AI Agents Trading Accounts: Who Approves The Trade? Seven Brokers, Seven Answers, And The First Supervisory Frameworks
B2Broker has given white-label brokers MCP access for AI agents with three access levels, Finance Magnates counts at least six firms letting ChatGPT and Claude work in client accounts under different approval rules, and on 4 November the US OCC published its first supervisory framework for agentic AI. Who approves an AI agent's trade now depends on the broker, and that choice is becoming a regulatory question.
AlchmAI Editorial7 min read
6+
Brokers and trading technology firms letting AI assistants work in client accounts since June (Finance Magnates, 4 November 2026)
3
Access levels B2Broker gives brokers for AI agents: off, read-only or full (B2Broker, 4 November 2026)
5.68m
Interactive Brokers client accounts in October, up 34% year on year, as it widens autonomous AI trading (Finance Magnates, 2 November 2026)
38
Pages in the DFSA and Cambridge paper on frontier and agentic AI risks in finance (Finance Magnates, 4 November 2026)
AI agents trading accounts no longer follow one rule. At some brokers the agent can only read; at others it prepares an order the client must confirm; at a few it trades on its own within limits. On 4 November 2026 B2Broker gave white-label brokers Model Context Protocol access with three settings, Finance Magnates counted at least six firms letting ChatGPT and Claude into client accounts, and the US Office of the Comptroller of the Currency published its first supervisory framework for agentic AI.
The common question, put plainly by B2Broker's founder Arthur Azizov, is this: adding AI to a terminal "raises an immediate question: who decides what an assistant can actually do?" Our first look at this market, on 27 October, covered Interactive Brokers removing per-trade approval. A week later the more useful story is the spread of answers, and the fact that regulators have started to describe what good control looks like.
How Do Brokers Let AI Agents Trade Client Accounts Now?
Through MCP connections with very different approval rules. Finance Magnates' 4 November round-up, alongside B2Broker's release the same day and Interactive Brokers' own release notes, sets out the range. The table below summarises what each firm says its agents can do and who approves the trade.
| Firm | What the agent can do | Who approves | Limits stated |
|---|---|---|---|
| PrimeXBT | Prepares market, limit and stop orders from ChatGPT or Claude | Client confirms each order after seeing price, fees, margin and liquidation level | Read-only or Read & Trade; cannot withdraw or move funds |
| B2Broker B2TRADER | Analyses positions and, with permission, places, modifies and closes trades | Broker sets the level; full-access agents trade without confirming each order | Each agent sees only the authorising trader's accounts |
| Leverate MCP for Traders | Restates size, price, stops and margin impact | Client says yes before execution | Not stated |
| Capital.com (MENA) | Previews each trade | Two-step confirmation | Not stated |
| Axi | Claude acts on the account | Client can set an action to always allow, after which Claude stops asking | Not stated |
| ThinkMarkets | AI can execute orders since June | Not stated | API cannot move money |
| Interactive Brokers | Autonomous Trading for eligible clients, plus AI-built screeners | No confirmation for each order for eligible clients | Restrictions and availability conditions apply |
Two details stand out. First, the line between preparing and placing an order is where most firms differ: B2Broker's full-access setting and Interactive Brokers' Autonomous Trading have moved past it, and Axi's always-allow option comes close. Second, where firms state a hard limit, it is about money: PrimeXBT and ThinkMarkets both say their agents cannot move funds out of the account. PrimeXBT's Jonatan Randin described the aim as bringing "real trading data and functionality into the AI assistants people already know and use".
The controls around the agent also differ. PrimeXBT's integration covers CFDs and crypto futures on demo and live PXTrader 2.0 accounts, with a separate trading API behind two-factor authentication and optional IP restrictions. Interactive Brokers' Connector release notes add AI-built stock screeners, described in plain language and covering US, European and Hong Kong stocks, plus earnings insights inside the connected platform. In other words, the agent is becoming a full interface to the account, not a chat window beside it.
Why Does The Approval Model Matter For AI Agents Trading Accounts?
Because it decides who is accountable for a wrong trade. B2Broker's answer is to hand the decision to each broker: brokers can disable MCP, allow read-only access for analysis, or grant full access so authorised agents can place, modify or close trades. Its own in-terminal AI Chat still requires explicit confirmation for any trade it proposes. Azizov framed the trade-off directly: "A trader might need analysis in one situation and automated execution in another. The broker has to define those boundaries."
Scale makes the choice harder to defer. Interactive Brokers reported 5.68 million client accounts in October, up 34% on a year earlier, and 4.31 million daily average revenue trades, according to Finance Magnates. Its Autonomous Trading lets eligible clients authorise connected AI platforms to submit supported orders "without approving each order individually", across more than 75,000 stocks, options and ETFs. B2TRADER, which B2Broker sells to other brokers, can process up to 3,000 requests per second. Decisions made in a product meeting now apply to a very large number of accounts.
