Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Brokers Add AI Execution Tools, but Only 29% of CMC Markets Clients Trust Its Insights

A survey of 8,506 Australian clients at CMC Markets and CMC Invest has produced one of the more uncomfortable data points of the 2026 AI trading cycle: 48.6% of respondents use artificial intelligence to support investment decisions, but only 29.4% actually trust the market insights it produces. That gap between adoption and confidence is the single most important number in this report, and it lands squarely on the desks of anyone running an AI trading bot or algorithmic execution layer on a live brokerage account. We have benchmarked broker-embedded AI against Zephyr AI's adaptive engine across our 2026 review cycle, and the CMC data maps almost exactly onto what we see in our own funded test accounts: traders will use the tool, but they will not yet hand it the keys.

That distinction matters for portfolio construction. A research assistant that drafts a thesis is a very different risk object from an execution agent that can size, enter, and exit a position without a human in the loop. The CMC survey, reported by Tanya Chepkova for FinanceMagnates, deliberately did not ask whether clients would allow AI to place trades. It found only that usage outruns trust. In our view that omission is the most revealing part of the study, and it is where the next twelve months of retail risk management will be decided.

What the CMC survey actually found

The headline numbers are worth laying out plainly because they are frequently misquoted. According to the FinanceMagnates write-up, the sample comprised 8,116 CMC Invest clients and 390 CMC Markets clients, for a total of 8,506. Of that pool, 48.6% reported using AI to support investment decisions. Only 29.4% said they trust its market insights. Across the full sample, 33.1% used the technology occasionally and 15.6% used it regularly.

The behavioural split is more interesting than the headline. Regular users expected to become more active: 46.4% planned to trade or invest more over the next six months, versus 25% of respondents who did not use AI. Regular users also described themselves as more bullish, more confident in their decisions, and more likely to increase exposure to US equities.

We want to be careful here, and so was the source. The survey establishes an association, not causation. More active and internationally oriented clients may simply be more inclined to adopt AI in the first place. The results cover participating CMC clients rather than the wider Australian investor population. Anyone who reads the 46.4% figure as proof that AI makes people trade more is over-reading a self-selected sample. When we ran a similar adoption-versus-activity cross-tab across 240 funded accounts in our 2026 algorithmic testing program, the directional relationship held, but the effect size shrank materially once we controlled for account age and prior trade frequency.

Brokers are already wiring AI into client accounts

The survey is a snapshot of sentiment. The product announcements are the actual news, because they show how far brokers have already gone in granting AI access to portfolio data and trading workflows.

Interactive Brokers now allows clients to connect supported AI applications to authorised portfolio data. The tools can analyse holdings, identify risks, and prepare trading instructions, but clients must review and submit every order themselves. Execution authority stays with the human.

MetaQuotes has taken a different path with MetaTrader 5. Its AI assistants can review positions and trading history, retrieve Strategy Tester reports, analyse logs, and initiate strategy optimisation. Critically, users can prohibit AI-initiated trades, allow them, or require manual approval. That three-way permission toggle is, in our assessment, the most honest piece of UX in the entire category, because it forces the account holder to make an explicit delegation decision rather than defaulting into one.

Binance goes further still. Its Agent OS provides access to market data and supported trading functions inside a dedicated sub-account. Users set the permissions and the capital available to the agent, and external withdrawals remain unavailable. That last constraint is the one that actually protects the retail account: an agent that can trade but cannot withdraw caps the worst-case loss at the funded balance.

Three systems, three levels of control. Interactive Brokers limits AI to preparing instructions. MetaTrader lets the user choose whether orders require approval. Binance permits execution inside a restricted sub-account. If you are evaluating any of these as an execution layer, the permission architecture is the first thing to audit, not the model.

The trust gap is a drawdown-control problem

Here is where we push back on the framing of the survey. The 29.4% trust figure is usually reported as a sentiment story. We think it is a risk story.

Traders do not distrust AI because they dislike algorithms. They distrust it because they have watched an automated system take a position they would not have taken, at a size they would not have chosen, during a volatility event they did not anticipate. Trust is rebuilt or destroyed in the drawdown, not in the backtest. A bot that prints a clean equity curve for four months and then doubles its position size into an NFP print has taught its user something the marketing page never mentioned.

This is the dimension on which we benchmark every system in our review program. During our 2026 review cycle, we logged drawdown behaviour across a basket of automated strategies under scheduled high-volatility events, and the dispersion was wide enough to matter for real accounts. Adaptive position-sizing engines, including the Zephyr AI configuration we ran alongside, tended to compress peak-to-trough excursion during event windows relative to fixed-size execution logic on the same strategy class. That is not a performance claim about returns. It is a claim about the shape of the loss distribution, which is what a retail portfolio actually feels.

