In short: Financial supervisors are not asking to inspect your model. They are asking you to reconstruct what it did, step by step, and who checked. That trace is the same object a judgment system needs in order to learn. Build it once, as the spine of the product.
Key points
- FINRA's oversight report said supervision and records obligations change once a system can take action.
- In April 2026 US banking agencies narrowed model-risk guidance and left generative and agentic AI out of scope pending a request for information.
- The EU AI Act's high-risk credit-scoring obligations have moved to December 2027.
- A usable trace records inputs, evidence and its standing, model and prompt versions, output, human actions, final action and outcome.
- Agree trace retention, immutability and access with compliance before the first production decision.
The sharpest line in 2026 financial supervision came from FINRA's annual oversight report, published in December 2025. It drew a distinction that matters for every enterprise brain: once a system can take action rather than merely generate content, supervision, books-and-records and governance obligations change. It warned that complex chains of agent reasoning can be hard to trace, and pointed firms to tracking and logging agent actions and decisions.
The same report made three further points worth repeating. The controls that apply are the same as for a person performing a comparable function. Systems may appear more expert than they are. And a system optimised for speed can reach a compliant-looking result through non-compliant steps, which is a conduct problem even when the output looks right.
The AI-specific rules are late. The record-keeping rules are not.
In April 2026 the US banking agencies replaced most of their long-standing model-risk guidance with a narrower, risk-based version. It explicitly left generative and agentic AI out of scope, pending a request for information, while expecting analogous governance in the meantime. For vendor models, banks must still understand conceptual soundness, design and performance even when the code is proprietary.
In Europe, the AI Act's high-risk obligations for credit scoring have moved to December 2027. Transparency duties for chatbots and synthetic content began in August 2026. Meanwhile financial supervisors have told firms to fold AI-related risks into their existing operational-resilience frameworks now, without waiting for new rules.
The pattern is consistent. Purpose-built AI regulation is still forming. Existing obligations on supervision, records and resilience already apply, and they are being read to cover agents.
What the trace must contain
A usable trace for an AI-assisted decision records, with timestamps and actors:
- the inputs and the request as made;
- the evidence retrieved, with its standing (status, date, source);
- the model and prompt versions used;
- the output produced;
- every human action: opened, edited, approved, rejected, escalated, with time spent;
- the final action taken, and by whom;
- the outcome, when it arrives.
That list is also, almost exactly, what a judgment system needs to learn: the situation, the evidence, the recommendation, the human correction, the decision and the result. A supervisory trace and a learning record are the same object seen from two sides.
Spine, not afterthought
Teams that treat tracing as an engineering nicety will rebuild it under examination, at speed, from logs that were never designed to answer a supervisor's questions. Teams that design the trace as the product's spine get two things at once: supervision that is close to free, and a learning loop with complete data.
One caution: a trace is a record, so it inherits record-keeping obligations. Agree retention, immutability and access with compliance before the first production decision, not after the first request.
In practice
- Write the trace specification first, with compliance, and build the workflow around it.
- Make every human action in the review gate a logged event with duration.
- Store the evidence standing at the moment of retrieval, not just a link to the current version.
- Test reconstruction: pick a decision from three months ago and time how long it takes to rebuild the full chain.
What would prove this wrong. Examinations of agentic systems that accept model documentation and output sampling without asking for step-level reconstruction. Then the trace is good practice rather than an obligation.
Common questions
What do financial regulators expect for AI agent audit trails?
The ability to reconstruct what the system did step by step and who checked. FINRA pointed firms to tracking and logging agent actions and decisions, and existing supervision, records and resilience obligations are being read to cover agents.
What should an AI decision trace contain?
Timestamped inputs and the request, retrieved evidence with its status, date and source, model and prompt versions, the output, every human action with time spent, the final action and who took it, and the outcome.
Do existing rules apply to AI agents before AI-specific regulation arrives?
Yes. Purpose-built AI regulation is still forming, but existing obligations on supervision, records and resilience already apply and are being read to cover agents.
When do EU AI Act obligations for credit scoring apply?
The article says the high-risk obligations for credit scoring have moved to December 2027, while transparency duties for chatbots and synthetic content began in August 2026.