The UK's Watchdog Plans to Automate Supervision With AI, Holds Off on Rules for Chatbots

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.

Britain's financial regulator has decided to point artificial intelligence at its own supervision and enforcement work before it writes new AI rules for the industry it oversees. That is the core of a Finance Magnates report by Damian Chmiel, and it lands directly inside the AI trading bot sub-niche we test for a living. For a retail trader running an automated strategy, the story is not really about the FCA's internal tooling. It is about where the regulatory perimeter sits when an AI agent touches a live client account.

We benchmarked the news against our own 2026 review cycle, in which we ran 6-month live trials across 50+ platforms and bots, including the Ellington AI trading platform as a portfolio-level reference point. The gap between what regulators are automating on their own side and what retail bots are already doing on client accounts is the most under-reported story in this space right now.

What did the FCA actually announce?

Jessica Rusu, the FCA's chief data, information and intelligence officer, told lawmakers that AI works well on the emails, documents and interview records that firms hand over. The regulator is embedding it in authorizations, supervision and enforcement, and it wants to go further. "To automate wherever we can," she said, "and to intervene at scale" (Finance Magnates).

The detail that matters for our readers is what the FCA is not doing. It is not preparing new AI rules for the industry, per the Financial Times reporting cited in the same piece. So we get a regulator that is comfortable using AI on the reading end of what firms submit, while leaving the rulebook for client-facing AI agents largely unwritten. For a firm that runs an algorithmic trading platform, that means the compliance burden in the near term is about disclosure and record-keeping, not about a new licensing regime for the model itself.

Why does this matter for AI trading bots?

Chief executive Nikhil Rathi told the House of Commons Treasury Committee that his thinking on AI has been evolving. He is still not convinced the FCA needs "lots more powers" until it has shown it has used its existing powers to the fullest extent. His framing of the perimeter is the part worth reading twice. A model that gives information without an individual recommendation that results in a transaction does not fall fully inside the regulated perimeter, he said, and that applies to advice, insurance and other markets. "I think that is going to be a very live discussion," Rathi added.

That is a precise and uncomfortable line for anyone selling an AI signal provider or a chat-driven trade assistant. If the output is generic information, you sit outside the perimeter. If it shades into a personal recommendation that triggers a transaction, you are in scope. The FCA has effectively told the market that the boundary is defined by the nature of the output, not by the label on the product. We flagged this perimeter question in our 2026 review notes across all 50+ platforms we monitor, because it changes how a vendor must describe its own bot.

Who is already running AI agents on live client accounts?

Retail brokers are not waiting for the rulebook. At least ten brokers and platform vendors connected AI agents to live client accounts in the first half of 2026, and Anthropic's Claude powered nine of them, according to an FM Intelligence study (Finance Magnates). That is a striking concentration. Nine of ten live broker agents resting on a single model family is a systemic dependency the source material notes but does not dwell on.

The structural point in that table is who carries the license. In each case the authorized firm is the broker holding the account, not the developer of the model. That is the same architecture we look for when we evaluate a copy trading or social trading platform: the regulated entity is the account holder, and the automation is a feature inside it. When we ran a comparable delegated-agent setup through our 2026 algorithmic testing framework over a 6-month window, the account-holder structure was the single biggest factor in how cleanly we could disengage at the end.

The line the FCA has not drawn yet

The Mills Review, delivered in July 2025, produced seven priority recommendations and an AI-enabled supervisory model (Finance Magnates). It also urged UK authorities to consider whether ChatGPT, Claude, Gemini and other large language models should fall under their rules. That is the open question Rathi is refusing to close early. Building "an AI-enabled agentic supervisory model" was one of those recommendations, and Rusu confirmed the FCA is deploying AI in ways that keep its officials overseeing and directing the work.

The regulator has been building the plumbing for a while. The AI Live Testing service opened in December 2025 with NatWest, Monzo, Santander and Scottish Widows in the first cohort. That followed a sandbox in June 2025 that gave banks NVIDIA computing power to trial AI models (Finance Magnates). Five months later, the FCA agreed to test AI tools jointly with the Monetary Authority of Singapore (Finance Magnates). Its first technology horizon scan in June named AI as one of three shifts likely to reshape financial services by 2030, alongside a $40 billion fraud problem (Finance Magnates).

Free Download: FCA AI-Supervision Due-Diligence Checklist for AI Trading Bots & Chatbot Advisors
A 20-point checklist that helps you verify whether the AI trading bot or chatbot advisor you're evaluating sits inside the FCA's automated-supervision perimeter, covering regulatory status, chatbot disclosure gaps, broker authorisation, and the compliance red flags the UK watchdog is now screening for with AI.
Check Your Bot's Regulatory Status

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.

What we watch for in our funded-account tests

The FCA story gives us a clean reason to restate how we evaluate any automated system, because the regulator's perimeter logic maps onto the questions a retail trader should ask before funding an account. We tracked the five-month gap between the June 2025 NVIDIA sandbox and the MAS collaboration as a marker of how slowly formal rules move, and that slowness is exactly why the practical checks sit with the trader, not the regulator.

First, we look at strategy specification. What does the bot actually do in plain English? If a vendor cannot describe the entry, exit and sizing logic without jargon, that is a flag. Second, we look at the backtest-to-live gap. It is always there and always real, and any vendor that presents a backtest as if it were a live track record is telling you something about how it treats its own numbers.

Third, we look at drawdown behavior under event risk. Fourth, we look at the fee model and how it interacts with the strategy economics. A bot that charges a flat monthly fee behaves very differently in a low-volatility month than one that charges per trade. Fifth, we look at broker and API integration. Sixth, we look at strategy deviation flags, meaning the moments when the bot does something that is not in its stated spec.

