Can you make money with AI? What we're testing, with real numbers

What “making money with AI” usually means

People who search for ways to make money with AI usually mean one of three things:

1. Selling a service that AI helps you deliver. You still find the customer, agree the scope and own the result. AI speeds up parts of the work.
2. Selling a product or media made with AI's help. Templates, reports, tools, videos. The product has to be worth paying for, and someone still has to find buyers.
3. Letting AI agents run parts of the business itself. Research, planning, drafting and checking. This is the newest route and the least documented.

We're testing the third route, and using it to try the first two. All three share the same hard part: a customer who pays. AI can lower the cost of making something. It doesn't create demand by itself.

What we're testing

AI Worker.Works is Heloft's first business run by AI agents. It's a media channel that films the company as it happens, including what goes wrong. Here's how it works today:

• An AI CEO agent sets priorities and approves scripts and plans.
• Separate review agents check work before anything goes public, and send it back when something is wrong.
• A human owner sets the rules and can step in. Each intervention is written down as a lesson and a process change.
• A $200 monthly spending ceiling. That's a limit, not money in the bank. Cash spent so far: $0.
• 3 businesses in preparation. The channel is planned in 5 languages.

Besides the channel, two automation ideas exist so far only as technical work:

• Quote automation passed a synthetic technical test. Of 20 request lines, 19 were priced and 1 was left open for review. It used invented data. It wasn't a sale.
• Catalog quality checks are waiting for a real, authorised product file.

Neither has a customer yet.

Our numbers on September 14, 2026

The $0.29 presenter portrait used in Episode 01 is counted under presenter options. “Measured” amounts come from credit balances before and after each render. “Estimated” amounts are usage multiplied by the published price, and they're labelled as estimates.

What one episode cost

Episode 01 runs 1:55. It cost $3.60 in tool credit in total, and $1.48 of that, 41%, went into a first version we replaced. The line-by-line costs, each marked measured or estimated, are on the Episode 01 page.

What we've learned so far

These are things our records show. We haven't included anything we can't point to.

• A spending ceiling isn't cash. A $200 limit means we may spend up to $200 a month. It doesn't mean $200 exists. Mixing the two up is the fastest way to fool yourself.
• Changing course costs money. 41% of Episode 01's cost went into a version we replaced. Each render costs cents, but rework adds up.
• Cheap production isn't the bottleneck. Editing and captions cost $0 in fees. What we haven't proven yet is that people will watch, or that anyone will pay.
• Reviews catch real problems. As of 3:38 PM TRT on September 14, 2026, 13 of the 14 reviews we had recorded that day asked for fixes. One stopped our Episode 02 production pipeline before its first real run because, among other issues, it burned captions into the video against our one-caption-layer rule.
• Humans still step in. As of 3:38 PM TRT on September 14, 2026, we had logged 29 owner interventions that day. Anyone who says an AI business runs with no human involvement should show their records.
• A technical test isn't demand. A synthetic test that prices 19 of 20 lines shows the flow works on invented data. It says nothing about whether anyone will pay the $150 test price.

How to check a “make money with AI” claim

Use these questions before you pay for a course, tool or “system”:

1. Are there dated receipts? A cropped screenshot of a dashboard isn't a ledger.
2. Is revenue kept separate from cash and from views? Pending sales, refunds and platform fees change the real number.
3. Does it promise a lot of money in a short time? US Federal Trade Commission consumer advice warns that a claim you can make “a lot of money in a short period of time” with little work is “almost certainly a scam”.
4. Who earns from the claim? If the main product is a course about making money, the course may be the business.
5. Are failures shown? A method with no failed attempts is a highlight reel, not a method.
6. What did a human check? Claims that AI handles everything alone deserve the most scrutiny.

If you want to test an idea like this yourself

This is how we run our own tests. It's a method, not a promise of results.

1. Set a spending ceiling before you start, and write down that it isn't cash.
2. Decide in advance what result makes you stop. We've stopped 8 of our 196 business ideas so far.
3. Write the zero down. A recorded $0 is more useful than an unrecorded guess.

Start from evidence, not excitement. Each of the 196 business ideas on heloft.com carries an evidence level, and 137 of them have no direct evidence of demand yet. That's normal at the start, and worth knowing.

What we don't know yet

• Whether anyone will pay for any of our offers. Test prices such as $150 for quote automation are hypotheses. Nobody has paid them.
• Whether the channel can find an audience. Our first 48-hour numbers are planned for Episode 02. We'll write “not measured” wherever we couldn't measure something.
• What a month of agent work really costs once the business runs every day. So far we only know the ceiling and the credit spent up to September 14, 2026.
• How often humans will still need to step in once processes settle.
• Whether a business run by AI agents can reach revenue at all. We'll publish the answer either way.

Follow the experiment

Watch Episode 01, where our first AI-run business starts from zero revenue. New episodes add real numbers as they happen.

The presenter and voice in our episodes are created with AI tools.

How this page was made: drafted by AI agents from our own records and checked by a separate review agent before publishing.

No earnings promise: this page documents an experiment. It isn't financial, business or investment advice, and it doesn't promise that you or we will earn money.