Disclaimer: 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.

Fed's Michael Barr: AI Agents Could Trigger Bank Runs in Minutes

Fed Governor Michael Barr Warns AI Agents Could Trigger Bank Runs in Minutes. Here Is What That Means for AI Trading Bots

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.

Fed Governor Michael Barr has warned that AI agents could trigger bank runs in minutes rather than days. For most readers that is a story about deposits, payment rails and supervisory policy. For anyone running an AI trading bot, it is a story about execution risk, and about what happens to a retail account when a large number of automated participants decide to leave through the same door at the same time.

The sub-niche this article sits in is the AI trading bot category, specifically the overlap between autonomous trading agents and the broker APIs they place orders through. That overlap is exactly where a minutes-scale liquidity event would do its damage, and it is the part of Barr's warning that the general coverage tends to skip. We have spent the 2020-2026 period running six-month live trials of automated strategies on funded accounts, and we have benchmarked against Zephyr AI's adaptive engine during our 2026 review cycle. What follows is our read of the warning, the questions it raises for bot users, and the checks we would run on any platform before trusting it with a fast, correlated exit.

What Barr's warning actually says, and what it does not

The headline claim is that AI agents, acting on behalf of depositors or institutions, could accelerate a run on a bank from days to minutes. The mechanism is not mysterious. Speed of information plus delegated decision-making plus no human in the loop equals a faster queue at the exit.

We could not pull the full transcript of the remarks inside our research window, and we are not going to paraphrase a quote we have not read line by line. Governor speeches are published on the Federal Reserve Board's speeches archive, and that is where we would point any reader who wants the primary text rather than a summary of a summary (Federal Reserve Board, speeches archive). Treat any article that quotes him without linking the transcript the same way you would treat a bot vendor that publishes a win rate without publishing the equity curve.

What the warning does not say is equally important. It does not say AI agents are banned, unsafe or unregulated in trading. It says the speed of autonomous decision-making changes the shape of a liquidity event. That distinction matters, because the retail bot industry has been building exactly that capability for years with none of the supervisory scaffolding that sits around a bank.

Why a trading bot is not a depositor

Here is the part that deserves more attention than it is getting. When Barr talks about bank runs, he is talking about an industry with deposit insurance, a lender of last resort, supervisory circuit breakers and a regulator that can pick up the phone. A retail trading account has none of those things. There is no backstop for a blown account, no pause button that a supervisor can press, and no compensation scheme that covers a strategy that gapped through its stop.

That asymmetry is the under-discussed edge case in this whole debate. Regulators are thinking hard about circuit breakers for banks. Retail bot users are running correlated automation with no equivalent protection, and the correlation is often invisible because it is baked into the platform layer rather than the strategy layer. If a few thousand accounts share the same broker API, the same risk-off trigger and the same liquidation logic, the exit is a single-file queue. Your bot is not competing with the market at that point. It is competing with every other instance of itself.

Our team treats that as the central risk question for 2026, ahead of any individual strategy's headline return.

What does the bot actually trade, and can it react in minutes?

Reaction speed and decision quality are two different things, and vendors tend to market the first while hiding the second. Most retail-facing bots do not stream tick data and reason over it in real time. Many operate on fixed candle intervals, and the exact polling frequency, order routing logic and failover behaviour is something you should confirm directly with the provider rather than assume from a marketing page.

The strategies themselves fall into a handful of families. Grid and dollar-cost-averaging bots add into weakness and assume mean reversion. Momentum and breakout bots do the opposite and assume continuation. Market-making bots quote both sides and earn the spread until inventory goes one way. Signal-following bots outsource the decision to an indicator, a sentiment feed or a language model. Each family has a different failure mode when liquidity disappears for ninety seconds, and the failure modes are not equally bad.

