Skip to content
NewSecurity Audit for AI-Built Apps. Fixed scope, clear report
GeekTech
AI Governance

AI Risk Register: How to Build One and Keep It Current

How to build an AI risk register: what to record for each AI use, how to score likelihood and impact, who owns each risk, and when to review it.

7 min read

GeekTech Engineering Team18 years building and securing software

To build an AI risk register, start from your list of AI tools and features. For each one, write down the specific things that could go wrong, how likely and severe each is, which control is in place, what you'll do about the rest and who owns it. Score against written anchors, set review triggers and keep it in one place. This is general information, not legal advice.

What an AI risk register is, and what it isn't

An AI risk register records risks and the decisions taken on them. The AI inventory, a separate record, lists what you run. The NIST AI Risk Management Framework (AI RMF) expects mechanisms to inventory AI systems (GOVERN 1.6), and our guide to finding the AI tools your team uses covers that discovery. The register starts where the inventory ends.

A data protection impact assessment (DPIA) goes deeper on one use of personal data than a register row can, so the row links to it. Our post on running a DPIA for AI explains when GDPR requires one.

The EU AI Act requires a documented risk management system that runs across the lifecycle of each high-risk AI system (Article 9), mainly an obligation for providers. A register is lighter and covers every AI use, including vendor tools you only deploy.

What to record in an AI risk register

Keep one row per risk per AI use. A tool used by sales and by HR gets separate rows.

ColumnWhat to writeExample
ID and AI useInventory reference and the use, not just the toolAI-07, help center support assistant
Risk statementCause, event and consequence in one sentenceSee the next section
CategoryOne label from your category listWrong output
Likelihood, impact, scoreValues from your anchors, before controls3, 4, 12
Current controlsControls that exist today and how you know they workAnswers limited to retrieved articles; weekly sample review
TreatmentMitigate, transfer, avoid or accept, plus the actionMitigate: refuse when retrieval finds nothing
Residual scoreScore after current and planned controls2, 4, 8
OwnerOne named roleHead of Support
ReviewNext date and events that force an earlier oneQuarterly; model change; new data source

The controls column is where registers go wrong. "Staff are trained" is an intention. "Prompt logs are sampled weekly and the result is recorded" is a control, because someone can check that it happened.

How to write an AI risk statement someone can act on

A row that says "hallucination" can't be scored or assigned. Write a cause, an event and a consequence. A hypothetical example:

Because the support assistant falls back on general model knowledge when no help article matches, it may state a refund policy that doesn't exist, and a customer may act on it and dispute the charge.

That tells the owner what to fix (the fallback), what to measure (answers without a source) and what harm to weigh.

Which risks to look for in each AI use

Run each use through a fixed category list. The NIST Generative AI Profile (NIST AI 600-1) defines risks unique to or made worse by generative AI, including confabulation, data privacy, information security, intellectual property, human-AI configuration and value chain and component integration. A company-level list can map to those:

  • Data exposure: personal or confidential data in prompts, logs, evaluation sets or vendor retention.
  • Wrong output: confident errors that a person or system acts on.
  • Security: prompt injection, tools with more access than the task needs, unsafe handling of model output.
  • Decisions about people: ranking or filtering candidates, customers or staff.
  • Over-reliance: reviewers who approve AI output without reading it.
  • Vendor and supply chain: model retirement, changed data terms, an outage with no fallback.
  • Legal and regulatory: transparency duties, high-risk classification, copyright.
  • Cost and availability: usage spikes, rate limits, runaway agent loops.

Mark categories that don't apply as "not applicable" with a one-line reason.

