AI Risk Register: Columns, Owners and a Sample Row to Copy

Ask three managers which AI tools touch customer data. If you get three different lists, your organisation has an AI risk register only on paper, or none at all. A working AI risk register is a short shared table. It has one row per AI use, a named person for each risk, a chosen response and a review date that people keep. Below are the columns, the owners, the review rhythm and a sample row you can copy.

Which signs show an AI risk register has stopped working?

Registers fail quietly. Each symptom below points to a specific fix.

• Rows describe tools, not uses. "ChatGPT" is one row, yet marketing drafts posts with it and HR screens CVs with it. The risks differ, so the rows must differ too.
• The owner column names a department. "IT" cannot answer a question in a meeting. Each risk needs one person.
• Review dates are in the past. If nobody touched a row after a vendor changed its model or its terms, the score describes an old system.
• Every risk is "medium". When every row scores medium, the scores tend to reflect habit rather than evidence, so test a few against what actually happened.
• Nothing is ever closed or accepted. A register that only grows becomes a list of worries, not a record of decisions.

Our position: fix the row unit first. We recommend one row per use case (a tool used for a purpose by a team), not one row per model or vendor. The trade-off is more rows. The gain is that each row has a clear owner and a risk you can actually test.

What do NIST and ISO expect an AI risk register to hold?

No framework prescribes your columns, but the main ones describe what the register must support. The NIST AI Risk Management Framework 1.0, released on 26 January 2023, is voluntary. NIST says it is now revising version 1.0, so check for a newer version before you rely on the item numbers below. In its Govern playbook, GOVERN 1.6 asks for mechanisms to inventory AI systems. GOVERN 1.5 asks for planned monitoring and periodic review of the risk management process and its outcomes, with roles and the review frequency defined.

The Manage playbook says documented AI risks are prioritised by impact, likelihood or available resources (MANAGE 1.2). MANAGE 1.3 says risk response options can include mitigating, transferring, avoiding or accepting. Those four words make a ready-made "response" column.

For generative AI, NIST's Generative AI Profile (NIST AI 600-1, July 2024) defines 12 risks that generative AI creates or makes worse. They include confabulation, data privacy and information security. Use them as prompts when you fill the risk column.

ISO/IEC 42001:2023, published on 18 December 2023, sets requirements for an AI management system. ISO/IEC 23894:2023, published in February 2023, gives guidance on managing AI risk. The register is where both meet daily work. Our AI governance framework for Saudi Arabia, the UAE and Jordan shows where it sits beside policy and decision rights.

AI risk register template: the 14 columns

Keep the register in a shared spreadsheet or your existing GRC tool. Do not build new software for it. Fourteen columns are enough:

We recommend a 3-point scale over a 5-point one. With three points there are fewer borderline calls to argue over, so reviews tend to move faster. The cost is less detail at the top. If several risks rate 9, rank them by hand.

Who owns each entry, and who owns the register?

Three roles keep the register alive. Name them in your AI policy, not only in the spreadsheet.

• Use owner. The manager who runs the process. They say what the AI does, which data it sees and when the use changes.
• Risk owner. The person who acts on a risk. For a data leak risk this might be the information security lead. For a wrong-answer risk it might be the customer service head.
• Register keeper. One person in risk, compliance or the AI team. They chase overdue reviews, check scores for consistency and report the top risks upward.

The register keeper does not own the risks. When the keeper ends up owning everything, business managers tend to stop reading the register. The keeper's job is to make owners answer.

New rows come from your AI inventory. If you do not yet know what staff use, start with our guide to finding and governing shadow AI, then add a row for each approved use.

A sample row from an AI risk register

This row is illustrative. Picture a mid-sized bank in Jordan that answers customer questions with an Arabic chatbot on its website.

Notice what the row does not say. It does not say "AI may hallucinate". It names the event, the effect and the test that shows whether the control works.

A second illustrative case: a family-owned retailer in Riyadh lets buyers draft supplier emails with Microsoft Copilot. The main risk is confidential pricing pasted into the wrong thread. The response is to mitigate with sensitivity labels and staff guidance, rated 4 before controls and 2 after. One row, two owners.

How often should you review it?

Use a fixed rhythm plus triggers. A fixed rhythm alone misses sudden changes, and triggers alone let quiet risks go stale.

1. Monthly, 30 minutes. The keeper checks overdue rows and any new uses from the inventory.
2. Quarterly, with owners. Rescore every row rated 6 or above, and close rows for uses that ended.
3. Twice a year, to leadership. Report the top risks, accepted risks and late controls to the executive committee or the board's risk committee.
4. Any time a trigger fires. Triggers include a model or vendor change, new data types, a new country of processing, an incident or complaint, or a new law or regulator rule.

Rules on personal data differ by country and sector. Saudi Arabia, the UAE and Jordan each have a personal data protection law, but scope differs. The UAE federal law does not apply in free zones with their own law, such as DIFC and ADGM, and leaves some sectors, such as health and banking data, to their own rules. Sector regulators can add requirements. This is general information, not legal advice; your obligations depend on your sector and data. Our PDPL and regulators page explains how we help teams prepare.

How to start your AI risk register this month

Begin small. Take the five AI uses that touch the most customer or personal data. Write one row for each with its owners, and book the first quarterly review now. Add the rest as the inventory grows.

If you want help, our AI Governance & Compliance Foundation (GB4) covers policy, decision rights, AI and risk registers, and PDPL alignment. It is part of the Governance Lab. To find a first step, book the AI Exposure Scan: a free 45-minute conversation that ends in a written next step.

Related lab: Governance Lab ↗