Checklist · AI governance

Is your Marketing Operations model ready for governed AI?

Review ownership, data, workflow, evaluation, privacy, and human-review conditions before introducing AI into marketing work.

10-minute checklist

Start with the work, not the model

AI readiness begins with a clearly defined operating problem. A team is not ready because it has access to a model or because a vendor has added an AI feature. It is ready when the work has an owner, the expected outcome is observable, the inputs are understood, and failure can be detected before harm spreads.

Use this checklist for one use case at a time. The controls required for summarizing an internal status report are different from those required for generating regulated customer communications or making audience decisions.

Use case, value, and decision ownership

  • Problem statement

    The use case names a specific task, delay, quality issue, or decision it is intended to improve.

  • Named owner

    One accountable business owner can approve the use case, controls, and operating changes.

  • Human decision boundary

    The team knows which decisions remain human and where AI may recommend, draft, classify, or summarize.

  • Value hypothesis

    Expected value is expressed as time, quality, risk reduction, capacity, or a measurable business outcome.

  • Baseline

    The current effort, defect rate, cycle time, or other comparison point is known well enough to evaluate the pilot.

Data and knowledge readiness

  • Approved sources

    The model receives information from named, permitted, and sufficiently current sources.

  • Data classification

    Personal, confidential, regulated, and licensed information is identified before it is supplied to a model.

  • Knowledge ownership

    Someone is accountable for updating the standards, examples, policies, and reference material used by the workflow.

  • Provenance

    Users can identify which source or rule supports a material recommendation.

  • Retention

    The team understands what the provider stores, for how long, and whether submitted data may be used for training.

Workflow and human-review controls

  • Entry criteria

    Only work that meets defined input requirements enters the AI-assisted path.

  • Review role

    The reviewer has the subject knowledge, time, and authority to reject or correct the output.

  • Exception path

    Uncertain, sensitive, or incomplete cases move to a clear manual queue.

  • Change record

    Material prompts, rules, models, and approval criteria are versioned.

  • Fallback

    The team can continue critical work safely when the AI service is unavailable.

Suggested control intensity by use-case consequence
ConsequenceExampleMinimum operating control
LowInternal summary draftUser review, source links, correction path
ModerateCampaign brief classificationSampling, confidence threshold, exception queue, owner review
HighCustomer-facing regulated copyMandatory qualified approval, version record, policy checks, rollback
Very highEligibility or sensitive audience decisionSpecialist risk assessment; automation may be inappropriate

Evaluation, monitoring, and stop conditions

  • Define acceptable performance

    Use a representative evaluation set and measures connected to the task: factual accuracy, classification precision, policy adherence, correction effort, or reviewer agreement.

  • Monitor the operating outcome

    Track whether cycle time, quality, capacity, or risk actually improves. Faster generation with more review and correction is not productivity.

  • Name stop conditions

    Pause or revert when critical errors exceed the agreed threshold, source quality declines, the workflow changes, a provider changes material behavior, or reviewers cannot keep up.

  • Reapprove meaningful changes

    A new model, data source, customer context, or decision boundary can alter risk. Treat material change as a new operating decision.

Make a clear go, pilot, or stop decision

AI readiness decision guide
DecisionUse whenNext move
Go to controlled pilotOwner, inputs, controls, evaluation, and stop conditions are clearTest with limited scope and named reviewers
Prepare firstThe use case is valuable but one or more operating conditions are weakFix the specific data, workflow, or governance gap
Do not automate yetFailure is hard to detect, decision rights are unclear, or consequences are unacceptableRedesign the work or retain a human-only process