Ask a language model where the next support level on EURUSD is and it will answer. It will give you a number to four decimal places, a short justification and a level of certainty that sounds like a desk analyst. The number is fiction. The model has no live price, no order book and no idea what happened this morning, so it produced something shaped like an answer because that is what it was asked for.
That single failure mode explains most of the disappointment traders report with AI tools, and most of the damage. The tools are genuinely useful. They are useful at a different set of tasks than the ones people try first.
Where they actually earn their place
The strongest use is reading things you already have. Export a year of trades and an assistant will find patterns you will not, because you stopped counting after the fortieth row. Which sessions produce your losses. Whether your average holding time collapses after a losing streak. How often you moved a stop and what happened afterwards. This is the analysis a trading journal is supposed to enable and almost nobody performs by hand.
The second strong use is language work on documents. Central bank statements, earnings releases, regulatory notices and platform documentation are long and repetitive by design. A model that summarises a policy statement and lists what changed against the previous one saves twenty minutes and rarely gets the wording wrong, because the source text is sitting in front of it.
The third is code. Writing an indicator, fixing a Pine Script error, translating an idea into a testable rule, converting a spreadsheet formula into something a platform can run. Code either compiles or it does not, and the output can be checked against a chart in seconds. That verification loop is what makes the task safe.
Where they break
They break on anything requiring current, precise, external facts. Prices, spreads, swap rates, this week's economic calendar, whether a broker still holds a given licence. A model without a live connection is working from a snapshot of the past, and a model with a live connection is only as good as the source it fetched. Either way, a number that matters should come from the price feed or the economic calendar, never from a chat window.
They also break on the thing that feels most impressive: prediction. Directional calls from a language model are not analysis, they are the most plausible sounding continuation of the prompt you wrote. If you ask leadingly, you will get agreement. That is not a defect you can prompt your way out of, and it becomes dangerous when the answer arrives in the confident register of a research note.
An AI assistant is not a signal service and it is not advice. Anything it produces about a market is information to check, not a reason to open a position. Trading carries a high risk of loss whoever or whatever generated the idea.
The guardrails worth setting before you start
Four rules cover almost every problem I have seen.
- No execution. The assistant never has credentials that can place, modify or close an order. Read only access to trade history is fine. Write access to an account is not, no matter how tidy the plan sounded.
- Every number that will size or price a trade comes from the platform, not the model. Positions get sized by a position size calculator against live account equity.
- Any rule the model proposes goes through the same backtesting and forward testing you would apply to a stranger's idea, because that is exactly what it is.
- Nothing sensitive goes into a third party chat: account numbers, broker credentials, client data, identity documents. If a firm is handling client information, that is a data protection matter with real consequences.
Assistant, or expert advisor
People conflate two different things. An expert advisor is deterministic. It runs the same rules on every candle and you can test what it did over ten years of data. A language model is probabilistic and will answer the same question differently on Tuesday. For anything that has to behave identically every time, the deterministic tool wins, and it is not close. The comparison in algorithmic indicators versus expert advisors holds here too: automation you can audit beats automation that sounds clever.
Where a model adds something an expert advisor cannot is judgement about messy text and messy history. Reading a rulebook. Explaining why a prop firm flagged an account. Turning a vague observation into a precise, testable rule you can then hand to the deterministic system.
Firms deploying assistants to clients
Different problem entirely. A broker or prop firm putting an assistant in front of clients takes on responsibility for what it says. Two controls are not optional. The first is a hard boundary on advice: the assistant answers questions about the platform, the rules and the account, and declines anything that resembles a recommendation to trade. The second is logging, because if a client claims the assistant told them to do something, the transcript is the only evidence that exists.
Cost control matters as well. Firms that run assistants at volume usually want their own model keys rather than a bundled allowance, which is the bring your own key pattern, so spend and data handling both stay under the firm's control.
A working setup
Mine is unremarkable. Trade history and journal notes go in weekly for pattern review. Long policy documents go in for summaries I then read against the original. Indicator and script work happens with a model open beside the editor. Market direction, entries, stops and sizing come from the chart, the rules in my trading plan and the account balance in front of me.
The tool is good at the parts of trading that are reading and writing. The parts that are risk stay manual, and they should.
"I use one every day and I have never let one touch an order ticket. The moment a model can place trades, its confidence becomes your position size, and models are confident about everything."
— Alex Onta, Executive Director, SINGUARD
Key Takeaways
- Language models are strong at reviewing your own trade history, summarising long documents and writing indicator code, because all three can be verified.
- They invent prices, levels and current facts on demand. Any number that sizes or prices a trade comes from the platform, never from a chat window.
- Set four rules up front: no execution access, no sensitive data, platform sourced numbers, and full testing of any rule the model suggests.
- For deterministic, repeatable behaviour an expert advisor beats a language model. Use the model to design the rule, not to run it.
Frequently Asked Questions
Can an AI assistant predict market direction?
No. A language model produces the most plausible continuation of your prompt, not a forecast, and it will agree with a leading question. Treat any directional output as an idea to test, never as a reason to trade. Trading carries a high risk of loss.
Is it safe to connect an AI assistant to a trading account?
Read only access to trade history is reasonable. Order placement access is not, because a confident but wrong output turns straight into a live position with no human check in between.
What should never be pasted into a public AI chat?
Account numbers, platform credentials, API keys, client identity documents and anything covered by data protection obligations. Firms handling client information should run models under their own keys with logging and retention they control.
About the Author
Alex Onta is an Executive Director at SINGUARD. He built eTrader, the terminal, the mobile apps, eTrader Broker, Copytrading, Business and Community, along with the worldwide clustered-server infrastructure it all runs on, with his brother Roman Onta helping on the design, and he leads that division today. Together with Roman he builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals, and the two of them carry worldwide compliance, payment processing and international business structuring side by side. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.