Turn your best repeatable service into an AI-assisted delivery playbook

Choose a service with a stable deliverable

Start with a service you have completed enough times to recognize good work: a landing-page review, campaign research brief, or monthly content analysis. Avoid beginning with your most ambiguous strategic engagement. You need real examples, known inputs, and a reviewer who can recognize an unacceptable result.

Write the deliverable in concrete terms. For a hypothetical landing-page review, the output might be a prioritized set of findings, annotated evidence, and a proposed test for each important issue. “Improve conversions with AI” is a promise, not a delivery specification.

Extract the decisions behind your best work

Choose a few completed examples you are permitted to reuse, removing client identifiers and sensitive information. Compare them. Which questions did you ask? What evidence changed your recommendation? Which findings did you reject? A playbook based only on polished final documents misses the judgment that made them useful.

Build a short sequence of stages: understand the brief, inspect evidence, develop findings, prioritize, and review. For each stage, specify the input, desired output, responsible person, and conditions that require clarification. Keep the steps connected to the client problem rather than documenting every click.

Assign a useful role to AI at each stage

AI can organize notes, compare a draft against a brief, suggest alternative explanations, and identify missing evidence. It should not invent interviews, analytics, or client approvals. Give it the material it is allowed to use and ask it to label every inference.

For the landing-page example, the assistant can group observed usability issues and draft a finding structure. A practitioner checks the page, validates the evidence, evaluates severity, and decides what matters. If the task depends on live functionality, a screenshot alone will not establish how the experience behaves.

Prompt to try

Convert these redacted examples and working notes into a draft service playbook: [materials]. For each stage, define required inputs, tasks AI can assist, human decisions, expected output, and a review standard. Identify gaps instead of filling them with invented process. Separate client-specific choices from reusable methods. Keep the playbook focused on this deliverable: [deliverable].

Write a quality rubric people can use

Use a few observable criteria. A strong finding should identify the affected user, the relevant evidence, the consequence, and a proportionate next action. Define what passes and what needs revision. Avoid criteria such as “excellent” or “strategic” without examples.

Add examples of a weak and a strong deliverable, clearly labeled as fabricated if you create them for training. For instance, “the page needs better copy” is too vague. “The primary button promises a trial, but the next screen requests a sales meeting; clarify the commitment before the click” describes an observable problem and an actionable response.

• Traceability: each finding links to supplied or directly inspected evidence.
• Relevance: the recommendation addresses the agreed customer and business problem.
• Specificity: the next action is concrete enough for an owner to implement.
• Restraint: unknowns remain unknown, and estimated impact is labeled.

Run a reviewer comparison before expanding

Have a practitioner produce or review a small set of deliverables using the playbook. Include an ordinary case, an incomplete brief, and a case where the evidence contradicts the expected story. If possible, review versions without identifying which method produced them.

Record accepted findings, important omissions, correction time, and client-specific adjustments. A faster first draft is not an improvement when senior review takes twice as long. Treat this as an internal operating check, not a statistically conclusive proof of universal performance.

Prompt to try

Review this deliverable against our brief, evidence, and quality rubric: [materials]. For each criterion, return pass, revise, or cannot assess, with the exact evidence. Identify unsupported claims, missing customer context, and recommendations that exceed the agreed scope. Suggest the smallest changes needed for approval. Do not lower the standard to make the draft pass.

Keep the playbook useful after the first version

Give the playbook an owner and a version date. Store the input template, prompts, examples, rubric, and exception notes together. When a reviewer repeatedly corrects the same issue, improve the relevant stage rather than adding another vague instruction to the final prompt.

Keep a record of model or tool changes that affect the process and rerun representative examples after meaningful changes. NIST’s framework treats evaluation and management as continuing activities. For an agency, the practical result is a maintained service method with visible accountability, rather than a prompt collection no one owns.

Commercially, the playbook should support clearer scope and dependable delivery. It does not guarantee a margin increase or justify promising clients an outcome the evidence cannot support. Measure the complete delivery effort and use the recovered capacity deliberately.

Try it on one real task

1. Select one repeatable deliverable and gather three permitted examples.
2. Document the five stages and identify one useful AI task per stage.
3. Test the rubric on a draft with a deliberately unsupported finding.

Check before you use the result

• The service has a concrete definition of done.
• AI assistance and human decisions are assigned explicitly.
• Examples are permitted, redacted, and labeled where fictional.
• The playbook has an owner, version, and maintenance process.

Take the next step

Download the Waymaker Agency AI practice pack for these prompts, exercises, and review checklists.

New to AI? Take the free first-workflow course at AI Teachers. Already using it? Practice building a repeatable workflow.

Sources and further reading

• NIST: AI Risk Management Framework
• NIST: Generative AI Risk Management Profile

Sources support the principles discussed; the examples and suggested workflows are original teaching exercises.