I want to take the word “AI” off the brochure for a minute and say specifically what is happening inside a modern retail trading product when it claims to use it. Including ours.
The category has reached the point where the phrase “AI-powered trading” tells the user almost nothing. Some products use a single off-the-shelf LLM call to format an explanation. Some run actual reinforcement-learning models trained on price history. Some are essentially decision-tree heuristics with a sticker on the front that says “AI.” The honest pitches and the dishonest pitches use the same words.
So: here’s the unsensational version of what AI is good for in a retail trading product, what it is not good for, and where I think the category is going.
What AI is genuinely good for here
There are three jobs in a retail trader’s life that a machine actually does better than the human, every time, and these are the jobs a serious product is automating:
1. Continuous monitoring of a defined rule set.
A human user says “if BTC drops below 58,000, exit.” That rule is trivial. The hard part is being awake at 3:47am on a Wednesday in March when BTC actually touches 58,000 for forty-five seconds before bouncing. A modern AI-driven execution engine is, at its core, a piece of infrastructure that holds the user’s rules between when they set them and when they’re triggered. The “AI” part is mostly about choosing among a set of pre-validated strategies given the current volatility regime, and adjusting parameters on the fly. The most valuable property of the system is not its cleverness. It is its discipline.
2. Emotion-free execution.
After a losing trade, the average human’s next trading decision is measurably worse than a coin flip. After three losing trades in a row, much worse. The AI’s next decision is statistically identical to the one before. This is, frankly, the single highest-ROI feature of any retail automation product. Not “smarter trades.” Not “better picks.” Just: the system does not get angry, scared, or hopeful, and the user can. This is the boring AI win and it is the one that matters.
3. Pattern-aware parameter tuning.
The thing genuine machine-learning models can do — that simple rule engines cannot — is adjust position sizing, entry timing, and stop-loss placement in response to the current market regime. A market making 0.5%/day moves is a different market than one making 4%/day moves, and a static strategy will underperform in both. ML models are reasonable, not magical, at noticing this and tuning. They are not magical, and any product saying otherwise is selling you the magical version of an unmagical thing.
What AI is not good for here
The three things “AI trading” products oversell, in increasing order of how much I want them to stop:
1. Predicting which asset is going up next.
There is no model that does this reliably, despite an enormous number of marketing pages claiming there is. Markets are not so kind that the same pattern keeps working once the pattern is known. The actual frontier-research models that try this are, at best, fractionally above coin-flip on short horizons, and that fractional edge is eaten by fees. If a product pitches “our AI finds the trades for you,” ask what their definition of “find” is. The honest answer is “we run an execution strategy on the trades you told us you wanted.” The dishonest answer is the brochure.
2. Calling tops and bottoms.
Same problem, more dramatic version. No AI calls tops and bottoms. No human does either. Anyone confident enough to promise this is either selling you a story or has not yet noticed they got lucky.
3. Replacing user judgment about risk.
The user has to decide how much of their money they are willing to lose. That is a values question, not a math question, and no model knows the answer for them. The temptation, as a product designer, is to “help” the user by setting that number for them. Resist this. It is the line between selling a tool and selling a managed fund, and the second one has rules that change everything.
What this looks like inside AutoCoin
At AutoCoin, the “AI” is doing three things and only three things:
- It selects among a known set of validated strategies — conservative, moderate, aggressive — based on user inputs and the current volatility regime.
- It executes those strategies 24/7 on the user’s own exchange account, with full audit trails and a non-custodial architecture so the funds never leave the user’s exchange.
- It tunes parameters within bounds the user set, never outside them.
Things AutoCoin’s AI is not doing:
- It is not predicting markets.
- It is not making the user money in any guaranteed way.
- It is not choosing how much risk the user takes. The user chooses. The system enforces.
This is a less exciting positioning than “our AI will trade for you and make you rich while you sleep.” It is also one of the only positionings that an honest operator can defend in front of a regulator or a serious customer, which is a useful filter on positioning in general. The reasons I built it this way are in Why I’m Building AutoCoin.
Where the category goes
My guess at the next 24 months, narrowly within retail crypto trading:
- The marketing-only “AI” products will get smacked, either by enforcement actions or by users churning when the strategies don’t live up to the brochure.
- The serious products — the ones treating AI as a discipline-enforcement layer rather than a prediction layer — will consolidate the high-quality end of the market, with materially lower CAC than the marketing-only products had at the same stage.
- The boundary between “trading bot” and “robo-advisor” will collapse, and the regulatory category will be roughly the same on both sides of that boundary. The companies that anticipated this will look obvious; the ones that didn’t will have a hard 18 months.
That last point is the whole argument of The Regulated Crypto Decade, if you want the long version.
If you’re a retail trader: ask the company you’re considering what specifically their AI does, and accept only specific answers. “Our model uses reinforcement learning to tune position sizing in real time” is a specific answer. “AI-powered next-generation trading” is not.
If you’re an operator: build the boring version. The boring version is the one that’s still alive in five years.
If you’re an investor: the operators who can clearly articulate what their AI does and doesn’t do are the ones who can be trusted with the harder, less articulate questions. Use the answer to this question as a filter.
I’d love to hear from anyone working on these problems from the model side. We learn faster comparing notes than competing on bullet points.
— James, autocoin.ai