The Fed Just Asked for Your Intraday AI Exposure. Most Desks Can't Answer.
Operator memo
The regulatory question has changed.
It is no longer:
“Do you have an AI policy?”
It is now:
“Can you reconstruct, in minutes, what your AI touched intraday, why it acted, who approved it, and whether your controls held?”
Most desks cannot.
They can produce an end-of-day risk report, a model inventory, a vendor questionnaire, and a policy PDF. Those documents may describe the intended control environment. They do not prove what happened when markets moved quickly, an agent made a recommendation, a trader approved an exception, or a vendor model changed during the session.
That gap is the finding.
What the Fed and BoE reportedly asked for
On September 21, Reuters reported that the Federal Reserve and Bank of England had stepped up scrutiny of banks’ exposures to large trading firms following Jane Street’s reported loss.
The reported questions were operational, not theoretical:
What is the firm’s risk appetite?
How did exposure change intraday?
Did controls actually work?
Could the firm explain its exposure as conditions changed?
This follows Reuters’ August 14 report that Jane Street took a roughly $15 billion hit in July tied to the AI-focused fund Situational Awareness, run by Leopold Aschenbrenner. The fund reportedly sold most of its public-equities portfolio after an AI and semiconductor stock selloff. Major banks were also reportedly subpoenaed by the SEC regarding the fund’s trading, leverage, margin calls, and lender communications.
The specific facts of that situation will be examined by regulators and investigators. The broader lesson is already clear:
A model inventory is not an exposure inventory. A policy is not a control. An end-of-day snapshot is not an intraday reconstruction.
Why intraday reconstruction is the hard part
AI-enabled trading and investment workflows do not move at the pace of committee meetings.
Agents can ingest new data, generate recommendations, alter priorities, route tasks, and trigger downstream actions in seconds. Exposure can move between scheduled risk snapshots. A vendor model can be updated. A prompt, data feed, or policy configuration can change. A trader can approve an exception in a chat channel that is disconnected from the formal audit system.
The evidence is usually fragmented across:
Order and execution management systems
FIX gateways
Market-data vendors
Risk and portfolio systems
Model registries
Agent logs
Email, chat, and approval tools
Vendor platforms
Post-trade and reconciliation systems
The result is a familiar failure mode: each system has part of the story, but no system has the decision chain.
A risk team may know the position. Compliance may know the policy. Technology may know which model ran. The desk may know who approved the trade. None of them can assemble the complete sequence quickly.
That is what makes the request difficult. The question is not simply, “What was your exposure at 10:00 a.m.?”
It is:
Which data did the system use?
Which model or agent produced the recommendation?
Which version was active?
Which policy applied at that moment?
What confidence, threshold, or exception triggered the action?
Who approved it?
What did the system do next?
Did the control block, pause, escalate, or permit the action?
What was the resulting position, P&L, and risk?
The Audit Bomb: one click, full decision chain
We call this requirement an Audit Bomb.
The name is intentionally blunt. An Audit Bomb is the ability to generate a complete, regulator-ready decision trail on demand, not after weeks of manual reconstruction.
It should capture, at minimum:
Inputs: market data, portfolio state, news, documents, prompts, and external signals.
Model identity: model, agent, version, vendor, configuration, and deployment status.
Policy applied: limits, permissions, thresholds, escalation rules, and data-use restrictions.
Reasoning path: the material factors and structured rationale behind the recommendation or action.
Approvals: named users, timestamps, roles, and approval scope.
Exceptions: overrides, failed checks, alerts, and policy deviations.
Action: order, allocation, communication, workflow change, or no-action decision.
Outcome: execution, exposure, P&L, reconciliation, and subsequent control response.
This is not a request for a model to reveal every internal token or produce a vague explanation after the fact. It is a requirement for 100% explainable, governed outputs: a deterministic record of what the system received, what controls applied, what decision was made, and what happened afterward.
The “we will show them the model card” answer fails.
A model card can describe intended use, limitations, training context, and evaluation results. It does not tell an examiner what the model touched during a volatile session. It does not show whether an agent exceeded its mandate. It does not connect the recommendation to the final order or exposure.
Three controls every desk needs
1. Kill Switch: scoped containment, not just shutdown
A kill switch should not mean pulling the plug on every automated workflow.
It should support a scoped pause of:
One mandate
One agent
One model version
One strategy
One asset class
One data source
One approval path
The control must preserve evidence while containing activity. A desk should be able to pause a strategy that is breaching a concentration threshold without disabling unrelated post-trade workflows.
A useful kill switch answers three questions immediately:
What exactly was paused?
Why was it paused?
What activity continued under the remaining permissions?
2. Audit Bomb: export the complete evidence chain
The Audit Bomb should be available to risk, compliance, and authorized desk leaders without requiring an engineering project.
