AI trading bots are booming: most quietly lose money

Three exchanges just handed AI the keys to your account. The fine print says you eat the loss.

Six of the best AI models in the world got $10,000 each and 57 days to trade. Every single one lost money. One dropped 30.8%.

That benchmark ran from January to March. Then, three weeks ago, Coinbase, OKX, and BNB Chain all shipped products that let an AI agent trade your crypto for you.

Nobody put those two facts side by side. I’m going to.

I’ve spent 10 years in financial services and fintech, and the last stretch building an algorithmic trading platform. So when the “AI trades for you” wave hit in June, I did what I always do. I read the disclaimers instead of the headlines.

What I found tells you more about the next 12 months than any launch keynote. AI trading bots are new. The risk transfer underneath them is very old.

The receipt: AI trading bots lost real money for 57 days

Arcada Labs ran a benchmark called Prediction Arena. They gave six frontier AI models $10,000 each and let them trade real money on prediction markets for 57 days, from January 12 to March 9.

Every model lost money. On Kalshi, losses ran between 16% and 30.8%. Not one of them beat sitting still.

These weren’t toy demos. They were the same class of system now being wired into live exchange accounts. The intelligence went up. The trading results did not.

Prediction markets are actually the easy case. They have clear outcomes and fixed settlement. Spot crypto, where these agents are now being pointed, is messier and more volatile. If the models struggled on the simple version, that should tell you something about the hard one.

The lesson isn’t “AI is dumb.” It’s that markets punish confidence, and language models are built to sound confident. A bot that’s certain it’s right, and wrong, loses faster than a human who hesitates.

I’ve watched real trading systems for a decade. The ones that survive aren’t the smartest. They’re the ones with the tightest rules about when to stop.

What actually launched in June

Between June 16 and July 1, three of crypto’s biggest names shipped agent products. The timing wasn’t a coincidence. It was a race.

Coinbase launched Advisor, the first in-app AI agent to hold SEC, CFTC, and NFA credentials at once. It gives round-the-clock trade ideas and can act on your account. Its own disclaimer says outputs “may be inaccurate or incomplete,” and the losses are yours.

OKX opened a marketplace on June 30 where AI agents hire each other and settle payment on-chain in stablecoins, after a closed beta of 50 providers. A day later, BNB Chain went live with Agent Studio, letting anyone spin up an on-chain agent with its own wallet in about 15 minutes.

There are already more than 150,000 AI agents on that chain. And BNB Chain, CoinMarketCap, and Trust Wallet are running a $36,000 hackathon specifically for AI trading agents that trade live on-chain. This isn’t a fringe experiment anymore. It’s the main stage.

I build on BNB Chain, so I watched this land in real time. The infrastructure is genuinely impressive. An agent with its own wallet, on-chain identity, and payment rails is a real unlock. But infrastructure that makes it easy to deploy a trader is not the same as evidence the trader makes money. Those are separate claims, and June only proved the first one.

Line up those three launches and the same shape appears. Every one moves the decision to a machine and the risk to you.

The sentence hiding in every disclaimer

I read the fine print on all three products. The wording changes. The structure doesn’t.

The agent decides. You sign. If it’s wrong, that’s your loss, not theirs.

A credential like an SEC registration tells you the operator followed a process. It does not tell you the bot will make money. Those are different promises, and the marketing blends them on purpose.

This is the oldest move in finance dressed in new clothes. For decades, the industry sold tools that pushed risk onto the customer while keeping the framing on the upside. Structured products did it. Copy-trading did it. An AI agent is just a faster, shinier way to run the same play.

And speed matters here. A human bag holder makes one bad decision an hour. An autonomous agent can make a hundred before you wake up. That’s not a bug they’ll patch. It’s the whole selling point.

The question was never “is the bot smart.” It’s “whose money is on the line when it’s wrong.”

To be fair, AI trading bots are good at one thing

I’m not here to tell you the whole category is a scam. It isn’t. There’s real utility in these tools, just not where the marketing points.

An AI agent is genuinely useful for research, summarizing market data, drafting a thesis, or watching for conditions you’d miss while asleep. That’s analysis support, and it’s valuable.

The failure starts when “help me think” becomes “trade my account unsupervised.” The Arcada results are what unsupervised looks like at scale. A co-pilot that flags ideas is a tool. An autopilot holding your wallet is a bet on a machine that’s rewarded for sounding sure.

Use the first. Be very slow to trust the second.

Why I built the platform on the opposite answer

When I started building BoBe, I kept coming back to that one question. So I inverted it.

BoBe runs a proprietary trading engine on BNB Chain, but it trades with its own capital, not yours. Users don’t hand over an account or pool funds into a bot. They acquire the platform’s utility token and lock it into a smart contract we call the Bakery. From there, 75% of the platform’s revenue is redistributed daily as USDT cashback, proportional to each participant’s share.

That cashback is variable. It can be zero on a given day. And every distribution is on-chain, so anyone can audit it without trusting my word for it.

The point isn’t the mechanism. The point is the risk boundary. You’re a platform participant, not a bot operator gambling your own stack on a confident machine. If you want to see how the cashback distributes on-chain, the contract activity is public at bobe.app.

I’m not telling you agents are useless. I’m telling you to know which side of the trade you’re standing on.

What the boring money is doing instead

While the agents were losing 30% in the benchmark, the unglamorous corner of crypto kept compounding. Kraken’s DeFi Earn was advertising up to 5.92% on stablecoins, with named risk operators disclosed and audited vaults underneath.

That’s the real competition for a $10,000 allocation. Not “which AI is smartest.” It’s “audited and boring” versus “autonomous and confident.”

For most people with money they can’t afford to torch, boring keeps winning. The best investors I know build systems with hard limits. They don’t hand the wheel to whatever demoed well last week.

Here’s the tell. When a product leads with how smart the AI is, it’s selling the demo. When it leads with what happens on a bad day, who holds the risk and how you verify it, it’s selling a system. The second kind is rarer and worth more.

None of this is financial advice, and crypto stays volatile no matter who or what is trading. But the frame matters more than the pick.

So here’s what I’d actually do

If you’re weighing any “AI trades for you” product, run it through five checks before you connect a wallet:

• Find the disclaimer first. If the loss is yours and the upside is theirs, you’ve learned the real deal. Read it before you watch the demo.
• Separate the credential from the performance. A registration is a process badge, not a profit promise. Don’t let one stand in for the other.
• Ask who trades what capital. Is the platform risking its own money, or routing yours through a bot? That single answer reorders everything else.
• Demand on-chain proof. Screenshots aren’t receipts. If you can’t verify the activity on a block explorer, treat the numbers as marketing.
• Compare against boring. Benchmark any agent against a plain audited alternative. If it can’t beat 5% without 30% drawdowns, the risk isn’t priced for you.

The next 12 months

The agent wave isn’t slowing down. More exchanges will ship them, and the demos will keep getting slicker. That’s fine. New tools are new tools.

Just remember the two facts nobody stacked together. The smartest models available lost money over 57 days, and the biggest platforms just made it easier to let one trade yours. The technology changed. The question didn’t.

Before you let anything trade for you, decide whose loss it is when it’s wrong. If you want to see what the other side of that looks like, where the platform trades its own capital and the cashback is on-chain and public, that’s the whole idea behind what I’m building at bobe.app.

AI trading bots are booming: most quietly lose money was originally published in Coinmonks on Medium, where people are continuing the conversation by highlighting and responding to this story.