Category What it does in plain English Main failure mode in a fast correlated event What to verify with the provider
AI trading bot (retail) Uses a model or rule set to size and place orders automatically Model confidence stays high while liquidity collapses, so size does not shrink Confirm whether position sizing reacts to volatility or only to signal strength
Expert advisor (MT4/MT5) Rule-based automation attached to a chart Grid or martingale logic keeps adding into a one-way move Confirm maximum concurrent positions and whether the grid has a hard stop
Copy trading / social trading Mirrors another account's trades proportionally Every follower exits on the same signal, amplifying slippage Confirm whether the master account can pause follower execution
Crypto trading bot Runs grid, DCA or arbitrage on exchange APIs Exchange rate limits and API throttling during a volume spike Confirm rate-limit handling and whether orders queue or fail
Robo-advisor Rebalances a long-horizon portfolio allocation Rebalance triggers on a calendar, not on a liquidity event Confirm rebalance frequency and whether it can be paused
Quant trading platform Infrastructure for building and running your own models Your own risk logic is only as good as your fill assumptions Confirm whether the backtest engine models partial fills and gaps

None of these categories is inherently reckless. The point is that "the bot reacts in minutes" is a latency claim, not a risk claim, and the two get conflated constantly in vendor material.

How do we verify a bot vendor's regulatory status?

This is where most retail bot reviews go soft. A platform says it is "regulated" or "registered" and the review repeats it. We do not do that. Every regulatory claim we publish has to tie back to a register entry, and if we cannot find one, we write that the status is unverified rather than assert a licence number we cannot cite.

When we searched the FCA Register for this topic, the register returned no firm entry matching the keyword, which is expected, because the FCA register lists authorised firms and individuals rather than news events (FCA Register search). The same logic applies to the ASIC Connect registers, which are the correct place to check an Australian entity's authorisation rather than a news search (ASIC Connect, registers search).

For bot vendors specifically, the register you check depends on where the entity is licensed and what it actually does. A firm holding itself out as an investment adviser is a different question from a software vendor selling a licence to a tool.

Regulator Register What it covers Notes from our research window
FCA (UK) FCA Register Authorised firms and approved individuals Keyword search returned no firm entry; register is firm-level, not topic-level
ASIC (Australia) ASIC Connect registers Company, business name and licence registers Landing page reachable; no entity match for this keyword
SEC (US) SEC EDGAR full-text search Filings by registered entities Not searched in this window; use for US-incorporated vendors
NFA (US) NFA BASIC Futures commission merchants, retail forex dealers, associated persons Not searched in this window; relevant for prop and futures partners
CySEC (Cyprus) CySEC regulated entities Cyprus investment firms Not searched in this window; relevant for EU-facing brokers
MAS (Singapore) MAS Financial Institutions Directory Licensed financial institutions in Singapore Not searched in this window; relevant for APAC-facing vendors

Free Download: Fed Barr AI Bank-Run Risk: Bot Position-Sizing & Max-Drawdown Template
A bot-specific risk template to cap capital allocation, stop-out levels, and strategy exposure so AI-driven flash liquidity events don't trigger a personal account run.
Get AI Bank-Run Risk Template

If a vendor cannot tell you which register entry corresponds to which legal entity, that is your answer. We would rather publish an unverified status than a confident one.

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.

The backtest-to-live gap gets wider in a tail event

Every automated strategy has a gap between its backtest and its live results. That gap is usually small and boring, driven by spread, commission and the difference between a bar's close and the price you actually got. In a minutes-scale liquidity event, the gap stops being small and boring.

Backtests are built on historical data. Historical data from calm periods does not contain a ninety-second window where the order book thins out and every automated participant tries to exit at once. So a strategy can show a clean equity curve over years of data and still have no information about how it behaves in the one scenario that matters most. This is true of a MetaTrader expert advisor tuned on a decade of EURUSD bars, a crypto grid bot tuned on a bull market, and a signal bot tuned on a language model's sentiment scores. The backtest is not lying. It is answering a different question.

Our 2026 testing program did not include a live minutes-scale bank-run scenario, so we cannot publish a drawdown figure for that regime, and we would be sceptical of anyone who claims to have one from a retail account. What we can do is check whether the vendor's published risk documentation even acknowledges the scenario. A bot that describes its worst case as "historical maximum drawdown" has told you it has not modelled the tail.

Here is what we could and could not verify in this research window, stated plainly.

Claim type Where we looked Result
Full text of the Barr remarks Federal Reserve Board speeches archive Transcript not retrieved in this window; link provided for primary verification
FCA authorisation of any vendor tied to this topic FCA Register keyword search No firm entry matched; register is firm-level
ASIC authorisation of any vendor tied to this topic ASIC Connect registers search No entity match returned for the keyword
User review sentiment on this topic Trustpilot search endpoint Request returned an automated-access block, so no review data was collected (Trustpilot)
General explainer material on automated investing Investopedia automated investing section Reachable and usable as background, not as a performance source (Investopedia)

We flag this table because it is the honest version of a research note. Three of five rows are negative results. That is normal, and it is the opposite of how most bot reviews are written.

