What the NIST AI Risk Management Framework is
Marketing teams now use AI in many places at once: drafting copy, suggesting subject lines, answering customers, summarising data and, increasingly, running agents that act inside workflows. Each use carries its own risks: a Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. that states a wrong fact fluently, a draft that drifts from the Brand VoiceThe consistent personality a brand expresses in its words across channels, with tone adjusting to the situation and the audience., Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. pasted into a tool whose terms allow training on it, an assistant that answers a customer from an out-of-date document. The question is how to manage all of them without a separate rulebook for each.
NIST's framework answers with a structure rather than a list of controls. It describes four functions. As the AI governance section puts them for a marketing team: govern (who is accountable), map (where AI is used and what could go wrong), measure (how it actually behaves) and manage (what you do about it). The same section notes that the vocabulary scales down well, which is why the site's guides use it. Among the qualities the guides draw from it are two that matter most in marketing: the validity and reliability of AI output, which is something to be measured rather than assumed, and responding to incidents, which is part of managing AI risk.
The framework does not tell a team which tools to use, which risks to accept or what its policy should say. It gives an order in which to think, and a way to check that no part of the work has been skipped.
The four functions in a marketing team
Each function asks one question. The table shows what that question covers in marketing work and where the site already treats it in depth.
| Function | The question it asks | Marketing work it covers | Where the site covers it |
|---|---|---|---|
| Govern | Who is accountable, and by what rules? | An owner for each use, an AI use policy, approved tools, review responsibilities, a review date | The AI use policy tutorial and the governance section |
| Map | Where is AI used, and what could go wrong? | Every use listed as a row, including features built into existing tools; the data each sees; the risk tier of each | The AI use register template and the policy tutorial's first two steps |
| Measure | How does it actually behave? | Trials on the team's own tasks before adoption; fact and claim checks; sampling of live output; logs of what review found | Tool evaluation, and accuracy and review workflows |
| Manage | What is done about what is found? | Review tiers sized to risk; fixes to prompts, sources and templates; pausing a use; responding to incidents | Review workflows and the guide to handling an AI mistake |
The functions are not a sequence to complete once. Mapping a new use triggers governing it (an owner and a tier), measuring it (a trial and then sampling) and managing it (the review it needs). What measuring finds feeds managing, and what managing changes is recorded where governing can see it.
Govern. The tutorial on writing a marketing AI use policy uses the four functions as the frame for the whole document: the register maps, the tiers and review rules manage, the incident and review clauses measure, and the owners govern. Accountability is the core of governing, and the RACI matrix is the tool for making it explicit, with exactly one accountable person for each use.
Map. The AI use register template is the map in practice: one row per use, with its tier, owner, tool, the data it sees and what it is allowed to do. The policy tutorial notes that the NIST framework's mapping function starts from knowing where AI is used, which is why the register comes first.
Measure. The guide to evaluating AI tools uses the framework's structure to check that no question was skipped before a tool is adopted, and measures performance on the team's own tasks rather than the vendor's demonstration. After adoption, the guide to accuracy and review workflows describes its habits as measuring and managing: observe how the system actually behaves, then act on it. GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. drafts in an approved source pack makes that measurement easier, because each claim can be traced to a source.
Manage. The guide to handling an AI mistake that reached customers covers the moment managing matters most: contain, correct, tell, trace and log, feeding what is learned back into how the system is governed and measured. The five whys helps with the trace step, following one failure back to a process the team can change.
Applying it at a small team's scale
A marketing team of a few people does not need a risk committee. It needs the four questions answered for each use, in the smallest documents that will actually be kept up to date.
Start by mapping. List every use, including the AI features that arrived inside tools the team already pays for, because those are the ones most often missed. For each, record what data it sees and where its output goes. A use whose output reaches customers without a person reading it first is a different risk from a draft that an editor rewrites.
Give every row an owner. The owner does not have to do the work, but answers for it: approving changes, booking the review, receiving incidents. A use without an owner is a use nobody will notice failing.
Size the measuring to the risk. Internal brainstorming may need no more than the policy's general rules; published claims need a fact check; customer-facing automation needs regular sampling of what it actually said. The comparison of rule-based automation and AI agents applies the four functions to agents in workflows: who is accountable, where the agent acts and what could go wrong, how its behaviour is measured, and what happens when it misbehaves.
Decide the incident response before the first incident. Who is told, who can pause a use and where the record goes should already be written down when something reaches a customer, not worked out on the day.
Then review on a date. A policy and register that were right when written drift as tools change their models and the team finds new uses.
Worked example: a recruitment agency governs its AI drafting (illustrative)
A recruitment agency governs its AI drafting
Grumvexel Recruitment, an invented agency with a marketing team of 3, drafts job adverts and candidate emails with an AI assistant, about 40 adverts and 600 emails a month.
Govern. The marketing lead is named accountable for both uses, and the policy names the operations manager as the person who can pause either.
Map. The register gets 4 rows: advert drafts, candidate email drafts, subject-line suggestions in the email tool, and a summarising feature in the CV database that nobody had listed. The last one sees candidates' personal details, so it moves to the highest tier and goes to specialist review before further use.
Measure. Each week a recruiter samples 10 adverts and 30 emails against a short checklist: wrong facts about the role, unsupported claims about the client and wording that could put some candidates off unfairly. In the first month the sample finds 6 adverts with salary ranges that did not match the brief.
Manage. Salary now comes only from the brief's source fields, and the advert template forces a check. When a wrong interview date later reaches 2 candidates, the team corrects it with each candidate the same day, traces the cause to an old template and logs the fix in the register.
Limits and criticisms
The framework is voluntary and general. It is written for organisations of every kind, so it offers a structure rather than marketing controls. It will not say which review a product page needs or what a chatbot may promise; the team still has to decide those things.
It does not say which risks to accept. Mapping and measuring show what could go wrong and how often it does, but deciding that a risk is worth taking for a given benefit remains the business's judgement.
Following it does not show compliance with any law. It is a way of organising risk management, and a team that has an owner, a register and a review routine may still fall short of a regulation in one of its markets. This page gives no legal or compliance advice.
A small team can turn it into paperwork nobody reads. Four functions can become four documents, each written once and never opened. The test of the structure is whether the owner books the review and the register changes when the tools do.
And the site describes it only at the level of the four functions and the qualities its guides cite. Anyone applying the framework beyond that should read NIST's own material, linked in this page's sources, rather than rely on a marketing summary.
Related frameworks
The RACI matrix turns "govern" into named roles: for each use of AI, who does the work, who is accountable, who is consulted before changes and who is told when something goes wrong. The five whys serves "manage" when a single AI mistake needs tracing to the step in the chain that let it through, so the fix lands on the process rather than the person. The marketing frameworks library lists the other planning frameworks by the decision they serve.
Common mistakes
- Starting with the policy instead of the map. A policy written before the uses are listed governs the uses someone remembered.
- Missing built-in features. AI inside existing tools is used daily and listed rarely.
- An owner for the tool, not the use. The same tool can carry a low-risk use and a high-risk one.
- Measuring once at adoption. Models and prompts change; sample live output on a schedule.
- No incident route until an incident. Decide who is told and who can pause a use before it is needed.
- Treating the framework as compliance. It organises risk management; legal questions need specialist advice.
- Four documents nobody opens. Keep the records small enough that the owner actually maintains them.
