13 Designing an AI-Supported Unit

Choose the evidence of learning before choosing the tool.

13.1 Start with a bounded pilot

Before redesigning a whole unit, test one activity with a colleague and, where appropriate, an ungraded student group. Check the source facts, rubric, access, feedback quality and effort. A successful prompt demonstration does not yet justify a high-stakes assessment.

The pilot pack provides an example. Use its P3 decision sheet to decide what should change before expansion. The route may remain non-AI if that better serves the students and outcome.

13.2 Backwards design

First name what students should be able to do. Then identify what evidence would demonstrate it, choose the assessment conditions and design practice that prepares them. AI can support some of that practice; it need not appear in every outcome or activity.

For a fictional third-year Workplace Conflict unit, possible outcomes are:

Distinguish accounts, documents and unresolved facts.
Plan a fair inquiry within role and authority.
Communicate a bounded next step respectfully.
Justify a recommendation using supplied policy and relevant disciplinary knowledge.
Evaluate feedback and apply the learning to a changed case.

These outcomes do not qualify students to conduct real legal investigations. Use appropriately reviewed sources and teaching where actual law is part of the curriculum.

13.3 An illustrative assessment design

A possible allocation is:

Practice analysis and revision, 25%: a response, selected evidence checks and an explained revision. Permitted help is declared.
Case portfolio, 40%: a bounded multi-source recommendation and evidence index, with appropriate individual contribution evidence.
Independent application, 35%: a supervised changed-case task and explanation under published conditions, with access adjustments.

These weights total 100%. They are a design example, not a prescribed split. The programme team should approve where independent assurance occurs and whether this design covers the intended outcomes.

The portfolio may be assisted, so its fluency is not treated as an unaided demonstration. The independent task should assess relevant understanding, not whether the student can reproduce an obscure fact from memory. If live communication is an outcome, include an observed interaction rather than relying solely on written work.

13.4 A 12-week teaching route

The following schedule is illustrative. It distinguishes preparation, practice, feedback and assurance rather than increasing AI complexity for its own sake.

Week	Learning focus	Student evidence
1	Sources, permissions, access and task boundaries	First-use decision and a short baseline
2	Facts, allegations and missing information	Annotated source cards
3	Questions and a fair first response	Paired or fixed-exchange practice
4	Apply the supplied policy; check feedback	Correction of P1-F with reasons
5	Revise and explain a decision	Response plus selected decision notes
6	First assessment and independent check	Changed-case explanation
7	Review errors; compare stakeholder accounts	Questions linked to missing evidence
8	Multi-source analysis	InnovateCo evidence index
9	Adaptation, access and ethical boundaries	Critique of a proposed process
10	Portfolio development and moderation	Recommendation with uncertainty
11	Portfolio submission and individual review	Contribution and reasoning evidence
12	Independent application and unit evaluation	New case and pilot decision

Critical verification starts in week 1, not after students have spent half a semester trusting generated feedback. The initial baseline and later tasks should be comparable enough to inform teaching, while acknowledging differences in assistance, task difficulty and marking.

13.4.1 A closer look at week 4

Students first judge P1-A against the supplied policy. Then they inspect the deliberately flawed feedback. They identify the invented warning, inappropriate approval and unnecessary disclosure request.

Only after documenting those checks do they write a revision. The teacher compares two explanations and discusses why one is better supported. A live model is optional. Asking students to retry until a model gives 100% would assess neither the source check nor independent judgement.

13.4.2 A closer look at week 8

Students use all four InnovateCo role cards and the documents. They identify what each source can support, formulate follow-up questions and recommend a bounded next step.

Do not ask the model to invent whatever evidence is needed to close the case. The supplied gaps are part of the task. A well-supported “not enough evidence” conclusion can be stronger than a confident finding.

13.5 Balance activities by purpose

There is no universal 60/40 balance between traditional and AI-supported teaching. Choose the mix according to the outcome, students, available support and assurance design.

A lecture or reading may introduce a necessary concept. Peer discussion can reveal differences in interpretation. A role-play can rehearse a question. An independent task can check transfer. AI may support some of these, but its presence is not itself a learning progression.

Make substantial activities possible without a model. For the planning task, the printed records and feedback are enough. For interaction skills, provide a structured peer/facilitator route. For AI-specific outcomes, provide suitable approved access rather than pretend the tool is optional.

