What this playbook covers and when to run it
The playbook covers the standing work of governing AI once a team uses it, from the first register row to the policy's review date, and teaches none of the controls itself. The AI marketing pillar runs a team's first adoption plan from rules to a quarterly review; the AI governance hub sets out the risk tiers and the four records governance depends on, a register of uses, a version history, review logs and an incident list; and the AI troubleshooting collection diagnoses one failure at a time and sends recurring ones back to the register. Each phase below hands off to the page that owns the method.
The NIST AI Risk Management Framework gives the programme its shape: as the framework page applies them, its four functions, govern, map, measure and manage, ask who answers for each use, where AI runs and what might fail, how the output really performs, and how the team responds to what it finds. They are not a sequence completed once; this playbook is the rhythm that keeps asking them.
Run it when the policy exists but nobody can say when it was last checked against what the team does; when a feature inside a paid tool started generating without anyone deciding it should; when an agency uses AI on the brand's work in tools the business never evaluated; when an assistant answers customers and nobody reads a sample of what it said; or when the same kind of AI error has reached customers more than once. Like the other playbooks in the playbooks library, it is a starting shape: in a small team one or two people may hold every role, while a large one spreads them across operations, content, service and data protection.
Roles and responsibilities
- AI governance owner. Owns the programme: the register, the route for new uses, the review calendar, the incident list and the review date of the AI use policyA team's written rules for AI: approved tools, what each use may do, what data may go in, what is reviewed, when customers are told and who owns problems.. Does not own every use.
- Use owners. One named person per register row. Answers for the use's output, approves changes to its prompts, instructions, sources and settings, books its reviews and receives its incidents.
- Reviewers and editors. Carry out the human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. each tier requires and record sign-off with the version reviewed. Nobody reviews their own output above the internal tier.
- Data protection owner. The person responsible for data protection in the business decides what Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. may reach each tool and anything that follows when an incident or a customer question touches personal data.
- Agency and contractor leads. The business-side owner of each supplier relationship. Makes sure the policy reaches work done for the business, the supplier's material uses appear in the register and AI settings in the business's accounts are switched on only as agreed.
- Team leads. Confirm the register reflects what their people actually do, release time for review and training, and decide with the governance owner which pilots become standard.
A RACI matrix writes these roles down per phase, naming a single accountable person per use, as the NIST framework page suggests. As the NIST framework page puts it, a use without an owner is one whose failures nobody notices.
Cadence, inputs and outputs
Treat the rhythm below as a first draft. Adjust it to the number of uses and to how many reach customers without a person reading each output first, rather than to an ideal calendar.
| Cadence | Owner | Inputs | Outputs |
|---|---|---|---|
| Weekly | AI governance owner | New requests, features found switched on in existing tools, agency declarations | Register rows written before go-live, each with a tier, owner and reviewer; tests opened |
| Per use | Use owner, reviewers | The output, its tier, the source pack, the review checklist | Output reviewed at the tier's depth, sign-off recorded, findings logged by failure type |
| Monthly | AI governance owner, use owners | The incident list, review logs, changes to the register | Patterns named, open actions chased, prompt and source fixes assigned, uses moved between tiers |
| Quarterly | AI governance owner, use owners, agency leads | The whole register, stored test sets, agency rules, platform AI settings | Every row confirmed or changed, retests logged, supplier and platform settings matched to the rules |
| Annually | AI governance owner, team leads, data protection owner | The year's register, incidents, review logs and impact measures | A reviewed policy, a measured view of what AI changed, uses kept, widened or stopped |
Each cadence feeds the next: weekly intake keeps the register complete, the monthly read decides what the quarterly walk must examine, and the quarterly findings give the policy review its evidence. AI tools share the request route of the marketing stack management playbook, and personal data in prompts is part of the routine in the privacy and consent operations playbook.
The playbook in phases
Map every use in the register
This phase produces a register with a row for every use, recording its tier, owner, data, permissions, reviewer, disclosure, versions and incidents. The AI use register template lays out and explains the columns. Map uses rather than tools, since one drafting tool may carry several uses in different tiers, and include AI features that came bundled with tools the team already subscribes to, which the NIST framework page says are the uses most often missed, plus the material uses of agencies working for the business. The register is the framework's mapping function in practice; every later phase reads it and writes back to it. The gate: no use goes live, and no feature stays switched on, without a row and a named owner.
