Most writing about AI risk management is for companies that build AI models. They test training data, measure drift and red-team their own systems. If your company buys ChatGPT, Copilot or an AI feature inside your HR software, most of that doesn't apply to you. What applies is simpler and more often skipped: a list of how you use AI, a judgment about what could go wrong with each use, and a person who answers for each one. That's an AI risk register, and you can start one this week.
What is AI risk management?
AI risk management is the routine of listing how your organization uses AI, judging for each use what could go wrong, how likely it is and how much harm it would do, choosing a response, and naming an owner who checks it again on a set date.
That's the same idea as any other risk management. The NIST AI Risk Management Framework, the main US reference, defines risk as "the composite measure of an event's probability of occurring and the magnitude or degree of the consequences." NIST released it on January 26, 2023, and says it's "intended for voluntary use."
This guide is for the people who end up owning AI risk at a company of 20 to 1,000 people: operations, IT, HR and finance leads. It covers the risks of using AI, not of building it. It isn't legal advice.
Why does a company that only uses AI need risk management?
Because the vendor manages the model's risks, not your use of it. What you put into a tool, what you let it decide and who checks its output are your risks, and the vendor can't see them.
NIST makes the same point. A company that deploys a model built by someone else "may not recognize that their particular uses could entail risks which differ from those perceived by the initial developer." The same assistant is low-risk when it drafts a newsletter and high-risk when it screens résumés or summarizes a patient's chart.
The incidents are already common. In OneTrust's 2026 AI-Ready Governance Report, a survey of 1,200 senior business decision-makers in eight countries published on September 14, 2026, 86% said they'd had at least one of the AI-related incidents the survey measured in the past year. Its examples include exposure of sensitive data or intellectual property, unapproved employee AI use, misinformation and data loss. A third (33%) had seen employees use unapproved AI, and 28% had two or more incidents in which AI systems or agents took unapproved actions. Only 27% slowed or paused AI deployment in response.
The lesson isn't to stop. It's that the risk is already in the building, and a written list is cheaper than finding it in an incident report.
What are the main types of AI risk?
For a company that uses AI, four groups cover most of what goes wrong: data that leaves where it should stay, wrong output that someone relies on, unfair or unaccountable decisions about people, and risks that arrive through vendors.
There's no single official list of four types. We use these four because each one has a different owner and a different fix. NIST's Generative AI Profile (NIST AI 600-1, July 2024) names 12 risks in more detail, and each one maps onto one of the four.
| Group | What it looks like at work | NIST Generative AI Profile risks it covers |
|---|---|---|
| Data | Client files pasted into a personal chatbot account; an assistant that can read every shared drive | Data Privacy, Intellectual Property, Information Security |
| Output | A made-up citation in a client memo; a wrong figure in a board pack; staff who stop checking | Confabulation, Information Integrity, Human-AI Configuration |
| People decisions | A screening tool that ranks applicants; AI-drafted performance reviews; a decision nobody can explain | Harmful Bias or Homogenization, Human-AI Configuration |
| Vendors | A tool that changes its data terms; an AI feature switched on inside software you already pay for; an agent that acts in other systems | Value Chain and Component Integration, Information Security |
Two of NIST's definitions are worth knowing by name. Confabulation is "the production of confidently stated but erroneous or false content," what most people call hallucination. Human-AI Configuration covers "automation bias, over-reliance," the habit of trusting output because a machine produced it. The first is the tool's fault. The second is a process failure, and it's the one your register can actually fix.
The profile's remaining risks (weapons information, violent or obscene content, environmental impact) mostly sit with the vendor. Note them once and move on, unless your use touches them.
If you want a longer list to check against, MIT's AI Risk Repository describes itself as "a living database of 1,700+ AI risks from 65 frameworks." Use it to check you haven't missed a category, not as a register.
AI risk register template: copy this
An AI risk register is a table with one row per use of AI and, for each, the risk, how likely and how bad it is, what you'll do about it, who owns it and when it's next reviewed. Paste this into a spreadsheet. One row per use, not one per tool: the same assistant used for drafting emails and for screening applicants carries two different risks.
