Practical guide
Analyze customer learning health with AI
Reconcile a fictional learning cohort into access, practice and demonstrated-understanding states before choosing a support response.
Customer education managers can use AI to reconcile learning activity while keeping access, practice and demonstrated understanding separate. This fictional twelve-person cohort exercise produces a support-decision brief, not an account-health score, a customer outcome claim or a judgment about any real learner.
Define the learning decision before the metric
Learning health is useful only when it points to a bounded support choice.
The fictional programme owner needs to decide whether to repair access instructions, revise one lesson or schedule a facilitated practice session. The cohort contains twelve assigned learners. This exercise does not assess contract value, renewal likelihood, product adoption or the performance of named people.
Ask AI to restate the three possible responses and the evidence each would need. Keep a no-change option visible. A low completion percentage alone cannot distinguish an access barrier from an unclear lesson, limited practice time or a logging problem.
Build mutually exclusive learning states
Each learner should occupy one current state so the total can be checked.
The fictional event packet places three learners in assigned but not accessed, two in accessed but not attempted, three in attempted but not yet passed and four in passed. The states total twelve. Passed means the supplied practice check met its exercise rule; it does not prove transfer to work outside the lesson.
Ask AI to map stable learner identifiers to the latest supported state, then verify every row against the event packet. Do not use names, demographic attributes or free-text personal details. Preserve earlier events in the audit trail rather than counting the same learner in several current-state totals.
Separate access evidence from learning evidence
No attempt after assignment can reflect a route problem, while an unsuccessful attempt creates different questions.
The three not-accessed records support an access follow-up question, not a conclusion that the learners chose not to participate. The two accessed-without-attempt records show that the lesson opened but do not reveal whether time, clarity or another condition prevented practice.
The three attempted-not-passed records provide a bounded opportunity to inspect which practice decision was difficult. Aggregate the missed step only if the exercise packet supplies comparable attempts. Do not ask AI to infer motivation, ability or customer sentiment from event absence.
Choose responses that match the state
A support plan should target the observed barrier without treating every learner the same.
For the not-accessed state, propose checking the invitation route and access instructions. For accessed-without-attempt, propose a reminder that names the practice task and available help. For attempted-not-passed, inspect whether the feedback explains the decision rather than merely marking an answer wrong.
Keep each action proposed until the responsible programme owner accepts it. Do not send messages, change access or rewrite the lesson in this exercise. The four passed records remain in the denominator and may still need a later transfer check, but they do not need the same immediate response.
Correct a duplicated learner identity
A record correction should change the state table without rewriting the original evidence.
A later fictional note shows that one not-accessed identifier and one passed identifier belong to the same learner after an account migration. Link the duplicate records, retain the correction reason and count that learner once in passed. The cohort now contains eleven distinct learners, with two not accessed, two accessed without attempt, three attempted not passed and four passed.
Ask AI to recompute the counts and confirm that two plus two plus three plus four equals eleven. Do not present the correction as improved learning. It repairs the unit of analysis and changes the denominator; it does not create a new attempt or alter the evidence attached to the other learners.
Deliver a learning-health brief with limits
The final artifact should make the support decision, arithmetic and unknowns easy to inspect.
Include the decision, state definitions, before-and-after counts, duplicate disposition, proposed response by state, owners and evidence still needed. Label every case and number as fictional. Keep the practice-check result bounded to this lesson and do not convert the state table into a general customer-health rating.
Review record identity, denominator discipline and whether each proposed response follows the observed state. The customer-education parent guide creates and tests a troubleshooting lesson. The customer-success guide prepares an account review. This page instead diagnoses a cohort-level learning pathway so a programme owner can choose a targeted support response.
Sources and scope
NIST: AI risk management. Skills England: workplace AI foundations.
These references provide background, not validation or endorsement of this exercise. The learning states, calculations and support choices are original fictional instructional material. Use them to discuss observable work, not to infer customer outcomes, learner traits, professional credentials or performance in every setting. Before adapting the exercise, confirm the relevant facts, approved tools, data permissions and decision owners. If the learning task or event definitions change, revisit the state rules and expected calculations as well.
Sources: [1] [2]
Sources
- 1.AI RMF Core · NIST
- 2.AI foundation skills for work benchmark · Skills England