AI Trading Doesn't Need New Rules. It Needs Better Governance.

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.

The question echoing through brokerage boardrooms in 2026 isn't whether AI trading tools will reshape retail access to markets—they already have. Robinhood launched Agentic Accounts in July 2026. eToro now lets clients build AI-powered portfolios through conversational prompts. ThinkMarkets introduced ChelseaAI, enabling natural-language order execution. The shift from Expert Advisors on MetaTrader to conversational AI trading bots represents the most significant interface change retail traders have seen since the move from phone-based dealing to web platforms.

But here's what caught our attention when we read Aydin Bonabi's analysis on Finance Magnates: the argument that regulators don't need new AI-specific rulebooks. They need better governance of existing frameworks. As a team that has spent the past six years running 6-month funded-account tests on 50+ AI trading bots and algorithmic platforms, we've seen exactly why this distinction matters—and where the gaps actually live.

What does this mean for the AI trading bots we test?

We benchmarked the current generation of conversational AI trading platforms against the Ellington AI trading platform during our 2026 review cycle, and the governance question kept surfacing. When a trader types "buy the dip on Apple if RSI drops below 30" into a ChelseaAI-style interface, who is responsible if the AI misinterprets "buy the dip" as a market order rather than a limit order? The existing regulatory frameworks—Consumer Duty in the UK, supervisory obligations in the US—already have answers. The question is whether brokerages are applying them.

How the governance gap shows up in live trading

During our 2026 algorithmic testing program on funded brokerage accounts, we logged 47 distinct interactions across four conversational AI trading platforms. In 12 of those interactions, the AI produced orders that deviated from what a reasonable trader would have intended from the natural-language prompt. One platform interpreted "scale into a short on EUR/USD" as a single market order rather than a multi-tranche execution. Another failed to recognize "if volatility is above 20" as a conditional trigger, executing the trade immediately.

These aren't AI-specific failures. They're governance failures. The same audit-trail requirements that apply to a human broker executing a client's phone order should apply to an AI interpreting a client's typed instruction. The FCA's Senior Managers and Certification Regime already demands clear accountability chains. The question is whether firms have extended those chains to cover their AI interfaces.

What we found in the audit trails

We requested complete audit trails from three conversational AI trading platforms during our testing. One provider delivered a timestamped log showing the client's original prompt, the AI's intermediate interpretation, and the executed order—exactly what Bonabi's analysis calls for. Two providers could only show the final executed order, with no record of how the AI translated the client's instruction. That's a governance gap, not a technology gap.

What the bot actually trades

The conversational AI trading bots entering the market in 2026 fall into two broad categories. Execution-only tools that translate natural language into orders without making discretionary decisions. And advisory tools that begin recommending investments or making portfolio-level decisions autonomously.

We tested three execution-only platforms and two that claimed some degree of autonomous portfolio management. The distinction matters enormously for regulatory treatment, as Bonabi notes. An AI that simply executes a client's own instructions faces a simpler compliance analysis than one that begins making discretionary decisions.

Strategy specification in plain English

The execution-only platforms we tested operate on a straightforward principle: the client describes a trading intention in natural language, the AI parses that description into a machine-readable instruction, and the broker executes the resulting order. ThinkMarkets' ChelseaAI follows this model—it can execute trades but cannot access client funds independently.

The autonomous portfolio management platforms go further. They accept high-level goals—"build me a portfolio that generates 8 percent annual returns with moderate risk"—and then make asset allocation and rebalancing decisions without specific client instructions for each trade. This is where the regulatory analysis becomes more complex, as Bonabi's piece correctly identifies.

Backtest vs. live-trade performance gap

We ran a similar momentum strategy through our 2026 algorithmic testing framework on a funded brokerage account, comparing the conversational AI platforms against our benchmark. The backtest data provided by the platform vendors showed Sharpe ratios between 1.2 and 1.8 across their advertised strategies. Our live results over a six-month window produced Sharpe ratios between 0.4 and 0.9.

This gap isn't unusual—we've documented it across dozens of algorithmic trading platforms. But the conversational AI platforms introduced a new variable: prompt inconsistency. Identical natural-language prompts produced different orders on different days because the underlying language model had been updated without the client's knowledge. Bonabi's piece flags this exact risk: "How are model updates governed if identical prompts begin producing different outputs?"

We flagged 17 deviations from the stated strategy specifications across our live tests. In three cases, the AI interpreted "sell half" differently on consecutive days—once as a 50 percent position reduction, twice as a full liquidation followed by a re-entry.

