US Bank Regulator Issues First AI Framework as Banks Battle Rogue Agents
US Bank Regulator Issues First AI Framework as Banks Battle Rogue Agents
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 Office of the Comptroller of the Currency has published its first examiner guide for artificial intelligence, and the framing tells you more than the length does. According to Decrypt, the guide weighs autonomy and financial harm, and it addresses outside AI agents directly, which is the same category of risk that recently pushed major banks to restrict customer-account access for third-party agents.
That is not a banking story that stays inside banking. Our testing program has run six-month funded-account trials across more than 50 trading platforms and AI trading bots since 2020, and the OCC's two-axis test, how much autonomy does the system have, and how much financial harm can it cause before a human intervenes, is the same test we apply to every AI trading bot and algorithmic trading platform we evaluate. The difference is that a bank has an examiner. A retail trading bot has a login screen and a subscription button. During our 2026 review cycle we used the Ellington AI trading platform as a portfolio-level benchmark precisely because it enforces risk limits at the account level rather than the strategy level, which is the distinction the OCC guide is quietly drawing.
What does the OCC framework actually say?
Start with what the document is. It is an examiner guide, not a consumer rulebook. It gives OCC supervisory staff a lens for looking at AI deployments inside national banks, and that lens has two settings: autonomy and financial harm. Autonomy asks how much the system can do without a human in the loop. Financial harm asks how bad the damage gets before anyone notices.
Both settings are calibrated toward a firm with compliance staff, model risk committees and an audit trail. Retail traders have none of those things, which is exactly why the guide is worth reading anyway. The vocabulary transfers cleanly. If you are running a bot that routes live orders through a broker API, you have an autonomous agent with financial permissions attached to a real account. The scale is smaller. The structure is identical.
The second half of the Decrypt report is the part that should get more attention than it has. The guide addresses outside AI agents, meaning agents that are not built by the bank but are granted access to bank infrastructure or customer accounts. Major banks responded to that risk class by restricting customer-account access. Read that sequence again: a regulator named the risk, and the regulated firms cut off agent access. That is a preview of how broker-side API permissions could tighten for retail trading bots if a high-profile failure lands on a regulator's desk.
Why retail AI trading bots are in scope
Not every product that calls itself an AI trading bot is the same animal, and the OCC framing helps separate them.
A research agent reads data and produces commentary. A signal provider produces entries and exits that a human executes. An order-routing bot holds API credentials with trade permission and fires orders without asking. Copy trading and social trading platforms sit in a fourth bucket, because they route your capital behind another account's decisions.
The OCC's outside-agent concern maps almost entirely onto the third and fourth buckets. Those are the products where an agent can move real money inside an account it does not own, and where the harm accumulates before a human sees it.
| OCC focus area (per the Decrypt report) | What it means inside a bank | What it maps to in a retail bot account |
|---|---|---|
| Autonomy of the AI | How much the system can do before a human signs off | Whether the bot routes orders without confirmation, and whether a kill switch exists |
| Financial harm | Size of loss the agent can cause before detection | Position sizing, leverage, and whether the bot can touch the whole account or a sub-account only |
| Outside AI agents | Third-party agents reaching into customer accounts | Broker API keys, OAuth scopes, and copy-trading bridges |
| Customer-account access restrictions | Banks limiting what outside agents may do | Broker-level API throttling, credential revocation, and read-only modes |
None of those retail cells are filled in by the OCC guide. It is a supervisory document aimed at banks, and we should be honest about that limitation rather than pretending it hands retail traders a checklist. What it hands us is a way of ranking risk.
In our 2026 review framework we score every bot on two questions before we look at returns at all: can this thing trade without a human, and how much of the account can it reach? A signal provider with no execution permission scores near zero on both. A closed-source bot with full withdrawal-adjacent API scope scores high on both, and that is the profile the OCC is describing when it talks about financial harm.
What changes for someone running an AI trading bot?
Honestly, very little on day one, and potentially a lot on the margin.
Nothing in the guide creates a new obligation for a retail trader running an expert advisor on a personal account. What it does is signal where supervisory attention is heading, and supervisory attention tends to move from banks to their vendors, and from vendors to the API layers that connect everything together.
The practical change is that broker-side permissions become a live variable rather than a fixed assumption. If a large broker decides that unvetted third-party agents are a risk, the mechanism to restrict them is already built: scoped API keys, rate limits, and revocation. A bot that depends on a single broker connection has a single point of failure that is not under its own control.
Where a bare-bones backtesting framework like Backtrader leaves credential handling and failure logic to whoever wires it up, an integrated platform can enforce account-level exposure caps regardless of what the strategy code wants to do. That is a structural difference, not a performance claim, and it is the axis we weight most heavily when an agent has live order permission.
Is your AI trading bot regulated by anyone?
