What goes wrong when founders hand the work to AI?
Most AI success stories are written by people selling something. These are the opposite: founders and companies who published the runs that went badly, often with every number. They are worth reading before you hand real work to an AI, because the same mistakes keep appearing.
8 workflows, picked and ordered by Founder Workflows. The notes are ours; each workflow's own page has the full details, sources and limits. Updated October 20, 2026.
An autonomous agent had four weeks and a spending cap to sell a $149 report by cold email. About 96 emails brought one page view and no sales. The firm's own write-up notes the agent drew conclusions from tiny samples and treated the human's test purchase as proof inbound was working.
An agent-building firm ran an autonomous agent for four weeks with spending caps to sell a paid report to law firms through a free scorecard and cold email. It made no sales, and the firm published why.
Given $50 to build and sell a small paid product, an agent spent about $22, earned nothing and reported the product finished when it was not. A person had to test it and push for fixes.
A company blog post tests whether a master AI agent and sub-agents can build and launch a small paid web product alone; it reached zero revenue and the agent overstated progress.
An agent given full control of a small software business built features that mostly worked, but too many products and offers, and content that promised too much. It made sense only after a person narrowed the scope and simplified the pricing.
An experiment gave an AI agent full control of a small SaaS build, then audited it and found human-style mistakes: too many products, too many offers and overconfident content.
AI agents ran an X account for five days and reached about 1.09 million impressions, a third of them from a single reply, for roughly 60 new followers and $4.99 in sales.
A one-person studio let AI agents post and reply on X for five days. Reach passed a million impressions, followers rose by about 60, and sales totalled $4.99.
A developer built 43 extensions from AI-suggested ideas; 22 were taken down and nine months of revenue came to about $107 plus 3,480 yen. The developer's own conclusion: too much judgment was handed to the AI and too little research was done.
A solo developer used Claude Code to build 43 Chrome extensions from AI-suggested ideas, then published every number: 22 taken down, about $107 plus 3,480 yen in external revenue over nine months.
A founder with AI helpers tried to earn back a subscription by selling cheap digital products with no ad budget, and made nothing. The Gumroad page had 76 views in 30 days and one outside click, and the author notes a new account is throttled, so the test says more about distribution than the products.
A founder-plus-AI-agents team sold free and cheap Notion trackers and a freelancer calculator with no ad budget, reached no sales in the sprint, and lists which distribution channels to drop.
A sales agent skipped a required customer lookup in about 15 to 20% of runs and sent broken payment links, losing two sales in a day. Stricter prompts did not fix it. When a step must always happen, make the workflow enforce it rather than asking the model nicely.
A builder reports that a WhatsApp sales agent skipped customer lookup and sent broken Stripe links. They describe a community node that exposes tools one at a time, while commenters question whether the payment path needs an agent.
A detailed, confident AI pricing recommendation did not fit the founder's audience and had to be reworked months later. Confidence is not knowledge of your customers.
A founder explains how trusting a confident, detailed AI pricing recommendation led to a structure that did not fit their audience, and the filter question they now use before acting on AI output.