The AI Herding Problem: Why Faster Trading Systems Could Make FX Markets Less Diverse in 2026
Zeeshan · 2026-09-27
AI may diversify decisions inside each firm while making the FX market more correlated through shared models, data, vendors, and objectives.
By Zeeshan | YouYaa Intelligence | 27 September 2026
A market can become faster without becoming more diverse.
The controversial thesis
Artificial intelligence is entering one of the world’s most automated markets: foreign exchange.
The scale is enormous. Average daily FX turnover was shown as $9.6 trillion for April 2025 in New York Fed FX Committee material.[1] Electronic execution in spot FX is near 75%, including dealer-to-customer trading.[1]
The controversial risk is not only that one AI makes a bad trade.
It is that many firms may use similar models, similar data, similar vendors, and similar profit goals. Each firm may pass its own model tests while the market becomes more correlated.
That is the AI herding problem: individual systems look controlled, but collective behaviour becomes harder to predict.
This is not a prediction of an AI-driven crash. The official sources describe early adoption and emerging risks, not a current market failure. But the control question is urgent:
Who tests what happens when thousands of “independent” systems react in the same direction at machine speed?
The adoption numbers
The market is not starting from zero.
The New York Fed material cites a Bank of England/FCA survey in which 75% of firms already used AI, and 55% of AI use cases involved some automated decision-making.[1]
FMSB says market-facing AI is still at a relatively early stage. It is normally embedded in existing trading infrastructure and remains subject to direct or indirect human supervision.[2]
The result is a market in transition: broad experimentation, heavy electronic execution, and control frameworks designed for systems whose logic was more stable.
| Signal | 2026 evidence | Why it matters |
|---|---|---|
| Average daily FX turnover | $9.6tn | Small behavioural changes can operate at huge scale |
| Electronic spot-FX execution | Near 75% | AI acts inside an already automated market |
| Firms using AI | 75% | Adoption is no longer limited to a few experiments |
| AI use cases with some automation | 55% | Human review may not cover every decision |
| Fully autonomous trading | Not currently present, according to FMSB | Future risk is not the same as current fact |
Why herding can happen
Traditional algorithm governance assumes that logic is relatively stable. A firm can document a system, approve it, test it, and monitor it.
Adaptive AI is different. Retraining, recalibration, data changes, and third-party model updates can alter behaviour between formal approval points.[1]
A small change may appear harmless inside one system. Across many systems, it can shift liquidity, execution, pricing, or risk appetite at the same time.
| Source of alignment | Possible collective effect |
|---|---|
| Shared data vendors | Similar signals arrive at similar times |
| Shared foundation models | Different firms receive related outputs |
| Similar optimisation goals | Systems prefer the same liquidity or risk response |
| Common cloud or model providers | A vendor change affects many firms |
| Similar risk limits | Firms withdraw or hedge together |
| Rapid retraining | Behaviour changes faster than governance cycles |
The market can therefore become less diverse even when every firm believes it is acting independently.
The speed problem
FX already operates at electronic speed. Adding AI does not simply automate a manual process. It can compress the time between information, price formation, execution, rejection, hedging, and post-trade action.
The New York Fed material identifies speed and automation as factors that can increase opacity and execution asymmetries.[1]
For a CFO or treasury team, that matters because a hedge can fail for reasons that are difficult to explain after the fact. For a fintech operator, it matters because a model update can change execution behaviour without changing the product interface.
For an HNWI or family office, it matters because a supposedly liquid market can behave differently when many systems respond to the same shock.
The accountability gap
FMSB says clear human accountability remains essential even as AI systems become more adaptive. It also says fully autonomous trading systems are not yet present in financial markets.[2]
That distinction is important.
A human does not need to click every trade to remain responsible. But the firm must know which decisions the system can make, which boundaries apply, how changes are approved, who can stop the process, and how the decision can be reconstructed.
| Control question | Minimum evidence |
|---|---|
| What can the AI change? | Documented decision and parameter boundaries |
| Who approves updates? | Named accountable owner and change log |
| How is drift measured? | Continuous performance and behaviour monitoring |
| What happens when the vendor updates its model? | Revalidation and rollback process |
| Can the system be stopped? | Tested kill switch and human override |
| Can the firm explain a bad execution? | Meaningful decision records |
| What if several firms use the same model? | Market-wide concentration assessment |
The last question is often missing from firm-level governance.