13.6 Two genuine disciplinary adaptations

For supply chain, use S1’s quantity, deadline, capacity and budget. The learner must calculate costs and reject an option that fails a constraint. A suitable independent task changes capacity or deadline, so the student must recompute the decision.

For marketing, use M1’s campaign figures. The learner distinguishes measured rates from unsupported profit and causal claims. A changed task can supply costs but retain non-comparable audiences, testing whether the learner separates financial calculation from causal inference.

Do not simply replace employee names with customer names. Adapt the evidence, disciplinary criteria and independent check. A specialist should review added technical assumptions.

13.7 A unit design record

For each assessment, record:

Target outcome and prerequisite knowledge.
Supplied sources and their version.
What students do before receiving assistance.
Permitted help, data and approved routes.
Criteria, complete weights and any critical requirements.
What independent evidence is collected and under what conditions.
Access arrangements and a capability-matched fallback.
Staff/student workload and moderation plan.
What a successful and an unsuccessful response look like.
Review date and unresolved limitations.

Map the record to programme outcomes with the institutional alignment appendix. Do not assume a tool-supported activity automatically satisfies an accreditation requirement.

13.8 Explain the design to students

Tell students why they receive support during practice and why a later task has different conditions. Publish the distinction in the brief.

For example: “During rehearsal, use the supplied records, peer feedback or the approved AI route. In the independent task, apply the criteria yourself to a new record. We assess the decision and its basis, not the amount of chat.”

Clarify the process for questions, access adjustments and technical failures. An outage should not send students to an unapproved service or force a private account requirement at the last minute.

13.9 Evaluate cautiously

Compare work under sufficiently similar criteria, inspect persistent errors and record access and effort. Higher final marks may reflect assistance, changed assessment difficulty or marking changes, not necessarily learning.

A small pilot can identify problems and promising practices. It cannot establish that an entire approach works for all disciplines. Share the materials, assumptions and limitations along with any results.

13.10 Activity: map one assessment

Allow 20 minutes; no AI required. Complete the unit design record for one assessment. Test it against P1-A, P1-B and a response that correctly states insufficient evidence. Check whether the marks reward the intended capability.

Finish with one change you would make before launch and one question a colleague should challenge. Expand the design only after those questions have a satisfactory answer.