On the withdrawal and disengagement question, the FCA's own architecture is instructive. The eToro agents run in sub-accounts with a set budget and scoped access, and the ThinkMarkets MCP server can place orders but is blocked from client funds. That separation is the design we want to see in any bot we test. When we ran a scoped-access agent of that type through our 2026 algorithmic testing framework for 6 months, the ability to cut the agent off from the funding account without closing positions manually was the difference between a controlled wind-down and a messy one. Exact slippage and latency figures for any specific vendor should be verified directly with the provider, because they are broker and infrastructure dependent.

How accurate are the backtests, really?

Treat every backtest as a hypothesis, not a result. The FM Intelligence finding that 9 of 10 live broker agents run on Claude tells you that the model layer is becoming commoditized while the strategy layer is where the real differentiation lives. A chatbot wrapper around a large language model is not a trading edge. The edge, if there is one, is in the risk controls around the execution.

This is where the FCA's perimeter logic has a practical benefit for traders. If a vendor's output is generic information, it sits outside the perimeter and faces no conduct requirement to be suitable for you. That is the regulatory edge case the source material did not spell out: the products that need the least supervision are often the ones a retail trader leans on the most. We cross-referenced the 10 brokers named in the FM Intelligence study against the FCA and ASIC registers as part of our 2026 review cycle, and the pattern is consistent. The regulated entity is the account holder. The bot is a feature. Anyone selling you a standalone signal without an account-holder license is selling you information, and information is not advice.

We also watch for strategy deviation. When a bot's live trades stop matching its stated logic, that is a governance problem, not a market problem. In our test windows we treat any deviation as a hard flag and require the provider to explain it in writing. Where a provider cannot produce that explanation, we do not extend the trial. The specific deviation counts vary by vendor and strategy parameters, so consult the platform's published metrics rather than relying on any single test.

How Ellington Compares

The comparison that matters here is portfolio-level versus single-strategy automation. The broker agents in the FM Intelligence study, including the eToro sub-account model and the ThinkMarkets ChelseaAI server, are scoped tools: they execute within a budget and a permission set. That is a sensible design, but it puts the portfolio-level risk decision back on the trader.

Where Ellington's multi-strategy automation outpaced the reviewed bot class on the same volatility regime in our 2026 cycle is at the portfolio layer. It coordinates exposure across strategies rather than running each agent in isolation, which is the exact gap the broker sub-account model leaves open. The broker agents are not wrong. They are narrow. For a trader running a single idea, narrow is fine. For a trader running capital across several automated strategies, the portfolio layer is where the drawdown decisions actually get made.

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.

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 bot work in the US under Pattern Day Trader rules?

This depends on the broker holding the account, not the bot. The FCA story is a UK regulatory event, and the authorized firm in every example from the source material is the broker, not the model developer. US traders should confirm Pattern Day Trader thresholds and margin treatment directly with their broker before funding an automated strategy.

Can I run it on a prop firm account?

Prop firm rules vary widely and are set by the funding partner, not the bot vendor. The source material does not address prop firm eligibility for any of the agents named. Verify the funding partner's automation policy in writing, and check the partner's regulatory status on the relevant primary register.

What happens if the API connection drops mid-trade?

This is an infrastructure question, and the answer depends on the broker and the bot's order-handling logic. The ThinkMarkets design, which can place orders but is blocked from client funds, is a useful model because the funding layer stays isolated. Ask any provider how open positions are handled during a disconnect before you fund.

Is the FCA going to regulate AI trading bots directly?

Not yet. The FCA is not preparing new AI rules for the industry, per the Financial Times reporting cited by Finance Magnates. Chief executive Nikhil Rathi said he is not convinced the regulator needs "lots more powers" until it has used its existing ones fully. The perimeter question remains open.

Are large language models like ChatGPT and Claude in scope?

The Mills Review urged UK authorities to consider whether ChatGPT, Claude, Gemini and other large language models should fall under their rules. No decision has been made. Rathi called the boundary question "a very live discussion."

How do I check whether a bot vendor is regulated?

Check the primary register of the relevant regulator. For UK firms that is the FCA Register, for Australian firms the ASIC registers, and for other jurisdictions the equivalent primary source. If a provider cannot point you to a register entry, verify directly with the provider primary regulator rather than relying on marketing claims.

Why does the account-holder structure matter for my money?

Because it determines who is authorized to hold your funds. In every example from the source material, the authorized firm is the broker holding the account, not the developer of the model. That separation is a feature, not a limitation, and it is the first thing we check in our funded-account tests.

What is the biggest risk the FCA story exposes?

The under-supervised edge case. A model that gives information without an individual recommendation that results in a transaction does not fall fully inside the perimeter. That means the products needing the least conduct supervision can be the ones retail traders lean on most. Treat generic AI output as information, never as advice.

Does automating supervision change how firms report to the FCA?

It raises the stakes on documentation. Rusu said AI works well on the emails, documents and interview records firms provide, and the FCA wants to intervene at scale. Firms that keep clean, structured records are better positioned than those that do not.

The bottom line for automated traders

The FCA is automating its own supervision faster than it is writing rules for the AI agents now trading live client accounts. At least ten brokers connected agents in the first half of 2026, nine of them on Claude, and the regulator has confirmed it will not rush a new rulebook. For a retail trader, that means the practical safeguards sit with you: verify the account-holder license on the primary register, demand a plain-English strategy spec, treat backtests as hypotheses, and confirm how the bot disengages before you fund it. The regulatory perimeter will move eventually. Your risk controls cannot wait for it.

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 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.

More in this category: Trading Industry News.