15 Implementation Practicalities

Plan the fallback before the activity needs it.

15.1 Check readiness

Before introducing a student-facing activity, confirm the capability, criteria, permitted data, approved tool or non-AI route, support arrangements and expected time. Test the material against contrasting responses, including a plausible but unsafe answer.

A prompt that worked once in your office is not evidence that it will behave consistently for a cohort. Keep fixed source cards and answer checks so that the activity survives a failed model response.

The data gate should appear in the student instructions, not only in the lecturer’s planning notes.

15.2 A capability-matched fallback

For the pilot’s planning-and-justification task, students can use the printed records and write or explain a response. For an outcome requiring live communication, offer an appropriate peer/facilitator interaction with access adjustments. A written case analysis alone cannot demonstrate every aspect of a spoken consultation.

Do not switch confidential material to an unapproved platform during an outage. Use a previously approved alternative or the offline route. If neither supports the assessment conditions, pause and apply the institution’s extension or alternative-assessment process.

Problem	Bounded response
Tool unavailable	Use approved fallback or pause; do not improvise a new data destination
Persona invents facts	Discard invention; return to source cards
Feedback conflicts with criteria	Check against records; seek educator review
Save/export fails	Use local notes or printed response sheet; avoid unnecessary screenshots
Student needs an adjustment	Arrange equivalent capability evidence through the appropriate process
Sensitive information entered	Stop further sharing and follow incident-reporting procedures

15.3 Access is more than account availability

Check mobile and desktop use, keyboard navigation, assistive technology, language support and the time needed to complete the task. A free service can still have limits that make an assessment unreliable.

Do not require a personal account or paid upgrade when an equivalent route can meet the outcome. Where AI use itself is assessed, provide suitable approved access and explain the requirement in advance.

Offer clear instructions, a short demonstration and a fixed worked example. Students should not have to discover the task’s hidden expectations by repeatedly asking a chatbot what the lecturer wants.

For a spoken response, assess the relevant reasoning or communication outcome, not eye contact, accent, speed or confidence unless a particular feature is genuinely part of the stated professional capability and assessed fairly.

15.4 Student wellbeing

Give advance notice of complaint, discrimination or performance scenarios. Use fictional data, a stop instruction and an agreed neutral alternative. Do not ask students to disclose their own experiences to make the activity authentic.

A role-player may disagree, pause or ask a question. The exercise should not escalate emotional pressure merely to make the learner “work harder”. Debrief the decisions and evidence, not whether the student persuaded the persona to become agreeable.

15.5 Budget the whole task

Record design, testing, support, moderation and marking time separately. An initial prompt takes little time; a defensible teaching activity takes more.

For 24 students, eight minutes of review per submission is 192 minutes. A 30-minute moderation discussion brings that illustrative marking budget to 222 minutes, or 3 hours 42 minutes. Setup and student support are additional. Replace these assumptions with observed times after a pilot.

For 100 students, a five-minute individual check takes 8 hours 20 minutes before administration. A sustainable design may distribute checks across tutorials, use several staff or choose another appropriate assurance format.

Do not offset those costs with an invented claim that AI always saves a fixed number of minutes per learner. Measure the actual task.

15.6 Keep an approved activity record

Save the scenario version, student brief, rubric, tested examples, permitted routes, access arrangements and known limitations. Record material changes between cohorts.

For model-supported use, note the product/account and relevant features tested. A later tool change may require rechecking the activity. Do not assume a well-worded prompt remains equally reliable across every model.

Keep student evidence only for the authorised purpose and retention period. An audit trail does not require retaining every private conversation indefinitely.

15.7 Evaluate the pilot

Use P3 in the pilot pack. Compare the baseline and independent response cautiously, examine errors and record obstacles.

Separate satisfaction, task completion, output quality and learning evidence. Students may enjoy an activity that does not improve transfer, or find a challenging activity useful despite lower satisfaction. Neither result should be simplified into a universal endorsement.

A reasonable decision can be to continue, revise, pause or stop. If a model repeatedly invents case facts, use the fixed route while redesigning. If a particular group cannot access the task, fix the access problem before increasing its stakes.

15.8 Activity: rehearse failure

Allow ten minutes; no AI required. Imagine the service fails halfway through your activity. Write what students do next, which evidence remains assessable, how they save their work and who approves a changed assessment condition.

Then try the fallback yourself. Check that it preserves the capability, does not send data somewhere new and does not penalise students for the outage. That rehearsal belongs before the first graded use.

