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.
| Consequence | Example | Minimum operating control |
|---|---|---|
| Low | Internal summary draft | User review, source links, correction path |
| Moderate | Campaign brief classification | Sampling, confidence threshold, exception queue, owner review |
| High | Customer-facing regulated copy | Mandatory qualified approval, version record, policy checks, rollback |
| Very high | Eligibility or sensitive audience decision | Specialist 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
| Decision | Use when | Next move |
|---|---|---|
| Go to controlled pilot | Owner, inputs, controls, evaluation, and stop conditions are clear | Test with limited scope and named reviewers |
| Prepare first | The use case is valuable but one or more operating conditions are weak | Fix the specific data, workflow, or governance gap |
| Do not automate yet | Failure is hard to detect, decision rights are unclear, or consequences are unacceptable | Redesign the work or retain a human-only process |