| Field | What to write | Example |
|---|---|---|
| Use | The task, in plain words | Summarizing client contracts for the account team |
| Tool and plan | Product, vendor and the plan you're on | Microsoft 365 Copilot, company tenant |
| Data it touches | The most sensitive data in this use | Client contracts, including pricing and personal data |
| Risk group | Data, output, people decisions or vendors | Output |
| What could go wrong | One sentence, specific to this use | A summary misses a termination clause and the team relies on it |
| Likelihood (1 to 3) | 1 unlikely, 2 possible, 3 likely or already happened | 2 |
| Impact (1 to 3) | 1 minor and fixable, 2 costly or embarrassing, 3 legal, financial or harm to a person | 3 |
| Rating | Likelihood times impact | 6 |
| Response | Mitigate, transfer, avoid or accept, and the action | Mitigate: the account lead reads the termination and liability clauses in full before any summary is shared |
| Owner and backup | Named people, not teams | Head of client services; deputy |
| Off switch | How to stop this use quickly | Remove the contract library from Copilot's reach in the admin center |
| Next review | A date | February 15, 2027 |
Keep the scale small. A 1 to 3 scale is enough to sort a register for a company of 200 people, and it stops arguments about whether something is a 6 or a 7. NIST's framework doesn't prescribe a scale. It asks that the "likelihood and magnitude of each identified impact" be identified and documented (MAP 5.1), and that treatment be "prioritized based on impact, likelihood, and available resources" (MANAGE 1.2).
The register starts from a list of uses. If you don't have one, build that first: our AI inventory guide has a template whose rows feed straight into this one.
How do you rate an AI risk?
Rate each use on two questions, how likely the harm is and how bad it would be, using what you know about the use, past incidents and the data involved. Multiply the two, and act on the highest scores first.
Three rules keep the ratings honest.
Rate the use, not the tool. "ChatGPT" isn't a risk. "Account managers pasting client contracts into personal ChatGPT accounts" is. The data and the decision drive the rating far more than the product name.
Rate people decisions and sensitive data higher from the start. NIST says higher initial priority "may be called for" where a system uses "sensitive or protected data such as personally identifiable information, or where the outputs of the AI systems have direct or indirect impact on humans." In practice that's anything touching hiring, pay, discipline, health data or client confidences.
Use evidence where you have it. NIST lists the evidence for a rating: "expected use, past uses of AI systems in similar contexts, public incident reports," feedback from outside the team, and other data. If something has already gone wrong once, it's a 3 for likelihood.
Set your risk tolerance before you rate
Risk tolerance is how much risk you'll accept to get the benefit. NIST describes it as an organization's "readiness to bear the risk in order to achieve its objectives," and says the framework "does not prescribe" one. Decide yours in one line before the first rating, for example: "Anything rated 6 or more needs a written response and an owner above team lead. Anything rated 9 stops until it's fixed." NIST's own line for the extreme case is that where risks are unacceptable, use "should cease in a safe manner until risks can be sufficiently managed."
NIST also warns against the opposite mistake: "Attempting to eliminate negative risk entirely can be counterproductive." A register where every row is red gets ignored.
What are the four responses to an AI risk?
NIST names four: mitigating, transferring, avoiding or accepting the risk. For each high-rated row in the register, pick one, write down the action, and give it to the row's owner.
NIST lists them in MANAGE 1.3: "Risk response options can include mitigating, transferring, avoiding, or accepting." Here's what each looks like for a company that uses AI.
| Response | What it means | Example |
|---|---|---|
| Mitigate | Reduce the likelihood or the impact | Require a named reviewer to check every AI-drafted client deliverable; turn on the admin setting that keeps data out of training |
| Transfer | Move the cost to someone else by contract or insurance | Negotiate data-handling and indemnity terms with the vendor; check what your cyber policy says about AI tools |
| Avoid | Don't do it, or not with that data | No client names or health data in any AI tool outside the company tenant |
| Accept | Live with it, in writing, with a review date | AI-drafted internal meeting notes: low impact, accepted, reviewed quarterly |
Most rows end up as "mitigate", and most mitigations are a person checking output, a rule about data, or an admin setting. Our AI usage policy template turns the "avoid" decisions into rules staff can follow.
Who owns AI risk management?
One named person owns the register and its review dates, and each row has its own owner: the person who answers for that use, approves changes to it and can switch it off.
NIST's framework puts this under GOVERN: "Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear" (GOVERN 2.1). For a mid-size company, the register owner is usually in operations, IT or risk. Row owners are the people who run the work the AI is part of, such as the head of recruiting for a screening tool or the finance lead for AI-assisted reconciliations.
Decisions above your risk tolerance go to whoever approves AI uses. In many companies that's a small AI council that meets monthly.
A risk register only works if someone can stop the use
The weakest part of most risk registers is the column nobody fills in: what happens when the risk turns real. A rating of 9 on paper protects nothing if no one knows who can pull the tool, how, or who to tell.
NIST asks for exactly this. Organizations should be able "to supersede, disengage, or deactivate AI systems" that behave outside their intended use (MANAGE 2.4), and "incidents and errors are communicated to relevant AI actors" (MANAGE 4.3). For a company that buys its AI, that means an off switch per row (an admin toggle, a revoked license, a disconnected data source) and a short routine for when something goes wrong.
That routine is a guardrail in our framework: stop, tell, fix, learn. Stop the use, tell the people affected, fix the cause and change the register so it doesn't happen twice. The free lesson Stop, tell, fix, learn walks through it, and the lesson on owners and backups covers who holds each switch.