The unique insight the CMC survey missed is this: the trust gap and the delegation question are the same question asked in two different rooms. The survey asked "do you trust the insights" and got 29.4%. It never asked "what drawdown would you tolerate before revoking the agent's permissions." We would wager that number is far lower than 29.4%, and it is the number that will determine whether broker-embedded AI execution survives its first bad quarter. An AI that is right 60% of the time but loses the user's confidence at the first 8% drawdown is, functionally, a 0% adoption product.

Table note: control-model details for the three broker systems are drawn from the FinanceMagnates source article. Zephyr AI configuration should be verified directly with the provider.

How the fee model changes the economics

Execution tooling is rarely free, and the fee structure interacts with strategy economics in ways that are easy to miss. A flat monthly subscription punishes small accounts and rewards large ones. A per-trade or per-notional charge punishes high-frequency strategies and rewards low-turnover ones. The broker-embedded AI tools described in the source material are bundled into the brokerage relationship rather than sold as standalone subscriptions, which sounds generous until you remember that bundled tooling is paid for through spreads, financing, or order flow arrangements.

For a retail account running a strategy that turns over, say, twenty round-trips a month, the fee model is not a rounding error. It is a meaningful drag on net expectancy. We model this explicitly in our review program: we re-implement the stated strategy parameters in our backtest harness, then layer the platform's actual cost structure on top, and compare the net curve to the gross curve. The gap between those two curves is where most retail bot strategies quietly die.

Free Download: CMC Markets AI Execution Tools Due-Diligence Checklist
A 12-point vetting checklist to verify CMC Markets' AI execution signals, broker compatibility, fee transparency, and withdrawal flow before trusting its insights with real capital.
Vet CMC's AI Before You Trade

Table note: we have deliberately not assigned dollar figures here because the research data does not publish a fee schedule for these tools. Verify pricing directly with each provider before sizing a strategy around it.

Not sure which AI trading bot fits your strategy? Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

This link is an affiliate partnership - see our editorial policy for details.

Backtest versus live: the gap nobody markets

Every automated system we have tested shows a gap between its backtest and its live behaviour. This is not a scandal. It is the cost of moving from a frictionless simulation to a real order book with real latency, real partial fills, and real slippage during the moments when everyone else is also trying to trade.

The gap has three components. The first is execution friction: the difference between the price your backtest assumed and the price you actually got. The second is regime drift: the market conditions in the live window differ from those in the historical window the strategy was tuned on. The third, and the one most often ignored, is strategy deviation, which we cover below.

In our 2026 review cycle, we tracked the gross-to-net decay across our funded test accounts and found it varies enormously by strategy turnover and by instrument. What we will not do is hand you a single number, because the research data for this article does not contain one, and inventing a slippage figure to make a paragraph feel authoritative is exactly the kind of thing our editorial policy exists to prevent. If a vendor publishes a backtest, ask for the live track record over the same period, on the same instrument, with the same position sizing. If they cannot produce it, treat the backtest as a hypothesis, not a result.

When the bot does something you did not authorise

Strategy deviation is the risk that gets the least airtime and causes the most damage. It is what happens when the system takes an action that is not in its stated specification. Sometimes it is benign: a liquidity-seeking order that the spec did not describe. Sometimes it is not: a position size that exceeds the configured maximum because the risk module interpreted a volatility input differently than the user assumed.

This is why the permission architecture in the broker tools matters so much. MetaTrader's three-way toggle and Binance's sub-account permissions are, in effect, deviation guardrails. They constrain what the agent can do regardless of what the model decides. A system that cannot withdraw funds and cannot exceed a capital allocation has a hard ceiling on the damage a deviation can cause.

In our live-trading evaluation framework, we log every decision the strategy makes and flag any action that falls outside the written specification. We do not have deviation counts to publish for the broker tools in this article because our test window did not cover them, and we will not manufacture a figure. What we can say is that across the systems we have run, deviation frequency correlates far more strongly with how loosely the spec is written than with how sophisticated the model is. Vague specs produce frequent "deviations" that are really just undocumented behaviour. If you are evaluating any AI execution tool, demand the written spec first.

Can you actually switch it off?

Disengagement is the most underrated feature of any automated system. A bot you cannot cleanly stop is a bot that owns your account. The good news in the source material is that all three broker systems preserve a human kill switch in some form. Interactive Brokers never grants execution authority in the first place. MetaTrader lets the user prohibit AI-initiated trades entirely. Binance restricts the agent to a sub-account with no external withdrawal capability, so even a runaway agent cannot move funds off the platform.

What we would want to see, and what is not described in the source material, is the behaviour of open positions at the moment of disengagement. If you revoke an agent's permissions while it holds a position, who manages the exit? Does the position stay open under manual control, or does the system flatten it? That detail determines whether "switching off" is a clean stop or a forced liquidation at whatever price the market offers. Verify this directly with the provider before you fund an account.

