LLM Outputs as Legal Liability: What Every Compliance Officer Needs to Know

The compliance officer at a mid-size asset manager did not set out to create a liability. She approved an LLM deployment for internal research summarization. The model was good. The outputs looked right. Analysts were faster.

Eighteen months later, during a routine SEC examination, the examiner asked to see documentation of how AI-generated summaries were reviewed before informing client recommendations. There was no documentation. Not because anyone had been careless — but because no one had built the infrastructure to create it.

This is the compliance gap that exists in most firms deploying LLMs today. It is not a question of intent. It is a question of infrastructure.

What the regulators have actually said

The regulatory language is moving faster than most compliance teams realize.

SEC (2023–2024): Staff guidance on AI in investment advisory and broker-dealer contexts makes clear that AI-generated outputs that inform client-facing decisions are subject to the same documentation and review requirements as analyst outputs. The "but the AI said it" defense does not exist.

FINRA (2024 Examination Priorities): Explicit inclusion of AI governance in examination priorities, with specific focus on whether firms have policies, supervision structures, and documentation practices for AI use in customer communications and research.

FCA (Consumer Duty, 2024): Consumer Duty requires firms to demonstrate that AI systems used in customer-facing contexts produce outcomes that can be explained and defended. "The model is a black box" is an explicit non-starter.

OSFI (B-10 Guideline): Canadian federal financial institution guideline on technology and cyber risk includes AI governance as a first-class concern, with requirements for model risk management that extend to LLMs.

The pattern is consistent: regulators are treating AI outputs as artifacts that must be documented, traceable, and reviewable — not as automatic or unreviewed decisions.

The three liability categories

Category 1: Client-facing outputs

Any AI output that directly reaches a client — whether a summary, a recommendation, a communication draft — is the highest-risk category. Regulators will ask: was this reviewed? By whom? What was the review process? Can you produce the original AI output alongside the final delivered content?

If you cannot answer all four of those questions with documentary evidence, you have a Category 1 liability.

Category 2: Decision-informing outputs

AI outputs that inform human decisions — research summaries, risk assessments, document extractions — are Category 2. The human made the decision, but the AI shaped it. Regulators are increasingly asking about the inputs to human decisions, not just the decisions themselves.

The documentation requirement here is: can you show that the AI input to a decision was accurate, and that the human reviewer had the information to evaluate it?

Category 3: Process outputs

AI outputs used in internal processes — compliance screening, contract review, policy checking — are lower immediate visibility but carry accumulating risk. If an AI-assisted compliance screen misses a violation, and you cannot demonstrate how the screen was conducted, the process failure looks worse, not better, for having used AI.

What defensible AI documentation looks like

Defensible documentation is not a paper trail of disclaimers. It is a technical record that allows reconstruction of how a specific output was produced.

For each AI output, the record should capture:

1. The exact prompt and any retrieved context (the inputs)
2. The model version and configuration (the system)
3. The output as produced, before any human editing (the raw artifact)
4. A verification score against source documents, with flagged claims (the quality record)
5. The reviewer identity and timestamp if human review occurred (the governance record)
6. The final delivered form if different from the AI output (the comparison record)

This is not aspirational. It is the minimum documentation package that satisfies current regulatory guidance across SEC, FINRA, FCA, and OSFI frameworks.

Most LLM deployments produce none of it by default.

The cost of inaction

The average regulatory fine for AI governance failures in financial services reached $2.4M in the past 18 months. That number understates the cost: examiner remediation requests, remediation work, enhanced supervision periods, and reputational effects with institutional clients are not in the fine total.

More immediately: firms that cannot produce AI documentation during examinations face follow-up examinations. Follow-up examinations are expensive in time and distraction, even when they do not result in enforcement action.

The compliance officer who approved the research summarization tool did not make a bad decision. She made a decision that was not yet documented. The infrastructure to document it is now standard expectation, not exceptional practice.

The practical first step

Before any larger infrastructure investment, conduct a workflow inventory. For each LLM workflow in production, answer:

• What outputs does it produce?
• Which of the three liability categories do those outputs fall into?
• What documentation currently exists for a specific output on a specific date?
• How long would it take to produce that documentation for an examiner?

The inventory will show you where the gaps are. The gaps will tell you where to start.

Fact AI Lab instruments the gap — the distance between "we use AI" and "we can prove what our AI did." If you want to talk through your specific workflows, the conversation starts here.

This post is not legal advice. For specific regulatory guidance, consult qualified legal counsel familiar with your jurisdiction and regulatory context.