Bloomberg recently ran a feature on the wave of retail traders using AI tools to build "automated money machines" — systems meant to take on the hedge funds. It opens with Joel Rieger, a software sales executive in Los Angeles, who spent more than a year hunkered down in his home office building an automated options-trading program like the ones that "mint money" for Wall Street funds staffed by small armies of Ph.D.s.
He wrote the Python. He ran the simulations. He tracked hundreds of stocks. And when he was done, the results weren't much better than simply parking his money in an S&P 500 index fund. So he decided it wasn't worth the work.
Rieger's outcome is the most interesting part of the story — and the part the "AI money machine" headlines tend to skip.
The enthusiasm is easy to understand. The tools have genuinely gotten powerful. Frontier models can now generate code, reason about markets, and produce strategy skeletons at near-zero cost. The barrier to "building a trading bot" has collapsed. A generation of traders who would once have bounced off MQL5 can now, in principle, get a working prototype.
But that progress has conflated two very different activities.
Writing code is not building a trading system. Rieger could write the code. That was never the hard part. What he found — a year in — is that the hard part is whether the mechanism survives contact with real markets: costs, spread, slippage, session reality, partial and rejected fills, disconnects at the worst moment. An indicator that looks clean in a backtest is not a system. It's a hypothesis that hasn't been tested yet.
The "automated money machines" the headlines hype are usually something quieter: a person coded an idea, and the idea reproduced the market's baseline once costs were applied. Not because the coder was naive. Because building a durable system is a different discipline than writing a script.
The AI tools changed the part that was never the bottleneck.
The analyst, the indicator, the "edge" you code — that's the intelligence layer, and it's the part being commoditized by the week. A model that can summarize markets or draft a setup at near-zero cost makes the signal cheaper, not more valuable. Products and projects built around a clever model have always leased that edge, and the lease expires on the model's release schedule.
Notice what didn't get easier: making a rules-based mechanism run reliably, until it behaves the same way every session, across broker quirks and platform updates and market regimes. Session-aware logic tuned per instrument. Execution that doesn't change when the trader is scared, greedy, or down on the day. Honest accounting of every trade — including the ones that didn't close the way the plan said.
None of that improves when the model improves. All of it is what separates a dashboard that lies from one that holds.
Here is the uncomfortable truth underneath the boom: the tools got powerful, but the discipline required to deploy them responsibly did not get easier. If anything, it got more demanding — because the threshold for "good enough to accidentally risk real money" is far lower when anyone can generate a plausible-looking strategy in an afternoon.
A trading bot is discipline expressed as software. It observes, checks conditions, acts — or waits. No opinion, no discretion, no "just this once." That is the whole value, and it is not a feature an AI tool can hand you. It has to be designed, tested, and relied upon.
Joel Rieger wasn't wrong to try. He was right to stop. His real lesson — the one worth more than a year of backtests — is that the edge was never in the code. It's in whether the mechanism actually does something durable past a pretty curve.
The tools changed. The equation didn't.
The market changes. The rules don't.
