AI Programme Design Templates: From Decision to Evidence

Template one: programme decision charter

Record the decision the programme must inform, target work, participating population, current process, desired behavior, constraints, accountable owner, collaborators, exclusions, and review date. Add a “not trying to prove” field. For example, a small post-training exercise may test whether a simulation surfaces useful verification behaviors; it does not establish enterprise productivity or workplace transfer.

The CDC Program Evaluation Framework begins with context, programme description, focused questions and design, credible evidence, supported conclusions, and action on findings (CDC evaluation framework). This structure supports a coherent plan, while the organization must adapt it to its own domain and governance.

Sources: [1]

Template two: outcome and assumption map

Use one row per proposed pathway: activity, immediate output, near-term behavior, later outcome, critical assumption, external factor, evidence source, and disconfirming signal. Keep access, participation, use, capability, work quality, risk, and business outcome separate. If the programme assumes managers will provide protected practice time, name how that assumption will be checked.

Rank assumptions by uncertainty and consequence. Test the most decision-relevant uncertainty before expanding the programme. Do not convert a plausible pathway into a causal claim.

Template three: learning and work-practice blueprint

Describe authentic task, inputs, permitted tools, time, expected artifact, ambiguity, evidence trap, verification opportunity, feedback method, and transfer prompt. Add accessibility and alternative-format needs. Link each activity to the behavior it is meant to exercise. Remove content that has no clear relationship to the decision charter.

OPM’s planning and evaluation guidance notes that training needs analysis should ask what needs to be done and why it is not being done, and that training is not always the best or only solution (OPM planning and evaluating). The practical point is to distinguish skill gaps from unclear expectations, missing access, workflow friction, or incentives.

Sources: [2]

Worked pack: a verification practice pilot

Imagine a synthetic six-week pilot for analysts who use AI to draft research summaries. The decision charter asks whether a controlled exercise reveals verification gaps worth addressing in the next curriculum. The outcome map separates attendance from the ability to find an outdated claim. The blueprint uses a document set containing conflicting publication dates and requires a short evidence table.

The control card names approved tools, prohibited data, escalation routes, and reviewers. The evidence plan includes exercise artifacts, source-opening records when available, reviewer rationales, participant reflections, and a later manager conversation. Missing event capture lowers confidence. The decision rule says the curriculum owner may revise the source-checking module when repeated evidence shows an unclear method; it does not claim broader impact.

Template four: responsibility and control card

List programme owner, learning designer, functional reviewer, security or privacy reviewer, facilitator, evaluator, participant support, and decision approver. For each role, record responsibility, authority, input, output, and escalation condition. Add data classification, approved environment, retention, incident handling, and change-control fields.

Template five: evidence and interpretation plan

For each question, record construct, indicator, unit, source, collection window, denominator, missingness, quality check, reviewer, analysis, limitation, and intended decision. Prewrite an interpretation note with spaces for finding, counterevidence, confidence, alternative explanations, and next test. This makes overclaiming visible before results arrive.

Template six: learning register

After each cycle, record observation, evidence, interpretation, decision, owner, due date, and what would change the decision. Version every modified assumption, activity, rubric, or data definition. Preserve the prior version so later comparisons do not mix different programmes silently.

This pack is educational guidance. It does not prove programme effectiveness, create a validated assessment, determine legal compliance, or establish that AI caused an observed outcome.