# Designing an AI-Supported Unit {#sec-unit-design} > Choose the evidence of learning before choosing the tool. ## Start with a bounded pilot Before redesigning a whole unit, test one activity with a colleague and, where appropriate, an ungraded student group. Check the source facts, rubric, access, feedback quality and effort. A successful prompt demonstration does not yet justify a high-stakes assessment. The [pilot pack](../appendices/pilot-pack.qmd) provides an example. Use its P3 decision sheet to decide what should change before expansion. The route may remain non-AI if that better serves the students and outcome. ## Backwards design First name what students should be able to do. Then identify what evidence would demonstrate it, choose the assessment conditions and design practice that prepares them. AI can support some of that practice; it need not appear in every outcome or activity. For a fictional third-year Workplace Conflict unit, possible outcomes are: 1. Distinguish accounts, documents and unresolved facts. 2. Plan a fair inquiry within role and authority. 3. Communicate a bounded next step respectfully. 4. Justify a recommendation using supplied policy and relevant disciplinary knowledge. 5. Evaluate feedback and apply the learning to a changed case. These outcomes do not qualify students to conduct real legal investigations. Use appropriately reviewed sources and teaching where actual law is part of the curriculum. ## An illustrative assessment design A possible allocation is: - **Practice analysis and revision, 25%:** a response, selected evidence checks and an explained revision. Permitted help is declared. - **Case portfolio, 40%:** a bounded multi-source recommendation and evidence index, with appropriate individual contribution evidence. - **Independent application, 35%:** a supervised changed-case task and explanation under published conditions, with access adjustments. These weights total 100%. They are a design example, not a prescribed split. The programme team should approve where independent assurance occurs and whether this design covers the intended outcomes. The portfolio may be assisted, so its fluency is not treated as an unaided demonstration. The independent task should assess relevant understanding, not whether the student can reproduce an obscure fact from memory. If live communication is an outcome, include an observed interaction rather than relying solely on written work. ## A 12-week teaching route The following schedule is illustrative. It distinguishes preparation, practice, feedback and assurance rather than increasing AI complexity for its own sake. | Week | Learning focus | Student evidence | |---|---|---| | 1 | Sources, permissions, access and task boundaries | First-use decision and a short baseline | | 2 | Facts, allegations and missing information | Annotated source cards | | 3 | Questions and a fair first response | Paired or fixed-exchange practice | | 4 | Apply the supplied policy; check feedback | Correction of P1-F with reasons | | 5 | Revise and explain a decision | Response plus selected decision notes | | 6 | First assessment and independent check | Changed-case explanation | | 7 | Review errors; compare stakeholder accounts | Questions linked to missing evidence | | 8 | Multi-source analysis | InnovateCo evidence index | | 9 | Adaptation, access and ethical boundaries | Critique of a proposed process | | 10 | Portfolio development and moderation | Recommendation with uncertainty | | 11 | Portfolio submission and individual review | Contribution and reasoning evidence | | 12 | Independent application and unit evaluation | New case and pilot decision | Critical verification starts in week 1, not after students have spent half a semester trusting generated feedback. The initial baseline and later tasks should be comparable enough to inform teaching, while acknowledging differences in assistance, task difficulty and marking. ### A closer look at week 4 Students first judge P1-A against the supplied policy. Then they inspect the deliberately flawed feedback. They identify the invented warning, inappropriate approval and unnecessary disclosure request. Only after documenting those checks do they write a revision. The teacher compares two explanations and discusses why one is better supported. A live model is optional. Asking students to retry until a model gives 100% would assess neither the source check nor independent judgement. ### A closer look at week 8 Students use all four InnovateCo role cards and the documents. They identify what each source can support, formulate follow-up questions and recommend a bounded next step. Do not ask the model to invent whatever evidence is needed to close the case. The supplied gaps are part of the task. A well-supported “not enough evidence” conclusion can be stronger than a confident finding. ## Balance activities by purpose There is no universal 60/40 balance between traditional and AI-supported teaching. Choose the mix according to the outcome, students, available support and assurance design. A lecture or reading may introduce a necessary concept. Peer discussion can reveal differences in interpretation. A role-play can rehearse a question. An independent task can check transfer. AI may support some of these, but its presence is not itself a learning progression. Make substantial activities possible without a model. For the planning task, the printed records and feedback are enough. For interaction skills, provide a structured peer/facilitator route. For AI-specific outcomes, provide suitable approved access rather than pretend the tool is optional. ## Two genuine disciplinary adaptations For supply chain, use S1's quantity, deadline, capacity and budget. The learner must calculate costs and reject an option that fails a constraint. A suitable independent task changes capacity or deadline, so the student must recompute the decision. For marketing, use M1's campaign figures. The learner distinguishes measured rates from unsupported profit and causal claims. A changed task can supply costs but retain non-comparable audiences, testing whether the learner separates financial calculation from causal inference. Do not simply replace employee names with customer names. Adapt the evidence, disciplinary criteria and independent check. A specialist should review added technical assumptions. ## A unit design record For each assessment, record: - Target outcome and prerequisite knowledge. - Supplied sources and their version. - What students do before receiving assistance. - Permitted help, data and approved routes. - Criteria, complete weights and any critical requirements. - What independent evidence is collected and under what conditions. - Access arrangements and a capability-matched fallback. - Staff/student workload and moderation plan. - What a successful and an unsuccessful response look like. - Review date and unresolved limitations. Map the record to programme outcomes with the [institutional alignment appendix](../appendices/institutional-alignment.qmd). Do not assume a tool-supported activity automatically satisfies an accreditation requirement. ## Explain the design to students Tell students why they receive support during practice and why a later task has different conditions. Publish the distinction in the brief. For example: “During rehearsal, use the supplied records, peer feedback or the approved AI route. In the independent task, apply the criteria yourself to a new record. We assess the decision and its basis, not the amount of chat.” Clarify the process for questions, access adjustments and technical failures. An outage should not send students to an unapproved service or force a private account requirement at the last minute. ## Evaluate cautiously Compare work under sufficiently similar criteria, inspect persistent errors and record access and effort. Higher final marks may reflect assistance, changed assessment difficulty or marking changes, not necessarily learning. A small pilot can identify problems and promising practices. It cannot establish that an entire approach works for all disciplines. Share the materials, assumptions and limitations along with any results. ## Activity: map one assessment **Allow 20 minutes; no AI required.** Complete the unit design record for one assessment. Test it against P1-A, P1-B and a response that correctly states insufficient evidence. Check whether the marks reward the intended capability. Finish with one change you would make before launch and one question a colleague should challenge. Expand the design only after those questions have a satisfactory answer.