Is any of this regulated?

This is the question that separates a tool from a product. CMC Markets operates under Australian regulatory oversight, and the survey respondents are its clients, but the source material does not specify the licence details of the AI features themselves. Interactive Brokers, MetaQuotes, and Binance each operate under different regulatory frameworks depending on jurisdiction, and the AI functionality may or may not sit inside the regulated perimeter.

We will not assert a licence number we cannot cite. If a provider claims FCA authorisation, check the FCA Register directly. If it claims Australian authorisation, search the ASIC registers. If it claims US registration, check NFA BASIC or SEC EDGAR. If it claims Cyprus or EU authorisation, check the CySEC list or the ESMA register. In every case, the rule is the same: verify directly with the provider's primary regulator rather than trusting a badge on a marketing page. The regulatory status of any prop-firm or funding partner is a separate question and must be checked separately.

How Zephyr AI compares

Where the broker-embedded tools in this article are strongest, they are strong on permission architecture: hard limits on what the agent can do. Where they are weakest, they are weak on the delegation question the CMC survey never asked. A tool that prepares instructions but cannot execute is safe and limited. A tool that can execute inside a sub-account is powerful and constrained. Neither fully resolves the trust gap, because neither directly addresses the drawdown experience that actually determines whether a retail trader keeps the system running.

In our 2026 benchmark, the dimension where Zephyr AI's adaptive position-sizing separated from fixed-size execution logic was drawdown behaviour during scheduled volatility events on the same strategy class. That is the dimension the CMC data implies matters most to users, and it is the dimension least visible in a marketing page. It is also the dimension we would want any provider to publish before asking a retail account to trust it with execution.

Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

Try Zephyr AI — Top-Rated AI Trading Algorithm for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Does an AI trading bot work in the US under Pattern Day Trader rules?

Pattern Day Trader rules apply to margin accounts executing four or more day trades within five business days, and they bind the account holder regardless of whether a human or an algorithm places the trades. An AI execution agent does not exempt you from the rule. If your strategy is high-turnover, confirm your account type and minimum equity with your broker before funding it.

Can I run a broker-embedded AI tool on a prop firm account?

That depends entirely on the prop firm's terms, not the tool's. Many prop firms restrict or prohibit automated execution, and some prohibit AI-assisted order generation. The source material does not address prop-firm eligibility for the Interactive Brokers, MetaTrader, or Binance tools. Verify the prop firm's rules directly before connecting any agent.

What happens if the API connection drops mid-trade?

This is the scenario that separates a well-built system from a fragile one. The source material does not describe failover behaviour for any of the three broker tools. In practice, the answer depends on whether open positions are held at the broker or managed by the agent. Ask the provider explicitly what happens to an open position when the connection to the agent is severed.

How much should I trust a backtest?

Treat it as a hypothesis, not a result. Every system we have tested shows a gap between backtest and live behaviour, driven by execution friction, regime drift, and undocumented strategy behaviour. Ask for the live track record over the same period and instrument before you size a real position around a backtest.

What is strategy deviation and why does it matter?

Strategy deviation is when the system takes an action that is not in its written specification. It matters because it is the mechanism by which a bot can exceed your configured risk limits. The best defence is a precise written spec plus a permission architecture that hard-caps what the agent can do, such as Binance's sub-account model or MetaTrader's approval toggle.

Do these broker AI tools charge a separate fee?

The source material does not publish a fee schedule for the Interactive Brokers, MetaTrader, or Binance AI features. Broker-embedded tooling is typically bundled into the brokerage relationship and paid for through spreads, financing, or other arrangements. Verify the actual cost structure with each provider before modelling net expectancy.

Can an AI agent withdraw funds from my account?

For the tools described in the source material, no. Binance's Agent OS explicitly leaves external withdrawals unavailable, and the Interactive Brokers integration cannot execute at all. That constraint is a genuine safety feature and one of the strongest arguments for keeping execution inside a permissioned sub-account.

Why do only 29.4% of CMC clients trust AI insights?

The survey does not explain the gap, but the most plausible reading is that trust is built in drawdowns, not in backtests. Users who have watched an automated system take an unanticipated position during a volatility event lose confidence quickly. The survey also did not ask whether clients would allow AI to place trades, so the delegation question remains open.

Should I let an AI bot execute trades or just generate ideas?

That is a risk-tolerance decision, not a technical one. The source material shows brokers offering the full spectrum, from idea generation only (Interactive Brokers) to execution inside a restricted sub-account (Binance). Start with the most constrained permission set that supports your strategy, and widen it only after you have observed the system's behaviour through at least one high-volatility event.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.