It should export structured evidence with consistent identifiers across data, model, agent, order, approval, and outcome records. If every system uses a different transaction ID, timestamp convention, or model name, the audit trail will break at the joins.
3. Governed workflow gates
Controls must exist inside the workflow, not in a policy document stored beside it.
Examples include:
Pre-trade notional and concentration limits
Funding-cost and liquidity thresholds
Escalation when market volatility exceeds a defined range
Human approval for mandate changes or exception trades
Restrictions on autonomous order generation
Vendor-model change approval
Post-trade reconciliation before the workflow is considered complete
Automatic pause when lineage or monitoring data is missing
These gates are especially important as policy shifts, tariffs, rates, and funding conditions change the operating environment. A tariff announcement may alter sector exposures within minutes. A rate move may change the economics of leverage and financing. A liquidity event may turn a previously acceptable strategy into a funding-cost problem.
The system must be able to adjust the workflow and retain the reason for the adjustment.
The evidence says adoption is ahead of governance
The industry is already using AI. The control layer is lagging.
According to Acuiti survey reporting published by Finance Magnates, 44% of surveyed firms reported quantifiable time and cost savings from AI. Yet only 18% said they could definitely rely on AI to make better decisions during volatile markets.
About one-third of firms had no formal AI governance policy. Roughly half of proprietary trading firms allowed staff to use AI relatively freely. Two-thirds of banks had dedicated AI risk committees, while hedge funds more often routed oversight through existing risk committees.
That is a useful distinction: governance activity is not the same as operational control.
A committee can approve a use case. It cannot reconstruct every decision made by an agent population during a stressed market.
The risk is also not limited to individual model failure. The September 3 arXiv paper, “Why Better Models Can Create Riskier Systems,” argues that more capable models may behave more similarly. Shared training and architectures can create correlated actions, reducing the diversification that firms assume they have.
In capital markets, model heterogeneity becomes a control objective.
If ten agents receive similar data, use similar models, and apply similar reasoning, they may converge on the same action at the same time. A well-governed AI control plane should monitor not only accuracy and drift, but also behavioral convergence across agents and strategies.
What a defensible answer requires
A defensible answer to the Fed, BoE, SEC, or an internal examiner requires a control layer spanning the full operating environment.
A unified control layer
The control plane must connect data, models, agents, people, and systems. It must work across the OMS, EMS, FIX gateways, risk platforms, vendor models, communications, and post-trade tools.
This is the foundation of modern asset management technology, capital markets AI, and investment management AI. The point is not to replace every system. It is to create a governed layer across them.
Per-decision lineage
Every material decision should have a unique identity tied to the underlying data lineage.
That means the firm can trace a recommendation from the original market-data feed or document through the model, policy check, approval, order, and outcome. This is data lineage for financial services applied at decision level, not simply a database catalog.
Real-time model monitoring
Monitoring must cover more than uptime.
Teams need visibility into:
Input drift
Output drift
Confidence changes
Abnormal action frequency
Vendor model changes
Agent-to-agent convergence
Policy violations
Latency and stale data
Changes in realized risk
This is the practical role of an AI model monitoring platform and MLOps for financial services: ensure that the production system remains within its approved operating envelope.
Policy-as-code
Limits must be enforced, not described.
A policy that says “human approval is required for material exceptions” is incomplete unless the workflow can identify a material exception, route it to the right person, block further action, and record the decision.
That is AI governance for finance implemented as executable logic.
Automated evidence generation
Risk and compliance teams cannot hire their way out of this problem. Labor scarcity is a control issue.
If every examination requires analysts to reconcile logs manually across six systems, the firm will either respond slowly or accept evidence gaps. The productivity objective is therefore not just to automate decisions. It is to automate the evidence around decisions.
Productivity is the product.
QUANTEX: the Governed Brain for capital markets
QUANTEX approaches this as a neurosymbolic system: a Governed Brain that combines neural pattern recognition with explicit symbolic controls.
The neural layer can identify patterns, summarize information, prioritize work, and support decisions. The symbolic layer defines what the system is allowed to do, under which conditions, with which approvals, and within which limits.
The result is an AI control plane for capital markets:
Supervised agents across trade lifecycle, compliance, and operations
Human-in-the-loop approval gates
Policy enforcement and role-based permissions
Model and agent monitoring
Complete decision lineage
Scoped containment through kill switches
One-click audit exports
Post-trade reconciliation and exception management
The commercial value is not “AI with more features.”
It is controlled execution that can survive volatility, staffing constraints, vendor changes, and regulatory examination.
Explore the QUANTEX AI Control Plane, or book a demo.
If governed, auditable AI workflows are on your roadmap, contact the QUANTEX team.
The next examination will not ask whether your firm experimented with AI.
It will ask what AI did at 10:17 a.m., and whether you can prove that the controls worked.
Comments