send events from the live agent, then a dry run — classify everything, send nothing.

The right response, personalized

Each category gets its own branded template, rendered with the lead's extracted first name and details — trial requests get qualification requirements built in, registrations get an onboarding welcome. (Templates carry the client's branding, so they're not shown here.)

Report back, every evening

At 5 PM the owner gets the report shown at the top of this page: every reply sent that day with timestamps, spam ignored, dedupes skipped, errors (none), and running totals since launch.

All screenshots are from the live production system. Customer email addresses are redacted in every image and the client's branding is removed at their request; volume figures (82 replies that day, 7,693 lifetime) are the system's own reported numbers, unedited.