How to score likelihood and impact for AI risks

Scores only compare if everyone uses the same written anchors. This scale is an example to adapt:

ScoreImpactLikelihood
1Internal inconvenience, fixed within a dayNot seen in testing; only trusted staff provide input
2Rework or delay visible to one customerSeen rarely in testing
3Several customers affected, or confidential data seen by the wrong internal teamSeen in testing at a measurable rate
4Personal data exposed outside the company, or a wrong decision about a personSeen in production, or untrusted users control the input
5Regulatory notification, legal claim, or harm to someone's job, credit or healthHappening now, or easy to trigger on purpose

Multiply likelihood by impact and set thresholds. For example: 15 or above needs a treatment plan and leadership sign-off before launch, 8 to 14 needs the owner's plan within a quarter, and below 8 can be accepted and reviewed on schedule.

Two adjustments are specific to AI. Score likelihood by volume, because a rare error rate across thousands of runs a month still produces regular incidents. And raise it when untrusted people shape the input, such as public chat users or inbound email.

Choosing a treatment and recording residual risk

The AI RMF lists the responses as mitigating, transferring, avoiding or accepting (MANAGE 1.3), prioritized by impact and likelihood (MANAGE 1.2), and expects remaining risks to be documented (MANAGE 1.4). For AI uses:

  • Mitigate: restrict answers to retrieved sources, or require human approval for actions that send, pay or delete.
  • Transfer: put retention, training-use and model-change notice terms in the vendor contract. The risk to your customers stays yours.
  • Avoid: drop the use or part of it, for example drafting job descriptions with AI but keeping it out of candidate screening.
  • Accept: record who accepted it, why and until when. An acceptance with no expiry date never returns for review.

If the residual score stays above your threshold, the decision belongs to someone with authority to stop the feature.

How to build an AI risk register in five steps

  1. Pull the inventory. List every approved AI tool and feature with its owner and data categories.
  2. Hold a short workshop per use. Invite the business owner, an engineer who knows how it works, someone from privacy or legal, and whoever would handle the complaints.
  3. Write and score the risks. Walk the category list and score each statement against your anchors.
  4. Assign treatments and owners. Every row above the acceptance threshold gets an action, an owner and a date.
  5. Review the top rows with leadership. Present the highest residual scores and all accepted risks, and record the decisions.

When to review an AI risk register

The AI RMF asks organizations to plan periodic review and decide its frequency (GOVERN 1.5). Quarterly is a reasonable default, but AI systems change between reviews, so add event triggers:

  • A new model, model version or provider.
  • A new data source, retrieval index or tool connection.
  • A new purpose or user group, such as opening an internal tool to customers.
  • Changed vendor terms on retention, training use or sub-processors.
  • An incident, a near miss, falling evaluation scores or rising complaints.

A question in the pull request template or procurement form ("does this change an AI use in the register?") catches more changes than a calendar reminder.

Common AI risk register mistakes

  • One row per tool with a generic label.
  • Scores without written anchors, so two reviewers rate the same risk 2 and 4.
  • Controls listed as intentions rather than checks.
  • Accepted risks with no expiry date.
  • Copying a vendor's risk list, which covers their product and not your data, users or decisions.

If you want an outside view, our AI Readiness & Governance Assessment covers your current AI use, EU AI Act risk classification and the security of AI tools and vendors, and delivers a risk register with governance and policy recommendations.

Frequently asked questions

Does the EU AI Act require an AI risk register?

Not as a named document for most companies. Providers of high-risk systems must run the Article 9 risk management system, and deployers have their own duties. A register shows you've assessed your AI use; our EU AI Act guide for SMBs covers the classification behind it.

What tool should we use for an AI risk register?

A shared spreadsheet works for a first version. If you already run a governance, risk and compliance (GRC) tool, add AI risks there so they follow the same review cycle. What matters is one location, named owners and a change history.

How many risks should each AI use have?

As many as the categories that apply. A customer-facing assistant with access to personal data will have more rows than an internal summarizer. A use with a single row usually means the statements are too general.

Insights

Related insights

More on AI Governance →

Get AI security and automation insights, monthly

AI security, automation and governance, written by engineers. No spam, unsubscribe any time.

Book a free 30-min call