Which AI risk management framework should you use?
For most US companies of 20 to 1,000 people, the NIST AI Risk Management Framework, with its Generative AI Profile as a checklist of risks. It's free, voluntary and the main US reference for AI risk.
| Framework | What it is | Use it when |
|---|---|---|
| NIST AI RMF 1.0 (AI 100-1) | Voluntary US framework with four functions: govern, map, measure, manage | You need a recognized structure for your register and policy. Start here |
| NIST Generative AI Profile (AI 600-1) | 12 risks of generative AI with suggested actions | You use chat assistants, copilots or agents, which is almost everyone |
| ISO/IEC 42001 | A certifiable management-system standard for AI | A customer or contract asks for a certificate |
| MIT AI Risk Repository | A searchable database of AI risks drawn from other frameworks | You want to check your register hasn't missed a category |
You don't need to adopt all of NIST at once. Our guide to the NIST AI RMF shows which subcategories matter for a company that uses AI rather than builds it, and how to write a one-page profile. The ISO 42001 guide covers when certification is worth it.
Is AI risk management required by law?
Not as a general duty in the US. NIST's framework is voluntary. But some existing rules already require a risk assessment that your AI use falls inside. The clearest is HIPAA: covered entities and business associates must "conduct an accurate and thorough assessment of the potential risks and vulnerabilities to the confidentiality, integrity, and availability of electronic protected health information." An AI tool that touches that information is part of the assessment. Our AI compliance guide sorts the state and sector rules that apply to AI use by layer.
How often should you review AI risks?
Review the whole register on a fixed date, quarterly for most companies, and review a row early whenever something changes: a new use, a new vendor, a change in vendor terms, or an incident.
NIST asks for "ongoing monitoring and periodic review," with the frequency decided and roles defined (GOVERN 1.5). It leaves the frequency to you. Quarterly suits most companies of 20 to 1,000 people because vendors change plans, defaults and AI features often enough that a yearly review misses them.
These events should trigger an early review of the affected row:
- a new use of an existing tool, especially one that touches people decisions or sensitive data;
- a vendor changing its data, retention or training terms, or switching on a new AI feature;
- an agent getting access to a new system;
- an incident or a near miss, here or at a similar company;
- a new law or rule that reaches the use.
How to start AI risk management in 30 days
01 · Name the register owner and set the tolerance
Pick one person and a backup. Write your risk tolerance in one line: which score needs a written response and which score stops a use. This gives every later step its rules.
02 · List the uses
Start from your AI inventory, or build one: what staff use, what's switched on inside existing software and what's been bought. Split tools into uses. This gives you the rows.
03 · Rate each row and sort
Fill in the risk group, what could go wrong, likelihood and impact for each row. Rate people decisions and sensitive data first. Sort by rating. This tells you where to spend the first month.
04 · Choose a response and an owner for every high row
For each row above your tolerance, pick mitigate, transfer, avoid or accept, write the action and name the owner. Add the off switch. This turns the register from a list into a set of jobs.
05 · Set the review dates and the triggers
Put the quarterly review in the calendar and give each high row its own date. Tell row owners which events mean an early review. This keeps the register true after the first month.
If you'd rather have help building the register and the routine around it, that's what our AI governance work does, and it's part of deploying AI in a way that lasts.
What's still open
- NIST's framework is being revised. NIST's page says "The AI RMF 1.0 is being revised as part of the White House AI Action Plan," without a date. The four functions and the risk definitions we've used here may change in wording. We'll check each new version against this guide.
- Agent risks are moving fastest. Agents that act in other systems, make purchases or change records add risks the 2024 Generative AI Profile describes only in general terms. Expect more specific guidance, and rate agent uses conservatively until it arrives.
- State laws keep changing. New state rules on AI in employment decisions may turn parts of a voluntary register into required records.
We'll update this article when NIST publishes the revised framework or new guidance changes what a company that uses AI should record. The date at the top of the page shows when we last checked.
Sources
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, NIST, January 2023: the definition of risk, risk tolerance and prioritization, and the GOVERN, MAP and MANAGE subcategories quoted.
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, NIST, July 2024: the 12 generative AI risks and their definitions.
- AI Risk Management Framework, NIST, checked November 2026: release dates, voluntary use and the note that AI RMF 1.0 is being revised.
- 45 CFR 164.308, Administrative safeguards, Electronic Code of Federal Regulations, checked November 2026: HIPAA's required risk analysis.
- OneTrust research: 86% of organizations experienced AI-related incidents, yet few slowed deployment, OneTrust, September 2026: incident, unapproved AI and agent figures from the 2026 AI-Ready Governance Report (1,200 senior business decision-makers, eight countries, fielded by Sapio Research). A vendor survey.
- MIT AI Risk Repository, MIT FutureTech, checked November 2026: the size of the repository.