A practical framework for AI-enabled work in education

Define the educational decision

Name the setting, learners, objective, stakes, and accountable educator. Distinguish brainstorming, explanation, feedback, grading support, student services, and institutional administration because they require different evidence and controls.

UNESCO's guidance for generative AI in education and research emphasizes a human-centred approach, capacity building, privacy, and validation of pedagogical and ethical suitability. Those themes support a simple operating question: what educational purpose does this use serve, and what evidence would show that it remains suitable in this context?

Write a narrow use statement. “Use AI for personalized learning” is too broad. “Generate two optional explanations of a worked statistics example for adult learners, using the approved course source and preserving the instructor's notation” can be inspected.

Sources: [1]

Build a learning-purpose evidence card

The card should include the learning objective, learner group, approved sources, data allowed, AI task, prohibited task, educator review, accessibility check, disclosure, retention, and evidence of the learner's own work. Add a stop condition for unsupported claims, sensitive data, inaccessible output, or a task where automation would remove the intended learning practice.

The U.S. Department of Education report on AI and the future of teaching and learning frames education AI around centering people, protecting human agency, and developing guidelines and guardrails with evidence. The report does not approve a particular product. Use it to structure questions, then test the actual tool, plan, settings, and learner context.

For every output, preserve the source and mark what the system generated, what the educator changed, and what the learner must still do. Do not treat fluent text as evidence that a learner understood the concept.

Sources: [2]

Work a synthetic course-design example

Imagine an adult-learning team preparing a module on interpreting confidence intervals. The approved source pack contains the instructor guide, three worked examples, a notation standard, and accessibility requirements. The proposed AI use is to generate alternative explanations and a short practice prompt.

The first generated explanation says that a 95 percent confidence interval means there is a 95 percent probability that the fixed parameter lies inside this particular interval. The instructor flags the explanation as misleading, links the approved wording, and narrows the generation task. The second version uses the course definition but introduces unfamiliar notation. The third version preserves the approved notation and includes an example with different numbers from the graded exercise.

The evidence card records:

The final artifact includes the reviewed explanation, change log, accessibility check, and a separate learner task. It does not claim that the generated explanation improves learning outcomes.

• purpose: optional explanation before independent practice;
• source boundary: approved course guide only;
• educator action: verify the statistical interpretation and notation;
• learner action: solve a new case and explain the conclusion in their own words;
• accessibility: equations have readable text alternatives and the activity does not depend on color;
• disclosure: learners are told which explanation was AI-assisted;
• stop condition: withhold the material if the source cannot support the explanation.

Test the workflow, not only the answer

Test a correct source, a stale source, a source that conflicts with local notation, a learner request containing personal information, and a prompt that asks the system to complete the assessed task. Verify how the workflow responds and whether a person can correct the output before it reaches learners.

Check different access conditions. A tool that works for a staff member may be unavailable to learners, require an unsupported device, or produce an inaccessible format. Record the actual product tier, account permissions, language, and integration rather than assuming capabilities.

Separate formative support from consequential decisions. Automated suggestions may help an educator notice patterns, but the organization should define who interprets them, what evidence is visible, and how errors can be challenged. Qualified reviewers determine applicable education, privacy, accessibility, labor, and procurement requirements.

Evaluate educational fit

Review the card against four questions: Did the use preserve the learning objective? Can decisive claims be traced to sources? Is educator or learner agency visible? Can errors, access failures, and uncertainty be corrected without hidden consequences?

Use fit, fit with conditions, redesign, or do not use for this task. Revisit the decision when the course, learner group, source set, model, or policy changes.

Two Prune can create a synthetic education-related work simulation and retain the submitted artifact plus available workflow evidence for human interpretation across its six AI-fluency pillars. It does not grade students, approve educational technology, certify compliance, or establish learning impact. Assessment evidence remains task-bounded and should include confidence and missing-evidence limitations.