What they did
Verdan built a bot to promote and sell their PDF guides. It writes and publishes its own posts on two social networks and leaves unprompted comments under posts by people it has never met, with a profile that says it is a bot. Each draft goes through a style check and then a separate AI review pass that can send it back or block it; a blocked answer to a question is replaced by a reply written in advance. Verdan switched it to publishing on its own on 13 July 2026, after 19 of 50 planned test samples, in 12 of which the review pass matched their own verdict. A daily report includes an alarm that counts messages still waiting for an answer and names the oldest. A separate agent handled Verdan's inbox but was told to leave social replies to the bot.
How it went
From go-live to 3 November the bot published 210 comments; 144 got no reaction. One stranger replied 17 times over ten weeks, in a thread that drifted into a security hole in his app. The bot pointed out a flaw in one of his conclusions and he ran the checks it named. One of his replies, from 7 October, never got its drafted answer: a usage limit hit during the night run, and answers waiting for review had no retry, so the draft was marked failed for good. The alarm named that reply in 14 daily reports, but Verdan had decided in September not to read the report and only found out on 3 November. A developer agent added the missing retry on 4 November.
Worth knowing
Unattended bots fail quietly at the step that has no retry, and an alert only helps if someone is named to read it. The bot deletes stored conversations after 30 days, so it lost the context needed to answer the stuck reply before anyone noticed. The post does not say which model or platforms it uses, or whether it sold any guides.