What Did The OCC And The DFSA Say About Agentic AI?
Both described the risk without writing new rules. The OCC's Bulletin 2026-54, issued on 4 November and titled Super Intelligence: Supervisory Framework, is a resource for bank examiners. It says examiners may classify AI use by risk that "may be based on potential consequences, transfer of agency, and the degree of human oversight", rather than by how sophisticated the technology is, and that control intensity should rise with the consequences of an agent's failure or misuse. It is explicit about its status: "This supervisory framework does not establish supervisory expectations for banks."
In Dubai, the Dubai Financial Services Authority published a 38-page research paper on 30 October with the Cambridge Digital Innovation and Regulation Initiative, Frontier and Agentic AI: Cyber Resilience in the Financial Sector. Finance Magnates reports that it flags operational errors and control failures when AI places trades without human approval, and significant liability for brokers that connect automated systems to client positions. The paper describes agentic systems that can use connected tools to act, sometimes without approval at each step, including accessing customer data, changing system settings or submitting trading instructions. It is not regulation either, but it names the same three questions the OCC does: who controls the agent, who is accountable when it errs, and how a firm responds.
Our View: Brokers Need An Approval Policy, Not An Approval Button
The table shows a market where each firm has made a product decision about approvals. What it does not show is the policy behind it: which clients may grant full access, which instruments an agent may trade, what happens when an agent hits a loss limit at 3am, and how a complaint is investigated when the client says the agent misread the instruction. The OCC's wording, consequences, transfer of agency and human oversight, is a good checklist for that policy even for firms it does not supervise.
We expect the firms that do well to make approval a setting with a record behind it: who enabled autonomy, when, for which account, with which limits, and every agent order tagged as agent-originated. That record is what a compliance lead will need the first time an agent's trade is disputed.
“Every broker now has an approval model for AI agents. Few yet have an approval policy, and that is what a regulator or a disputed trade will ask to see.”
What It Means For Trading And Fintech Firms
- Brokers and CFD providers: decide which of the four approval models you offer, to which clients, and write it down. Tag every agent-originated order and keep the permission record next to it.
- Prop trading firms: if traders connect agents to evaluation or funded accounts, your rule engine should know which orders came from an agent and apply the same limits, with the breach evidence to match.
- Compliance leads: map each approval model against client understanding. A client who switches on always allow should see, in plain words, what the agent may now do without asking.
- Fintech and platform vendors: give brokers the controls B2Broker describes, off, read-only and full, plus per-account limits and an audit log. Brokers will compare vendors on these settings.
- Research and quant teams: once agent orders are tagged, measure them. Clustering of agent orders around the same signals is the risk both regulators described.
The Bottom Line
AI agents trading accounts now sit on a spectrum from read-only to fully autonomous, and the choice differs at almost every broker: PrimeXBT confirms each order, Axi lets clients stop being asked, B2Broker hands the decision to the broker and Interactive Brokers lets eligible clients skip approval. The OCC and the DFSA have both described the same test: match the controls to the consequences. AlchmAI, a London trading technology and AI studio, builds agent access, approval settings and audit logs for brokers and prop firms, and this week made clear that the setting is the easy part; the record behind it is what matters.
References & Further Reading
- Finance Magnates - ChatGPT and Claude Reach More Trading Accounts, With Different Rules on Approval (4 November 2026). financemagnates.com/forex/brokers/primexbt-lets-chatgpt-and-claude-prepare-trades-that-clients-must-approve
- B2BROKER via GlobeNewswire - B2BROKER Expands B2TRADER AI Strategy With MCP Trading Agents and In-Terminal Chat (4 November 2026). globenewswire.com/news-release/2026/11/04/3395078/0/en/b2broker-expands-b2trader-ai-strategy-with-mcp-trading-agents-and-in-terminal-chat.html
- Office of the Comptroller of the Currency - Super Intelligence: Supervisory Framework, Bulletin 2026-54 (4 November 2026). occ.gov/news-issuances/bulletins/2026/bulletin-2026-54.html
- Finance Magnates - Dubai Regulator Raises AI Trading Risks as Brokers Expand Automated Execution (4 November 2026). financemagnates.com/forex/technology/dubai-regulator-raises-ai-trading-risks-as-brokers-expand-automated-execution
- Finance Magnates - Interactive Brokers Reports 4.3 Million DARTs in October (2 November 2026). financemagnates.com/forex/interactive-brokers-reports-43-million-darts-in-october-as-monthly-activity-rises-5
- AlchmAI Blog - Interactive Brokers drops per-trade approval for AI agents. alchmai.com/blog/ai-agents-trading-accounts-interactive-brokers-autonomous-trading-without-approval
- AlchmAI - Controls for AI features. alchmai.com/solutions/ai-controls
- AlchmAI - Broker trading platforms. alchmai.com/solutions/broker-trading-platforms
AlchmAI Editorial
Research and analysis, London
The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.
This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information