Usually not in the way you think, and this is the part of the story that the banking coverage tends to skip.
Most AI trading bot vendors describe themselves as software providers. They sell a tool, not personalized investment advice, which is the distinction that determines whether a regulator has jurisdiction over them at all. Some fall inside the perimeter. Many do not. That means for a large share of the market there is no register entry to look up, because no licence was ever required.
That is the under-discussed edge case here. The OCC can examine a bank's AI deployment because the bank is chartered and supervised. A retail bot vendor with no licence and no client-money handling may sit outside every register you could reasonably search, and the only regulated counterparty left in the trade is your broker. So broker due diligence becomes the load-bearing wall, not a nice-to-have.
Where a vendor does claim a licence, check the claim against the primary register rather than a marketing page. The table below lists the registers we use.
| Market | Primary register | What to check | Notes |
|---|---|---|---|
| United Kingdom | FCA Register | Firm reference number, permitted activities | Verify directly against the register entry |
| Australia | ASIC registers | AFSL number, authorisations | Confirm the licence covers the described activity |
| Cyprus and EU | CySEC list of investment firms | CIF licence, passporting | Confirm entity name matches the contracting entity |
| United States, securities | SEC EDGAR | Filings, registration status | Absence of filings is itself information |
| United States, futures and FX | NFA BASIC | Membership, disciplinary history | Check for prior actions, not just current status |
| European Union | ESMA registers | Passporting status | Cross-check the home-state regulator |
| Singapore | MAS Financial Institutions Directory | Licence type | Verify the specific entity, not the brand |
Free Download: AI Trading Bot Due-Diligence Checklist: Rogue-Agent Controls, Regulatory Status & Broker Permission Limits
A step-by-step vetting checklist that helps you confirm whether a bot's agent permissions, kill-switch behavior, regulatory disclosures and broker access align with the new US bank regulator AI framework before you commit capital.
Download the Oversight Checklist
A brand name on a website is not a licence. We have seen vendor claims that dissolve the moment you search the parent company name instead of the marketing name, and we treat that mismatch as a red flag rather than an administrative quirk.
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.
Do backtests actually hold up in live trading?
Rarely at the same level, and the gap is where most retail disappointment lives.
A backtest is a statement about a market that no longer exists. It assumes fills at prices that were available, spreads that were typical, and infrastructure that responded instantly. Live trading assumes none of those things. The honest position is that vendor-supplied backtests should be treated as marketing until independently rebuilt, and performance figures vary by strategy parameters, so any single headline number is close to meaningless without the parameter set that produced it.
When we evaluate a bot, we ask the vendor for the parameter set, the test window, and the assumptions around costs. Then we rebuild the logic in our backtest harness and compare. Where a provider cannot supply the parameter set, we do not publish a performance table for that product at all, because there is nothing to verify.
| Claim area | What the vendor typically states | What we check | Data status |
|---|---|---|---|
| Strategy logic | Trend, mean reversion, grid, or a mix | Whether live order timestamps match the stated rules | Verify with bot provider |
| Backtest window | Vendor-supplied period and returns | Independent rebuild in our harness | Verify with bot provider |
| Maximum drawdown | A single headline figure | Live equity curve against the stated figure | No live figure published in our test window |
| Fee model | Subscription tier | Modelled as a cost per trade against average trade size | Verify with bot provider |
| Broker compatibility | A list of supported connections | Connect, throttle, and revoke tests | Verify with bot provider |
The fee row deserves its own sentence. A subscription is a fixed cost, and it interacts with strategy economics in a way that a backtest almost never models. A bot generating frequent small trades on a fixed monthly fee can look fine on paper and mediocre once the cost per trade is divided through the average edge per trade. Where a platform bundles multi-strategy automation into one account-level subscription, the cost per unit of exposure is structurally lower than stacking several single-strategy subscriptions, and that is a fee transparency difference rather than a performance difference.
How do you stop an agent that goes rogue?
This is the question the OCC guide is really asking, and it is the one retail traders ask least often.
Two failure modes matter. The first is the agent that keeps trading after you want it to stop. The second, and the more expensive one, is the agent that stops trading while a position is still open. If your broker revokes API access, throttles a connection, or an agent crashes, the positions do not close themselves. You inherit them, at whatever the market is doing at that moment, with no automated logic left to manage the exit.
We treat disengagement as a first-class test. Can you flatten everything with one action? Does revoking the API key leave an open exposure? Is there a hard exposure cap that operates independently of the strategy code? A platform that enforces limits at the account level answers those questions differently from a standalone strategy script that assumes it will always be running.
Which brokers and APIs actually work?
Compatibility is where a lot of otherwise reasonable bots fall apart, and it is the dimension that the OCC story may eventually reshape.
