Three different things are sold as AI copy trading
The phrase covers products that have almost nothing in common, and the difference decides what happens to your account on a bad day.
- A follow button with an AI badge. The copying is still one to one. The “AI” is usually a ranking of traders, and once you follow someone you inherit their leverage and their worst week, exactly as before.
- A rule bot with a new name. Grid, DCA and Martingale bots follow rules written in advance. They are useful and predictable, and they are not intelligence. We explain the difference in how to tell real AI from relabelled automation.
- A model that decides each order. A language model reads the trade and its context, then returns a decision. This is genuinely model driven, and genuinely fallible, which makes one question more important than any other: what is the model not allowed to do?
CopyFi does the third, and builds everything around that question.
How it works on CopyFi
- You build a strategy. From visual blocks, in the strategy builder: a leverage cap, a size limit, a direction filter.
- You store a model key. Anthropic, OpenAI, or your own endpoint that speaks the OpenAI API. Storing a key changes nothing on its own.
- You set up a copy trading session and turn AI on for it. A session is a trader, a strategy and your exchange API key. You pick the model and, if you want, write instructions. They are sent with every decision, exactly as written.
- The trader you copy opens a position. CopyFi builds the question, calls your model, applies the session’s limits to its answer, and places what comes back.
- If the model is slow or unreachable, the session falls back to your strategy and keeps running on its own.
More careful, never more aggressive
This is the rule the whole feature hangs on. The strategy locked to the session is the only hard boundary. Your model can skip a trade, or take it smaller than your limit allows. It cannot raise the leverage, exceed the size, or trade against your direction filter, whatever reason it gives.
That design is deliberate, because a model can be wrong, and wrong convincingly. It may take a trade you would have refused, skip one you would have taken, and explain either choice in fluent, reasonable sentences. Limits written on a calm day are the only thing that does not argue back. Set them as though the model will eventually be wrong, because eventually it will be. The risk management guide covers how to size those limits.
Who decides, and who pays
- The decisions are your model’s, not CopyFi’s. CopyFi applies your limits and places the order. It does not evaluate or endorse what your model decides, and none of it is investment advice.
- You pay your provider for every decision, one call per copied order, at whatever your model costs. A fast cheap model and a slow capable one are both options, and the difference shows up on your provider’s bill.
- Your data goes to the provider you chose. Each question carries details of your session, strategy and broker account. Your exchange API keys are never sent.
Prove it on paper before it touches real money
Run the AI session on a testnet account with paper money first. Then do the comparison most AI trading products never offer: look at what your model decided against what the strategy alone would have done, on your own sessions, before you put real size behind it. If the model is not doing better than plain rules on paper, it has not earned your account.
One honest limit: backtesting on historical data is not shipped yet. Testing today means paper money against live prices, which is slower, and much harder to flatter.
The data it works from
A model is only as good as what it is shown. CopyFi tracks more than 1,800 traders with trading history data, and the length of that history varies a lot: some records span many months, others only a few. A few good months cannot separate a method from luck, for a model or for you. Read how to choose a trader to copy before letting any model decide on someone’s signal.
If you would rather have an assistant study your account than decide your orders, that is a separate, read-only feature: connect Claude or ChatGPT to your account.
Questions people ask
Does CopyFi use its own AI model?
No. You choose the model and supply the key: Anthropic, OpenAI, or any endpoint that speaks the OpenAI API. CopyFi builds the question for each copied order, applies your session’s limits to the answer, and places what comes back.
Can the AI raise my leverage or position size?
No. The strategy locked to the session still sets the leverage cap, the size limit and the direction filter. The model can be more careful than those limits, never more aggressive.
What happens if the model is slow or unreachable?
The session falls back to the strategy you selected and keeps running on those rules alone. Without a stored model key, a session simply runs its strategy.
Does AI copy trading guarantee better results?
No. A model can be wrong, and wrong confidently. Test on a paper account first, compare what the model decided against what the strategy alone would have done, and start small when you go live.
Trading involves risk, with or without a model. Past performance does not guarantee future results, and CopyFi makes no investment recommendations.