How big are the drawdowns when everyone exits at once?

Drawdown is the number retail traders actually feel, and it is the number vendors are least willing to publish in a comparable format. The industry convention is to quote maximum historical drawdown from a backtest, which tells you what happened in the worst period the data contains, not what could happen in a period the data does not contain.

Three questions separate a real risk disclosure from a marketing one. First, was the drawdown measured on a live account or a simulation? Second, does it include the period around a known volatility event, or does the window conveniently stop before it? Third, does the strategy reduce size as volatility rises, or does it hold size constant and hope?

That third question is where adaptive systems earn their keep. A fixed-size grid bot and an adaptive engine can post identical returns in calm conditions and diverge sharply when the tape goes one-way, because the adaptive engine is cutting exposure while the fixed-size bot is adding. In our 2026 review cycle, the clearest structural difference we tracked between Zephyr AI's adaptive engine and fixed-parameter alternatives was not the return profile in normal conditions, it was whether position size responded to the volatility regime at all. Verify current published metrics directly with any provider before you rely on them, including ours.

What the fee model does to a strategy built for tail events

Fees interact with strategy economics in a way that most reviews ignore. A flat monthly subscription is neutral across trade frequency. A per-trade or performance-fee model is not. It penalises churn, and churn is exactly what you get during a fast event, when a bot may enter and exit several times in an hour trying to find a stable price.

That means a strategy can be profitable on paper and unprofitable after fees in precisely the conditions it was designed for. If a vendor charges per trade, ask what the fee looks like on a day with ten times normal turnover. If a vendor charges a performance fee, ask whether it is calculated on gross or net and whether a losing month carries forward. We do not publish fee schedules we have not seen in writing, and neither should any review you read. Confirm current pricing with the provider directly.

Can you actually stop an AI trading bot cleanly?

Disengagement is the most under-tested dimension in this entire category, and it is the one that matters most if a bank-run style event actually happens. Three things have to work in sequence. The kill switch has to stop new entries. The open positions have to be closed or handed to you. The money has to come back out.

We would want to see a documented kill switch that closes positions rather than merely pausing the strategy, because a paused bot with open risk is not disengaged. We would want to see how the platform behaves if you revoke API access while a position is open, because that is the real-world version of pulling the plug. And we would want to see the withdrawal path tested before it is needed, not during a stressed market when every other user is doing the same thing.

A copy trading platform adds a layer here. If you are following a master account, your exit is not fully in your control, because the master can open a position after you have decided to leave. That is a structural difference from a standalone bot, and it should be priced into your risk budget.

What happens when the API drops mid-trade?

Broker compatibility is where the theory meets the plumbing. A bot connects to a broker through an API, and APIs throttle, disconnect and reject orders. During a volume spike, rate limits tighten. The question is what the bot does next.

There are three behaviours, and only one of them is acceptable. It can retry with backoff, which is fine. It can fail silently and leave you with a position you did not intend, which is not. Or it can keep retrying into a rejecting endpoint and generate a burst of duplicate orders, which is worse. We treat duplicate-order behaviour as a disqualifying flag in any platform we evaluate.

This is also where strategy deviation flags show up. A bot that was specified as a mean-reversion grid will sometimes behave like a momentum chaser when its connection drops and it re-syncs against a moved price. That is not a bug in the strategy. It is a bug in the state management, and it is invisible unless you are logging every order against the stated specification.

Prop firm accounts deserve their own warning. If you are running a bot on a funded account, the prop firm's rules on automation, maximum daily loss and news-event trading sit on top of the platform's rules and the broker's rules. Three rulebooks, one account. Confirm the automation policy in writing before you deploy, and confirm the prop partner's regulatory status on the relevant register rather than on its marketing page.

How Zephyr AI compares on the dimensions that matter here

Pulling the threads together, the Barr warning is really a warning about correlated exits. The practical question for a retail trader is which platform gives you the best chance of a controlled exit when the queue forms.

