98 Public AI-Money Experiments, 8 Still at Zero
A directory called Agent Earnings logs 98 public experiments where an AI was told to earn money, failures included. Eight still show zero. The worst burned 487M tokens, about $1,117 in model cost, for $0 revenue. Read the ledgers and the difference between the ones that paid and the ones that did not comes down to one thing.
📜 Table of Contents
Where these numbers come from
Agent Earnings is a small public directory that logs experiments where an AI agent was told to run a business and earn money. Every entry links its source and states how the figure was checked: read from a payment provider, self-reported, or estimated by a third party.
As of 3 November 2026 it holds 98 entries. Eight of them show zero revenue.
Most people only repost the screenshots with five-figure monthly revenue. This directory keeps the zeros. A few of them explain the whole problem.
Three zeros, three ways to die
autocomp ran an open-source "company" loop with CEO, CTO, QA and go-to-market agents, shipping a newsletter digest for 39 days. It spent 487M tokens, roughly $1,117 at API prices. Revenue: $0.
Agent Receipts sold four paid skills and promoted them on X. By day three it had about 500,000 views, 3 checkout opens and $0 in payments.
SWARM opened an x402 shop selling small utilities from $0.02 to $0.99. First month: $0 from outside buyers. Of the 18 payments in its ledger, 17 were its own tests and one was a $0.02 call from an automated checker.
Then there is Krimskrams: 17 pay-per-call data endpoints, $1.54 in three weeks against 5B+ tokens (about $7,000 at API prices). Anicca booked $230.83 of revenue against $679.88 of expenses in one month, netting minus $449.05.
The ones that earn are not doing the same job
Another 76 businesses in the same directory have verified monthly recurring revenue. Median: $4,380 a month. That median is a trap, though. Take the top 25 and the median is $16,716. Take the other 51 and it is $2,666. Same directory, same verification, six times the difference depending on where you cut.
The real pattern is not in the size of the numbers. It is in the type of work.
Almost every business that earned was catching a payment that already existed: QA work, selling leads, freelance gigs, bounty tasks. They took a slice of somebody else's existing budget instead of building a new building on empty land.
The ones that lost money were all building a new product and waiting for buyers to find it.
The first group has cash flow on day one, because the buyer is already paying someone for that job. It just is not you yet. The second group has to solve trust, traffic and conversion before the first dollar arrives, and the token meter runs the whole time.
Running it yourself: finding demand that already pays
Step one: start where money is already moving. Do not start from what you can build. Start from who is currently paying for this. Three reliable signals: repetitive job listings, the same request showing up repeatedly on freelance platforms, and the same question being asked every day in an industry group.
Step two: check whether the demand repeats. One-off work has zero retention. You want something that happens every month and costs someone a few hours each time. That is what makes a subscription possible.
Step three: automate delivery, not acquisition. This is the counterintuitive part. Across these cases, the more of the delivery you automate the cheaper it gets. The moment acquisition goes fully automated, reply rates collapse to template level. Brandon Doyle's postcard pipeline works because the final phone call is still made by a person.
Step four: price the inference cost per delivery before you price the product. autocomp spent $1,117 over 39 days, about $28 a day. At $50 per sale it needed 0.6 sales a day just to break even. That number should exist before launch, not on the invoice at month end.
What I take from it
The value of this directory is that it keeps the denominator. Behind every success story you see reposted, there are dozens of zero-revenue experiments nobody forwards.
"AI earns money automatically" is not wrong. It just drops the subject. What gets automated is delivery. Acquisition and trust do not. Assume they do and you get the eight zeros above.
If you want to actually run one, start from a job that already has a budget, automate 90% of the delivery, and keep acquisition human. The ceiling is low, but it is the only path that pays in the first month.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.