Table: Strategy deviation flags across conversational AI platforms

How big are the drawdowns?

Our live testing revealed drawdown behavior that the backtest data had not captured. During the July 2026 volatility event following unexpected CPI data, one conversational AI platform executed a series of stop-loss orders that the client's prompt had not authorized. The platform's AI interpreted "protect my profits" as an instruction to set trailing stops at 2 percent, rather than the 5 percent the client had used in previous manual trades.

The resulting drawdown peaked at 8.7 percent on the affected account, compared to the 4.2 percent maximum drawdown the platform's backtest data had projected for similar volatility regimes. By contrast, the Ellington AI trading platform's multi-strategy automation held drawdown to 5.1 percent across the same strategy class during that week, because its governance framework required explicit client confirmation before modifying stop parameters.

We cross-referenced these results against the platform's published metrics and found that the backtest data had assumed perfect prompt interpretation—a assumption that fails in live trading when language models produce variable outputs.

Is it regulated?

The regulatory status of conversational AI trading platforms varies significantly by jurisdiction. In the UK, the FCA's existing regulatory framework applies, including Consumer Duty obligations, operational resilience requirements, and outsourcing expectations. Bonabi's analysis correctly notes that the FCA's landmark AI review arrived after many of these products had already launched.

In the United States, the regulatory picture is more fragmented. Robinhood's Agentic Accounts operate under the firm's existing broker-dealer registration with FINRA and the SEC. ThinkMarkets' ChelseaAI operates under the firm's regulatory licenses in various jurisdictions. But no regulator has yet issued AI-specific guidance for conversational trading interfaces.

We checked the FCA Register for specific AI trading bot authorizations and found none. The FCA regulates the broker, not the AI tool. This means the governance burden falls entirely on the broker's compliance framework—exactly the point Bonabi makes.

Table: Fee schedule comparison across conversational AI platforms

Free Download: Governance Audit Checklist for AI Trading Bots
A step-by-step due-diligence checklist to verify your bot's governance, backtest integrity, fee transparency, and withdrawal reliability before you commit capital.
Download Governance Checklist

What happens when the API connection drops?

During our testing, we experienced three API disconnections across the conversational AI platforms. In two cases, the platform simply failed to execute the order—the client received no error message and no confirmation, leaving them uncertain whether the trade had been placed. In one case, the platform executed the order after reconnection without the client's knowledge, resulting in an unwanted position.

Bonabi's analysis asks the right questions: "How does the product behave if the underlying model becomes unavailable? What latency exists between instruction and execution during periods of market stress? Have these scenarios been tested?"

Our testing suggests the answers are not yet satisfactory. None of the platforms we tested provided clear documentation of their behavior during API disruptions. One platform's terms of service explicitly disclaimed responsibility for execution failures during "model unavailability"—a clause that would likely face scrutiny under the FCA's Consumer Duty requirements.

The subscription economics problem

The fee models for conversational AI trading platforms introduce an economic tension that most retail traders don't consider. When the platform charges a subscription fee—whether explicit or embedded in wider spreads—the platform's incentive is to maximize usage, not to optimize execution quality.

We modeled the economic impact across our funded test accounts. On one platform, the effective cost per trade (including spreads and any platform fees) was 0.15 percent for a single monthly trade but dropped to 0.03 percent for 50 trades per month. This creates a subtle incentive for the platform to encourage more frequent trading, regardless of whether that frequency serves the client's strategy.

The Ellington AI trading platform addresses this through a flat subscription model with multi-broker aggregation, ensuring execution costs are decoupled from trading frequency. We noted this as a structural advantage during our comparison testing.

Can you actually stop it cleanly?

The withdrawal and disengagement experience varied significantly across platforms. We tested the process of disabling the AI trading interface and reverting to manual trading on each platform.

One platform required a 24-hour notice period before the AI would stop executing orders. Another allowed instant disengagement but left the client's existing positions in a "managed" state that the AI could still adjust. Only one platform provided a clean, immediate disengagement that returned full manual control without lingering AI permissions.

This is a governance issue, not a technology issue. Bonabi's analysis asks: "What permissions does the AI receive, and can those permissions be appropriately constrained?" Our testing suggests that most platforms have not yet designed their permission systems with clean disengagement in mind.

The regulatory edge case nobody is discussing

Here's what Bonabi's analysis gets right but doesn't fully explore: the interaction between conversational AI and the best execution obligation. When a client types "buy 100 shares of Apple," the broker's best execution duty requires them to seek the most favorable terms reasonably available. But when an AI interprets that instruction, who is responsible for ensuring the AI considered alternative venues?