The market-wide blind spot
Most controls begin inside the firm. That is logical, but incomplete.
A bank may test its model against historical prices. A fintech may test its execution engine against stress data. A fund may test liquidity under a range of scenarios.
None of those tests alone answers what happens when many participants use similar responses at once.
The New York Fed material points to model herding, common model use, third-party concentration, and sharp market movements as issues for the industry.[1]
| Firm-level view | Market-wide view |
|---|---|
| Is our model accurate? | Are many models becoming similar? |
| Is our system within limits? | Do common limits trigger together? |
| Is our vendor reliable? | How many firms depend on the vendor? |
| Can our kill switch work? | What happens if many kill switches activate? |
| Is our execution explainable? | Can the market explain a synchronized move? |
The controversial conclusion is that AI governance cannot stop at model validation. It must include collective-behaviour monitoring.
What it means for fintech operators
Fintechs often sell speed, automation, and better execution. Those are valuable features. But AI changes the meaning of “better.”
A faster system may find liquidity quickly in normal conditions and withdraw quickly in stress. A model may improve execution costs for one client while increasing adverse selection for another. A third-party model may be updated without a visible product release.
Operators should treat model updates as product and market-structure events, not only engineering events.
What it means for CFOs and treasury teams
CFOs should ask more than whether a provider uses AI. They should ask how AI affects execution, hedging, liquidity, and records.
| CFO question | Why it matters |
|---|---|
| Does the system retrain or recalibrate? | Behaviour may change between approvals |
| Are third-party models used? | The firm may depend on external updates |
| What is the fallback execution route? | Speed is not resilience |
| How are rejected or delayed trades explained? | Audit and conduct risk may follow |
| Can the system be overridden? | Human accountability needs practical power |
| Are multiple providers using similar models? | Correlation can hide inside vendor diversity |
The aim is not to ban AI. It is to make the hidden dependencies visible.
Stress tests for the AI-herding problem
A useful stress test should not ask only whether one model fails. It should ask whether several systems converge.
Test scenarios might include a common vendor update, an unexpected data break, a sharp currency move, a liquidity gap, a false signal, a cyber event, and simultaneous model de-risking.
| Scenario | Test question |
|---|---|
| Common model update | How many strategies change at once? |
| Data outage | Which systems use the same backup? |
| Sudden FX move | Do models add liquidity or remove it? |
| False market signal | How quickly can the signal spread? |
| Vendor failure | Can firms operate independently? |
| Synchronized de-risking | Does the market become one-way? |
| Override event | Can humans slow the system before damage grows? |
Conclusion
The New York Fed material shows the scale: $9.6 trillion in average daily FX turnover, near 75% electronic spot-FX execution, 75% of surveyed firms using AI, and 55% of AI use cases involving some automated decision-making.[1]
FMSB adds an important limit: market-facing AI is not currently fully autonomous, and human accountability remains essential.[2]
The controversial conclusion is:
The biggest AI trading risk may not be a rogue machine. It may be a market full of well-controlled machines that learn the same lesson at the same time.
For 2026, the winning control framework will need two layers: firm-level governance and market-wide collective-behaviour monitoring.
FAQ
What is AI herding in financial markets?
AI herding occurs when many systems use similar data, models, vendors, objectives, or risk rules and begin reacting in similar ways.
Is AI already trading FX fully autonomously?
No. FMSB’s February 2026 review says market-facing AI is not currently autonomous and remains subject to direct or indirect human supervision.[2]
How large is the FX market?
The New York Fed material presents average daily FX turnover of $9.6 trillion for April 2025.[1]
How much FX execution is electronic?
Electronic execution in spot FX is near 75%, including dealer-to-customer trading.[1]
How widely are firms using AI?
The cited Bank of England/FCA survey found 75% of firms already used AI, and 55% of AI use cases involved some automated decision-making.[1]
Why are third-party models a risk?
A vendor update, model drift, outage, or data change can affect many firms at once. This can create concentration and correlation that firm-level testing misses.
What should CFOs monitor?
Monitor retraining, vendor dependencies, fallback execution, model changes, decision records, override controls, and whether multiple providers rely on similar models.
Is this investment advice?
No. This is general analysis of AI use in trading markets. Obtain appropriate legal, regulatory, and financial advice for specific circumstances.
References
[1] New York Fed FX Committee, Artificial Intelligence (AI) in FX Markets, 15 April 2026