Set the rules and extend them to suppliers
Every row points to one short set of rules. The tutorial on writing a marketing AI use policy drafts the policy from the register and ends with a review date and review triggers. The guide to disclosure and data privacy sets the rules for telling customers and platforms and for what data may enter a prompt; it describes the European approach, from the Commission's page on its regulatory framework, as risk-based, with transparency expectations such as telling people when they are dealing with an AI system. Treat that as direction to watch, not as a list of obligations; nothing here is legal advice. The guide to setting AI rules for agencies and contractors attaches the policy to suppliers' work, including which AI features they may turn on in accounts the business owns. The gate: the policy names its owner and its review date, and every agency doing material work has been told which rules apply.
Take new uses and tools through one route
New uses and tools arrive through the same front door as any tool request. The guide to handling requests for new tools triages AI requests against the register as well as the stack. A request that goes to evaluation takes the extra questions in the guide to evaluating AI tools: how it performs on the team's real work, what the vendor's terms say about data and training, the admin controls, how model updates are handled and how to leave. Before real material goes in, the tutorial on minimising data before prompting prepares it, from the GDPR principle that keeps personal data to what the purpose needs. The guide to testing prompts and models builds the test set and pass bar the use will be retested against. The gate: a use goes live only when it has passed its test, its data is minimised, its tier is set and its row is complete.
Review output at the depth its tier sets
The tier set at intake decides the depth of review. The guide to accuracy and review workflows sizes review to the risk tiers, from no record for internal work to specialist sign-off for regulated claims, and logs findings by failure type; groundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. each draft in an approved source pack lets reviewers trace every claim. The tutorial on fact-checking an AI answer checks a single research answer before it becomes a decision, and the AI-assisted content review checklist is the gate every article, ad or email passes before publication. Claims and endorsements follow the guide to AI claims and consumer protection, which states the direction in the FTC's business guidance that deception is judged the same way whatever technology produced the claim. Every sign-off leaves an Audit TrailThe record, attached to the work item, of what was reviewed, which version, who reviewed it, when, what they decided and the evidence relied on.: who reviewed, which version and what they decided.
Keep prompts, instructions and source documents current
A test that passed holds only for one PromptThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told., with one model, settings and sources, at one moment. The prompt library tutorial keeps versioned templates with owners and review dates. The tutorial on writing standing instructions for an AI assistant versions the instructions and names an owner, and the tutorial on preparing a document library for an AI assistant keeps one current source per fact, with updates tied to policy and price changes. The tutorial on reviewing AI enhancements in ad accounts records each platform setting from the help centres' own pages and sets the triggers that send the team back to it. The decision point: any change to the prompt, the instructions, the sources or the underlying Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. triggers a retest against the stored test set before the use carries on.
Handle incidents and customer questions
When output made with AI reaches customers containing an error, the guide to handling an AI mistake runs five moves, from containing and correcting the error to telling those affected, tracing the cause and logging it, and expects each incident to change something in the system. Mistakes that span channels or need customer service and communications in step run inside the marketing incident management playbook, which keeps one error logOne list across every channel of the errors that reached customers: when found and contained, what customers saw, the checks that missed it, causes and actions. across channels. Before an assistant answers its first customer, the AI customer service launch checklist is the gate, and the guide to AI in customer service decides which kinds of contact the assistant handles alone, which it helps staff with and which stay with people. When a customer asks if a bot answered them or where their details went, the guide to answering customer questions about AI finds the true answer in the register and routes data rights requests to the privacy route. The gate: no reply may contradict a use the register lists.
Build skills and measure the effect
Controls hold only if people can apply them. The guide to building AI skills in a marketing team builds training around the team's real tasks, treats review as a skill to practise as deliberately as prompting and spreads good practice through champions and the shared library. The guide to measuring AI's impact on team output sets a baseline before a change and measures time per accepted output, defects and rework rather than hours people say they saved. The decision point: a pilot becomes standard work when its measured effect, review time included, justifies it, and is dropped when it does not.
Review the register and the policy on their dates
On the policy's review date, go through the register one row at a time, as the AI use policy tutorial recommends, and confirm that the tier, the owner and the data rule are still right, even when nothing looks different. Start from the incident list: every incident since the last review should have led to a changed rule, tier, check or tool. AI steps inside automated journeys stay in step with the automation governance guide, under the same owner in both inventories. Rights in generated work stay open questions: the guide to copyright and training data questions frames them for counsel and vendors and describes the provenance records to keep meanwhile. The gate: a dated review recorded in the policy, with every row confirmed or changed.