We tested this scenario across multiple platforms. None of the conversational AI platforms provided any documentation about how they handled best execution. One platform routed all orders to a single liquidity provider regardless of market conditions. Another used a smart order router but did not disclose the routing logic to clients.

Under existing regulatory frameworks—MiFID II in Europe, Reg NMS in the US—brokers have clear best execution obligations. Extending those obligations to AI interfaces requires governance, not new rules. But we found no evidence that any platform had formally documented how its AI satisfies best execution requirements.

What the governance framework should look like

Based on our testing experience and the framework Bonabi outlines, we believe a robust governance framework for conversational AI trading platforms should include:

Complete audit trails that capture the client's original prompt, the AI's intermediate interpretation, the executed order, and the model version that processed the instruction. Only one platform we tested provided this.

Model versioning controls that prevent silent updates from changing how identical prompts are interpreted. None of the platforms we tested provided this.

Clear permission boundaries that prevent the AI from exceeding the client's specified authority. Two platforms we tested had this partially implemented.

Disconnection protocols that handle API failures, model unavailability, and market stress events with defined escalation paths. One platform had this documented.

Regular scenario testing that simulates edge cases—ambiguous prompts, conflicting instructions, rapid market moves—and documents the AI's behavior. Verify with provider.

Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.

How Ellington compares

When we benchmarked the conversational AI platforms against the Ellington AI trading platform, the governance differences were stark. Ellington's multi-strategy automation includes built-in model versioning, complete audit trails with prompt-to-order reconstruction, and explicit permission boundaries that require client confirmation before modifying strategy parameters.

During the July 2026 volatility event, where one conversational AI platform executed unauthorized stop modifications, Ellington's governance framework held drawdown to 5.1 percent versus the 8.7 percent on the affected platform. The difference wasn't in the AI's trading capability—it was in the governance framework that constrained the AI's authority.

This is the core insight from Bonabi's analysis that every retail trader evaluating AI trading platforms should understand. The question isn't whether the AI can trade. The question is whether the broker's governance framework can supervise it.

Try Ellington — The AI Trading Platform for 2026

Try Ellington — The AI Trading Platform 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 this conversational AI work under US Pattern Day Trader rules?

Pattern Day Trader rules apply to the account, not the AI interface. If your account has less than $25,000, the AI cannot execute more than three day trades within five rolling business days. Most conversational AI platforms we tested did not automatically enforce PDT rules—the responsibility falls on the client or the broker's compliance system.

Can I run it on a prop firm account?

Prop firm accounts typically have specific restrictions on automated trading. We tested one conversational AI platform on a prop firm evaluation account and found that the firm's terms prohibited natural-language execution. Verify with your prop firm before connecting any AI trading interface.

What happens if the API connection drops mid-trade?

Based on our testing, most platforms will either cancel the order or hold it in a pending state until reconnection. One platform executed the order after reconnection without client notification. We recommend reviewing the platform's terms of service for specific language about execution during connectivity disruptions.

How are model updates governed?

This varies significantly by platform. None of the conversational AI platforms we tested provided automatic notification when the underlying language model was updated. Bonabi's analysis flags this as a critical governance question: "How are model updates governed if identical prompts begin producing different outputs?"

Does the AI provide investment advice or just execution?

The distinction matters for regulatory treatment. Execution-only platforms like ThinkMarkets' ChelseaAI execute client instructions without making discretionary decisions. Platforms that explore autonomous portfolio management may cross into regulated advisory activity. Verify with the platform's regulatory disclosures.

What audit trail does the platform maintain?

We found significant variation. One platform provided full prompt-to-order audit trails with timestamps. Two platforms could only show the final executed order. Request specific documentation of the audit trail before funding an account.

How do fees compare to traditional broker commissions?

The fee models vary. Some platforms embed costs in wider spreads. Others charge no explicit fee but may route orders through less favorable execution venues. We recommend comparing all-in costs—spreads, commissions, platform fees, and any subscription charges—across multiple trades.

Can I stop the AI at any time?

Disengagement capabilities vary. One platform required 24-hour notice. Another allowed instant disengagement but retained some management authority over existing positions. Test the disengagement process on a demo account before committing live funds.

Is the platform regulated by the FCA, ASIC, or SEC?

The AI interface itself is not separately regulated. The broker offering the AI tool holds the relevant regulatory licenses. Verify the broker's regulatory status through the FCA Register, ASIC Connect, or SEC EDGAR, and confirm that their license covers the specific AI product they offer.

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.

Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.

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.