Do AI Trading Bots Work? I Tested Them With Real Money

I gave three AI trading bots $999 of my own money.

Real exchange. Real trades. Every single one shown live on a website anyone could watch.

Thirty days later they had $786.

That’s minus 21%, on a month when just holding Bitcoin lost about 3.6%.

So, do AI trading bots work? As software, yes: they watched the market around the clock and placed trades without me. As a way to make money, no. Doing nothing beat them, and so did 200,000 bots that just flip a coin.

And it’s the second time I’ve run this experiment. Here’s everything, with the numbers, the tools and what I think it means for anyone looking at an AI trading bot.

Key takeaways

• $999 in, $786 out after 30 days: minus 21.4%. Add the roughly $42 I paid for AI calls and it’s about minus 25%.
• Doing nothing beat the bots. Holding Bitcoin over the same 30 days was about minus 3.6%. Holding Ethereum was about minus 12%.
• They won 40% of their trades (51 of 129), and the best day of the whole experiment was day one.
• Every fix was “make the AI smarter”. Three rounds of smarter AI took the bots from losing money to making exactly 22 cents.
• 200,000 bots that just flip a coin did better. 35% of them made a profit over the same 30 days. Only about 3 in 100 did as badly as mine.

The experiment: three AI crypto trading bees and $999

I called them the bees.

Breezy, Bizzy and Boozy, $333 each, trading crypto on a real exchange from 25 September to 25 October 2026.

Every trade, every decision and every balance was live on beebots.tech. Over 75,000 people came and watched.

They even had their own live display on my desk: an ESP-Mosaico tile running firmware Gemini 4 Argon wrote.

The engine is open source if you want to check how it worked: github.com/imikerussell/beebots.

Here’s what ran them.

• Jev (by TypeSafe) picked every trade in week one.
• Zapier ran the Beekeeper every hour.
• Claude Opus 5.5 was the Beekeeper’s brain: an AI whose only job was to watch the bees and rewrite their rules.
• Gemini 4 Argon powered the final rebuild for the last 12 days.

Gemini 4 Argon and Zapier also run my AI sponsor finder.

Watch the experiment from 1:19

What happened: 30 days in one chart

Download the hourly data (CSV): every bee’s balance, every hour, 25 September to 25 October 2026.

The peak was $1,029.92.

On day one.

It never got back there. The low point was $781.65 on 18 October, about 24% below that peak.

One honest note on the chart: the flat line from 22 to 25 October is a 76-hour freeze in my engine. It stopped recording until the server restarted it. The balance was about $790 going in and about $789 coming out, so it doesn’t change the result.

Every fix was “make the AI smarter”

This is the part I think matters most.

Every time it went wrong, the answer was the same. Smarter AI.

Week one: the AI picks the trades. Minus 13.3%. One winning trade out of sixteen.

Then the Beekeeper. Opus 5.5, running on Zapier every hour, rewrote the bees’ rules 29 times. It was furious most of the time. Minus another 9.3%.

Brilliant lines. Worse results.

Then the rebuild on Gemini 4 Argon. Twelve days. It made 22 cents.

That’s the smartest version I built. It stopped losing. It never started winning.

Watch the “smarter AI” section from 2:58

How each bee finished

• Breezy was up 7.7% on 8 October. It gave it all back and finished minus 4.5%.
• Bizzy made 49 closed trades with 20 winners. Minus 14.9%.
• Boozy, my degen, went into meme coins in week one. Minus 44.6%. Its worst single trade lost $22.19.

Between them they traded $72,309 of positions on a $999 stake. That’s about 72 times the money. The exchange took a fee on every one: $37.64 in total.

I’d done this before: the OpenClaw lobsters

Back in February I ran the same idea with OpenClaw.

Three AI agents, trading on their own, with real money, live on my YouTube channel.

Twenty-two days in, they were about $100 up.

Sounds good.

Except it was costing me around $530 a month to run: API credits, servers and the X API. So really I was losing.

That series never got a final video. There was no happy ending to film.

And something I said back then has stuck with me. These agents report their own balances, and AI sometimes lies.

Watch the lobsters callback from 4:11

The control: 200,000 coin-flip bots

So why is your feed full of people turning $50 into $5,000 overnight?

I ran one more test.

I made 200,000 fake trading bots. No AI. No brain. Each one just flips a coin: buy, sell, or sit it out. Then I ran them on real Bitcoin prices over exactly the same 30 days as my bees, with trading fees and an AI bill taken off.

35% of them made a profit.

From a coin flip.

And only about 3 in 100 did as badly as my bees. To be fair, the coin-flip bots only traded Bitcoin, and my bees traded wilder coins.

I also ran it across eight years of Bitcoin prices. After any random 30 days, about 40% of coin-flip bots are in profit.

So give 1,000 people a trading bot that literally flips a coin.

After a month, about 400 of them have a profit screenshot.

Those 400 are the ones who post. The other 600 stay quiet.