The register as the programme's spine: what each review reads and changes
Every phase reads and writes the register; the version history, review logs and incident list hang off its rows. Each review reads it before starting and writes back before ending.
| Review | What it reads | What it may change | Who decides |
|---|---|---|---|
| Weekly intake | New requests, features switched on in existing tools, agency declarations | A new row with tier, owner, reviewer and data rule; a test opened; a request declined | AI governance owner, with data protection for personal data |
| Per use review | The row's tier, reviewer and source pack | Review logs; a request to move the use up a tier | The reviewer, then the use owner |
| After an incident | The row's versions, permissions and incident history | A rule, tier, check, prompt, source or tool | Use owner, with the AI governance owner |
| Monthly read | Incident list, review logs, register changes | Tiers moved, fixes assigned, overdue actions chased | AI governance owner |
| Quarterly walk | Every row, stored test sets, agency rules, enhancement settings | Owners replaced, retests logged, settings switched off, uses stopped | AI governance owner, use owners, agency leads |
| Annual policy review | The year's register, incidents and impact measures | Policy clauses, tiers and review depths; pilots made standard or dropped | AI governance owner and team leads, with data protection |
Two rules keep the register honest. A use missing from the register is treated as not approved, which follows the policy tutorial's habit of writing the row before anyone starts using a new feature. And the register lives where customer-facing staff can read it, because every answer to a customer's question about AI depends on it.
A holiday letting agency's first quarter of AI governance
Vuldorex Lettings, an invented holiday cottage letting agency with a marketing team of 3 and an agency writing its property listings, starts the playbook after a guest asks whether the website chat is a person and nobody can say for certain. The first register finds 9 uses: listing copy drafted in the agency's own tool, subject lines suggested by the email platform, assets created automatically in its ad accounts, the guest chat assistant and 5 internal drafting and summary uses. Only 4 had been thought of as AI. The marketing lead becomes governance owner and names an owner per row. In week 3 the agency's terms attach the policy, list the data it may use and require a declaration with each batch of listings. Every listing goes through the content review checklist, and the assistant's owner reads 30 conversations a week. In week 7 the chat vendor announces a model change; the stored set of 40 past guest questions is rerun, and 3 answers about pets now quote a superseded page, which is archived and retested before the assistant carries on. In week 10 a generated ad asset promises a sea view that 2 cottages lack; it is pulled the same day, logged as the quarter's 1 incident, and the rule changes so generated ad assets are reviewed before they run.
Measuring the playbook
The playbook works when every use is known, reviewed at the right depth and changed when review or an incident shows it should be. Operating signals show whether the rhythm holds: new uses entering the register before go-live, reviews and sampling done on their dates, a logged retest after every model or prompt change, and the monthly read and quarterly walk held. The best single test is how many uses each quarterly walk finds that the register did not know about; it should shrink over time.
Outcome signals reuse the measures the governance guides name: how many uses have a named owner today, how many were reviewed in the latest cycle, and how many incidents remain open and for how long, as the policy tutorial suggests; defects per reviewed item by tier, from the review workflows guide; and the proportion of incidents closed with a change to the system, with repeats from the same cause, from the guide to handling an AI mistake. Keep the number of incidents out of any target, since more reports often mean closer attention rather than worse practice. Put these signals into the reporting the team already produces each month and quarter, next to the measured impact, so governance sits beside the results AI is meant to help.
Common mistakes
- A policy without a programme. A signed policy nobody reviews drifts away from the team's tools. Give it an owner, a review date and a register that changes every week.
- Registering tools instead of uses. One tool can carry an internal summary and a customer-facing reply. Write one row per use, so each gets its own tier and reviewer.
- Missing the features you already pay for. Built-in drafting, generated ad assets and suppliers' tools are the uses most often left out. Look for them at every quarterly walk.
- The same review for everything. Full review of internal notes buries reviewers, and light review of customer-facing output lets errors through. Let the tier set the depth.
- No retest after a change. A model update or prompt edit can change output without warning. Rerun the stored test set before the use carries on.
- Incidents that change nothing. An incident closed without a changed rule, tier, check or tool will return. Make the change part of closing it.
