AI Model Risk Register

The idea behind it

A model risk register lists the ways a model can fail, how likely and how harmful each failure is, and what controls are in place. Banks formalized this discipline as model risk management. In the United States, the Federal Reserve and the OCC set it out in 2011 guidance known as SR 11-7, which covers model inventory, validation, ongoing monitoring and named ownership. Large language models raise the same questions, but their failure modes differ from those of a credit scoring model. An LLM can state false facts with confidence, leak personal data placed in its context, follow instructions injected by a user or a document, produce biased text, or answer from stale training data.

The register scores each pairing of a feature and a failure mode. The same model can carry low risk in a marketing draft tool and high risk in a contract summarizer, so risk is assessed per use, not per model. Likelihood and impact are each rated on a 1-to-5 scale, and their product gives a score from 1 to 25 that sorts risks into bands. Controls such as grounding with retrieval, output validation, human review, evaluation suites, audit logging and access controls are recorded against each row. An owner and a review date make sure the score is revisited when the model, the prompt or the use changes.

What's inside

• A register sheet with one row per feature and failure mode: model, likelihood, impact, controls, owner, review date, status and open actions.
• The risk score calculates as likelihood times impact, and its band (low, medium, high or critical) follows from the score. Rows scored 12 or more are highlighted.
• A scoring guide with 1-to-5 likelihood and impact scales, each level with an LLM example, and the score bands for low to critical.
• A sample register with ten rows across seven features, covering failure modes such as hallucination, PII leakage, prompt injection and bias.

Format: Excel workbook.

Who uses it

• AI governance leads, to keep one inventory of LLM features, their failure modes and their controls.
• Product owners of AI features, to show before launch which risks are high and what controls address them.
• Model risk and compliance teams, to apply existing risk practice to generative AI without a separate framework.
• CISOs and data protection officers, to track prompt injection and personal data exposure across every LLM feature.

When to use it

• A team plans to launch a chatbot, summarizer or other LLM feature to customers or staff.
• An AI feature is moving from pilot to production.
• A feature switches to a different model or model version.
• An incident shows an LLM output caused harm or was acted on wrongly.
• Auditors, regulators or the board ask for an inventory of AI risks.

How to use it

Add a row for each LLM feature before it launches and review the high scores on the date set in the register.

1. Read the Scoring Guide sheet so everyone uses the same likelihood and impact definitions and LLM examples.
2. In the LLM Risk Register sheet, add one row for each combination of feature and failure mode, naming the model used.
3. Select likelihood and impact from the 1-to-5 drop-downs; the Risk Score column multiplies them.
4. Record the controls in place, a named owner and a review date for each row.
5. Set Status to Active, Mitigated, Accepted or Closed, and list open actions for rows scored 12 or more.
6. Review the register at least quarterly and whenever a model, prompt or use case changes.

You leave with: A scored inventory of LLM failure modes per feature, with controls, owners and review dates.

Where it goes wrong

• Scoring the model once instead of each use; the same model has different impact in different features.
• Scoring likelihood without considering controls, or listing controls that are planned rather than in place.
• Leaving review dates unset, so scores go stale after a model upgrade.
• Treating hallucination as the only risk and missing prompt injection, data leakage, bias and over-reliance.

Get the register

It is free. Enter your work email and we send the download link.

Questions

Read more on this

• Risk in Enterprise AI: Hallucinations, Bias, and Systemic Failure
• US banking regulators just put generative AI outside model risk guidance, in a footnote
• Buyer’s Guide: AI Governance & Responsible AI