That’s before anyone fakes anything.

Watch the coin-flip test from 5:47

The posts you’re seeing

Some of them are faked. Some are copied.

One post, “A 19-year-old student built a trading bot with Claude Code in 2 days, first night: $6,732 profit”, went out word for word 14 times from 7 accounts. Together, over 2.1 million views.

In the first copy the student is American. A day later he’s Japanese.

Another post, a Grok bot that supposedly turned $50 into $5,273 in 48 hours, picked up 1.27 million views.

X readers added a Community Note saying the dashboard in the video was fabricated, and that the account was trading months before the bot supposedly started.

On 26 October it was still up.

I’ve hidden every name, handle and avatar. The point isn’t the people. It’s the pattern.

Is AI trading a scam? The regulators have been saying this for years

The US Commodity Futures Trading Commission put out an advisory with this title: AI Won’t Turn Trading Bots into Money Machines.

Their words: “Instead of getting automatic money-making machines, customers lost tens of millions of dollars.”

The biggest case they’ve brought, Mirror Trading International, took at least 29,421 Bitcoin, worth over $1.7 billion, from at least 23,000 people in the US and thousands more worldwide.

There was no bot. It was a Ponzi scheme.

In September 2026 the SEC charged two more schemes that sold AI trading bots. Neither did any trading at all. At least $15 million.

Here in the UK, the FCA says nearly two thirds of 18 to 29 year olds follow finfluencers, and three quarters of those trust their advice. The regulator’s own description of these followers: “often young and potentially vulnerable”.

When seven influencers with 4.5 million followers pleaded guilty to promoting an unauthorised trading scheme, the smallest fine was £600.

Three more were charged in 2025. Their trials are in 2027.

So is nothing being done? Not nothing. But nowhere near enough, and nowhere near fast enough.

Watch the regulators section from 7:24

Who is actually making money

Somewhere along the line, someone is making money from all of this.

Go back to the viral post about a Claude bot that made $78,083 on Polymarket in five days. Read to the end and the instructions are: comment a word, like and repost, follow so they can DM you.

That’s the product. Not the bot. You.

Your follow, your comment, your attention. Next week it’s a course, a bot template or a coin.

And the exchange gets paid on every trade, win or lose.

What people who build trading bots told me

The comments under the video were some of the best I’ve had. People who actually build these things, mostly agreeing, with their own numbers. Paraphrased, without names:

• A full-time trader is shadow trading his own bot on the live market for months before risking anything, and won’t sell it.
• Someone spent six months building and testing strategies, by AI and by hand, with a live paper-trading platform. After costs, the best one only matched a near risk-free benchmark.
• Another runs 20 bots side by side as an experiment. After two months, plain buy-and-hold Bitcoin was the clear leader.
• More than one said the same thing in different words: AI is great for building and testing the machine. It shouldn’t be making the trades.

Months of work, real testing, honest results. That’s the opposite of “plug it in tonight, wake up rich”.

My open challenge

If you’ve really got an AI trading bot that makes money on its own, show me.

Not a screenshot. Not a backtest. Not “first night, $6,000”.

Give me read-only access to a live account, starting from the day you accept. Leave it running for six months and don’t touch it. Beat simply holding Bitcoin, after fees and after your AI bills. Steady gains, not one lucky bet.

I don’t think it can be done. I’d genuinely love to be proven wrong.

FAQ

Do AI trading bots actually work?

They work as automation: they can watch markets around the clock and place trades without you. In my 30-day real-money test they lost 21.4%, while simply holding Bitcoin lost about 3.6%. Working software is not the same as a working strategy.

Do crypto trading bots work?

Not in my test. Three AI bots traded crypto on a real exchange with $333 each from 25 September to 25 October 2026. They turned $999 into $786, minus 21.4%, while simply holding Bitcoin over the same 30 days lost about 3.6%. They won 40% of their trades and still lost money.

Is AI trading a scam?

AI trading itself isn’t a scam, but the promise of easy money often is. The CFTC and SEC have shut down several “AI trading bot” schemes that never traded at all. If someone promises guaranteed profits while you sleep, treat it as a red flag.

What is the success rate of AI trading bots?

My bees won 40% of their trades and still lost money overall. For a sense of luck: 35% of 200,000 coin-flip bots made a profit over the same 30 days. A profit screenshot from one month tells you almost nothing about skill.

Can you make money with an AI trading bot?

Some people who already know how to trade use AI to build and test their systems, and the honest ones talk in months and small margins. Neither of my experiments made money once costs were counted. If you try it, paper trade first and compare against just holding.

Watch the full video

Everything above, with the bees, the lobsters and the posts, is in the full video: AI Trading Is a Lie: I Lost Real Money Twice.

This is not financial advice. It’s the opposite, really: I spent my own money so you don’t have to spend yours.

If you’d rather build with AI the honest way, that’s what the Creator Magic community is for.