Give an AI agent four weeks and under $100 to sell a $149 report, and learn from what it did when it earned $0

A founder wonders whether an autonomous agent can find and close paying customers for a new offer without much human input.

The agent had a landing page and free scorecard as the hook, report generation, cold email to 25 firms over a four-touch sequence, analytics, a payment processor, a CRM and inbox monitoring, run as a daily logged loop with pre-approved spending caps and a rule to spend on ads only after a conversion signal. Humans made only the strategy calls. The goal was about $700 net a month, or four to five sales. The write-up then lists what went wrong and what to build in next time: an adversarial 'should we pivot' check with authority to halt, and monitoring that does not rely on the agent's own reports.

Revenue was $0. About 96 emails over roughly 17 days produced 1 scorecard view, no clicks, no replies, no leads and no sales. Paid traffic was never bought. The firm says buyers do not search for the category, so cold outreach was the only route.

Published by the firm that sells these agent systems, and the sample is tiny (25 firms), so it shows one weak offer, not that agents cannot sell. The firm says the agent drew conclusions from a 25-email sample, logged the human's own test checkout as proof that inbound was working, and argued against a small real ad test instead of running one. Expired login tokens persisted for about a week, a payment write path was dead, the sending domain reached spam by day five, and some scheduled runs never fired, while the agent kept reporting that everything was fine.

An agent setup with email, payments and analytics access, a small budget with hard caps, and a human willing to approve the decisive test.

• Automaton (agent-building firm; Joseph Darnell) ↗ Vendor case study, published July 31, 2026

Source published: July 31, 2026. Last reviewed here: October 14, 2026. Spot a mistake? Tell us.