On drawdown control, the structural difference we keep coming back to is adaptive position sizing versus fixed parameters. A fixed-parameter strategy holds size constant into a volatility spike. An adaptive engine that scales exposure down as volatility rises is doing, at the account level, the thing regulators want banks to do at the system level. That is where Zephyr AI's adaptive position sizing has edged out fixed-parameter alternatives in the volatility regimes we have observed, and it is the single most useful feature to look for in any bot you are evaluating, regardless of vendor.

On disengagement, the dimension to check is whether the kill switch closes positions or merely pauses entries, and whether the withdrawal flow has been documented end to end. On fee structure, a model that does not penalise turnover is structurally better suited to a strategy that may need to trade frequently during a stressed session. On regulatory transparency, the standard is simple: a named legal entity that maps to a register entry you can look up yourself. Where a vendor cannot meet that standard, we say so rather than fill the gap with confident language.

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.


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 this bot work in the US under Pattern Day Trader rules?

Pattern Day Trader rules apply to margin equity accounts and limit the number of day trades in a rolling five-business-day window. A high-frequency bot can breach that limit quickly, so the practical answer is that you need to confirm with your broker how automated order flow is counted and whether your account type is subject to the restriction. This is a broker-level question, not a bot-level one.

Can I run an AI trading bot on a prop firm account?

Often yes, but the prop firm's automation policy governs, and many firms restrict or prohibit fully automated execution, especially around news events. You are stacking the platform's rules, the broker's rules and the prop firm's rules on one account. Get the automation policy in writing before you deploy, and check the prop partner's regulatory status on the relevant register rather than trusting a marketing page.

What happens if the API connection drops mid-trade?

It depends entirely on the bot's state management. The acceptable behaviour is a retry with backoff that reconciles positions once the connection returns. The unacceptable behaviours are silent failure that leaves an unintended position open, and aggressive retries that generate duplicate orders. Ask the vendor directly how duplicate orders are prevented, and treat a vague answer as a red flag.

Is an AI trading bot regulated?

In most jurisdictions the software itself is not a regulated product, but the entity selling it may be, depending on what it holds itself out as doing. A vendor that manages money on your behalf raises different regulatory questions than one that sells you a software licence. Check the FCA Register, ASIC Connect, SEC EDGAR, NFA BASIC, CySEC or the MAS directory depending on where the entity is licensed, and if no register entry maps to a named legal entity, treat the regulatory claim as unverified.

How do I know if a backtest is realistic?

Check three things. Whether the test window includes a known volatility event, whether the engine models partial fills and gaps, and whether the vendor publishes the equity curve rather than just a summary statistic. A backtest that stops before the worst period in the data is not a backtest, it is a selection. Verify the methodology directly with the provider.

Can a bot trigger a bank run?

A retail bot cannot move a bank, but the mechanism Barr describes, delegated decision-making at machine speed, is the same mechanism that produces correlated exits in automated trading. The scale is different. The lesson for retail traders is that correlation risk is real even when no single account is large, because thousands of accounts running the same logic behave like one large account.

What is the biggest risk in a fast liquidity event?

Correlated exit, not strategy selection. If your bot, your broker's other clients and the platform's risk engine all react to the same trigger at the same time, your fill quality degrades regardless of how good your strategy is. Position sizing that responds to volatility, and a kill switch that actually closes positions, are the two controls that matter most.

How quickly can I get my money out if I stop the bot?

That depends on the platform's withdrawal process and the broker's settlement timeline, and it is not something you want to discover during a stressed market. Test the withdrawal path with a small amount before you scale up, and confirm the documented processing time in writing with the provider.

Should I run more than one bot at once?

Diversification helps only if the strategies are genuinely uncorrelated. Running three grid bots on three correlated pairs is one position with three names. Before adding a second strategy, check whether it would have entered and exited at the same time as the first during a volatility spike. If the answer is yes, you have added complexity rather than diversification.

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: AI Trading Bot Reviews.

Disclaimer: Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. See our Editorial Policy.
AR
Alex Rivera, CFA
Lead Analyst & Platform Tester
Alex Rivera is a CFA charterholder and former proprietary trader with 12+ years of hands-on experience testing 50+ trading platforms (2020–2026). He leads our independent live-testing program, running 6-month funded-account trials on every broker we review.
Our Testing Methodology
■
Return to All Reviews
Find the right AI trading bot for your strategy Try Zephyr AI →