# Implementation Practicalities {#sec-implementation} > Plan the fallback before the activity needs it. ## Check readiness Before introducing a student-facing activity, confirm the capability, criteria, permitted data, approved tool or non-AI route, support arrangements and expected time. Test the material against contrasting responses, including a plausible but unsafe answer. A prompt that worked once in your office is not evidence that it will behave consistently for a cohort. Keep fixed source cards and answer checks so that the activity survives a failed model response. The [data gate](ethics-integrity.qmd) should appear in the student instructions, not only in the lecturer's planning notes. ## A capability-matched fallback For the pilot's planning-and-justification task, students can use the printed records and write or explain a response. For an outcome requiring live communication, offer an appropriate peer/facilitator interaction with access adjustments. A written case analysis alone cannot demonstrate every aspect of a spoken consultation. Do not switch confidential material to an unapproved platform during an outage. Use a previously approved alternative or the offline route. If neither supports the assessment conditions, pause and apply the institution's extension or alternative-assessment process. | Problem | Bounded response | |---|---| | Tool unavailable | Use approved fallback or pause; do not improvise a new data destination | | Persona invents facts | Discard invention; return to source cards | | Feedback conflicts with criteria | Check against records; seek educator review | | Save/export fails | Use local notes or printed response sheet; avoid unnecessary screenshots | | Student needs an adjustment | Arrange equivalent capability evidence through the appropriate process | | Sensitive information entered | Stop further sharing and follow incident-reporting procedures | ## Access is more than account availability Check mobile and desktop use, keyboard navigation, assistive technology, language support and the time needed to complete the task. A free service can still have limits that make an assessment unreliable. Do not require a personal account or paid upgrade when an equivalent route can meet the outcome. Where AI use itself is assessed, provide suitable approved access and explain the requirement in advance. Offer clear instructions, a short demonstration and a fixed worked example. Students should not have to discover the task's hidden expectations by repeatedly asking a chatbot what the lecturer wants. For a spoken response, assess the relevant reasoning or communication outcome, not eye contact, accent, speed or confidence unless a particular feature is genuinely part of the stated professional capability and assessed fairly. ## Student wellbeing Give advance notice of complaint, discrimination or performance scenarios. Use fictional data, a stop instruction and an agreed neutral alternative. Do not ask students to disclose their own experiences to make the activity authentic. A role-player may disagree, pause or ask a question. The exercise should not escalate emotional pressure merely to make the learner “work harder”. Debrief the decisions and evidence, not whether the student persuaded the persona to become agreeable. ## Budget the whole task Record design, testing, support, moderation and marking time separately. An initial prompt takes little time; a defensible teaching activity takes more. For 24 students, eight minutes of review per submission is 192 minutes. A 30-minute moderation discussion brings that illustrative marking budget to 222 minutes, or 3 hours 42 minutes. Setup and student support are additional. Replace these assumptions with observed times after a pilot. For 100 students, a five-minute individual check takes 8 hours 20 minutes before administration. A sustainable design may distribute checks across tutorials, use several staff or choose another appropriate assurance format. Do not offset those costs with an invented claim that AI always saves a fixed number of minutes per learner. Measure the actual task. ## Keep an approved activity record Save the scenario version, student brief, rubric, tested examples, permitted routes, access arrangements and known limitations. Record material changes between cohorts. For model-supported use, note the product/account and relevant features tested. A later tool change may require rechecking the activity. Do not assume a well-worded prompt remains equally reliable across every model. Keep student evidence only for the authorised purpose and retention period. An audit trail does not require retaining every private conversation indefinitely. ## Evaluate the pilot Use P3 in the [pilot pack](../appendices/pilot-pack.qmd). Compare the baseline and independent response cautiously, examine errors and record obstacles. Separate satisfaction, task completion, output quality and learning evidence. Students may enjoy an activity that does not improve transfer, or find a challenging activity useful despite lower satisfaction. Neither result should be simplified into a universal endorsement. A reasonable decision can be to continue, revise, pause or stop. If a model repeatedly invents case facts, use the fixed route while redesigning. If a particular group cannot access the task, fix the access problem before increasing its stakes. ## Activity: rehearse failure **Allow ten minutes; no AI required.** Imagine the service fails halfway through your activity. Write what students do next, which evidence remains assessable, how they save their work and who approves a changed assessment condition. Then try the fallback yourself. Check that it preserves the capability, does not send data somewhere new and does not penalise students for the outage. That rehearsal belongs before the first graded use.