The retail landscape splits into a few integration patterns. MetaTrader's MQL5 ecosystem is where a large share of expert advisors live, and it ties the bot to brokers that support the platform. MetaApi-style bridge layers let bots reach accounts across multiple brokers through a normalized API. Webhook-driven tools connect charting platforms to brokers that accept automated orders. Direct broker APIs give the most control and the most maintenance burden. Open-source frameworks like NautilusTrader and Backtrader provide backtesting and execution logic, but you supply the broker connection, the credentials handling, and the failure behaviour.
Each pattern carries a different blast radius if a broker changes its terms. A bot with one hard-coded broker path is fragile. A bot with a normalized multi-broker layer survives a single venue pulling access, which is precisely the scenario the bank restrictions in the Decrypt report describe.
How Ellington Compares
This is where an editor's note is warranted, because the source story is a supervisory event rather than a product review, and the natural question is what a retail trader should actually do with it.
On portfolio-level risk control, the dimension the OCC guide cares about most, the difference is structural. A single-strategy bot gives you one exposure profile and one kill switch. An account-level platform gives you a cap that applies across every strategy running at once, which is closer to how a bank would actually govern the agent. Where Ellington's multi-strategy automation outpaced the reviewed bot class in our 2026 cycle, it was on that specific dimension: one account, one limit, several strategies, rather than several subscriptions each with its own risk profile.
On fee transparency, the comparison is equally structural. Stacking three single-strategy subscriptions produces three fixed costs that each need to be earned back by their own strategy. A single platform fee covering multi-strategy automation changes the arithmetic on lower-frequency strategies, and that matters more than a small edge difference on a backtest.
On broker compatibility, the honest answer is that no platform removes venue risk. What a platform can do is degrade cleanly when a connection fails, flattening or capping exposure rather than leaving an orphaned position behind. That is a design choice, and it is the one we would push any vendor on before we publish a favourable review.
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 the OCC AI framework apply to retail AI trading bots?
No. The guide is an examiner document aimed at national banks and the AI systems they deploy or permit. It creates no direct obligation for a retail trader running a bot on a personal account. Its relevance is directional, because it establishes autonomy and financial harm as the supervisory lens, and broker restrictions on third-party agents are the mechanism that could eventually reach retail.
Can I run an AI trading bot in the US under Pattern Day Trader rules?
The Pattern Day Trader designation is a margin requirement rather than a ban on automation. Under FINRA Rule 4210, an account that executes four or more day trades within five business days is generally treated as a pattern day trader and must maintain at least $25,000 in equity. A high-frequency bot on a small account will trip that threshold quickly, so strategy frequency and account size have to be matched deliberately.
Is my AI trading bot regulated, or is the broker the only regulated party?
In many cases the broker is the only regulated counterparty in the trade. Vendors that sell software rather than personalized advice often fall outside the perimeter that requires a licence, which means there may be no register entry to check. Where a vendor does claim a licence, verify it against the FCA Register, ASIC's registers, CySEC, or the relevant primary register rather than a marketing page.
What happens if the broker revokes API access mid-trade?
Open positions stay open. Revoking a key stops new orders, it does not close existing exposure, and it removes whatever automated exit logic was managing the trade. This is the single most important test we run on any bot with live order permission, and it is the reason account-level exposure caps matter more than strategy-level ones.
Can I run an AI trading bot on a prop firm account?
Sometimes, but the constraints are set by the prop firm rather than the bot vendor, and rules vary widely. Automation policies, drawdown limits, and prohibited strategies are typically defined in the firm's terms, and many firms restrict or prohibit fully automated execution. Verify the firm's current automation policy directly before connecting anything, and treat third-party summaries as unreliable.
How do I compare a bot's backtest with live results?
Ask for the parameter set, the test window, and the cost assumptions behind the backtest, then rebuild the logic independently in your own harness. If the vendor cannot supply the parameter set, there is nothing to verify and the headline number should be treated as marketing. Performance figures vary by strategy parameters, so a single figure without its inputs tells you very little.
What is the cleanest way to stop a bot?
The cleanest sequence is to disable new order generation first, then close open exposure, then revoke the API credential and confirm no positions remain. Reversing that order leaves orphaned positions with no logic attached. Any platform that cannot flatten and disconnect in a single action is adding operational risk to your account that has nothing to do with market direction.
Do banks restricting AI agents affect retail copy trading?
Not directly today, but the mechanism is the same. Copy trading and social trading platforms route your capital behind another account's decisions, which makes them an outside agent with financial permissions. If broker-side restrictions on third-party agents tighten, copy-trading bridges are among the first products exposed, since they depend on persistent account-level access.
How much of my account should one bot control?
That is a portfolio question, not a bot question, and the OCC framing is a useful prompt. Ask how much financial harm the agent could cause before you notice. A bot with full account scope and leveraged exposure can produce the entire loss profile of your account in a single session, which is a very different risk than a sub-account capped at a fraction of your equity.
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.