AI can draft the course. It cannot prove anyone learned.

Faster authoring creates a release-control problem

SAP's October 12 SuccessFactors 2H 2026 announcement says the WalkMe Learning Arc integration can generate courses and quizzes from natural-language requirements inside the learning system while maintaining author control. That final phrase is the operating boundary. AI may create a draft quickly. It does not establish that the draft is accurate, teachable, accessible, fair, or effective.

WalkMe's current documentation describes AI-generated content, courses, quizzes, simulations, trusted input sources, translation, and SCORM export. The product capability is concrete. Public evidence of customer outcomes is not. This guide therefore uses the release as a catalyst, not as proof that organizations can turn weeks of learning design into hours without quality loss.

The independent evidence is more cautious. A 2026 systematic review and meta-analysis of AI-generated multiple-choice questions found no overall difference from human-authored items on average difficulty or discrimination, but heterogeneity was high and equity evidence was limited. Another 2026 review of 71 studies found error rates ranging from below 1 percent to 45 percent, incomplete validity evidence, and no support for unsupervised summative use. The practical conclusion is useful and narrow: AI can accelerate item drafting, but it does not remove expert validation or field evidence.

CDC's Quality Training Standards offer a durable control model: establish the need, write measurable objectives, keep content accurate and relevant, engage learners, design for accessibility, evaluate the training, assess learners, and provide follow-up support. An AI authoring workflow should make those obligations more visible, not compress them into a prompt.

Start with the problem, not the course request

Many requests arrive already framed as content: “build refresher training,” “turn this policy into microlearning,” or “make a quiz.” The first decision is whether training can change the observed gap. Employees may already know the rule but lack system access, time, authority, a usable process, or consistent management. Publishing another course can create completion evidence while leaving the failure untouched.

Document the business outcome and baseline. “Improve compliance” is too vague. “Within 30 days, payroll specialists classify five exception types with at least 90 percent accuracy and escalate the two reserved for counsel” can be taught and observed. The objective should describe a behavior, conditions, threshold, and timeframe.

Classify consequence before drafting. A voluntary product overview can tolerate more exploratory language than lockout-tagout, anti-bribery, privacy, financial reporting, clinical, licensing, or disciplinary training. High-consequence content needs stronger source authority, qualified reviewers, realistic practice, conservative scoring, and a narrower release.

Build one trace from source to workplace behavior

The source register is the foundation. Record title, owner, version, effective date, jurisdiction, population, exact relevant passages, and review or expiration date. Generated citations are not evidence. Retrieve the real source and verify that the cited language applies to the learner and current process.

Then map each objective to content, practice, assessment, feedback, and owner. This exposes three common defects. Orphan content consumes attention but supports no objective. Orphan objectives appear in the course description but are never practiced or assessed. Orphan questions test trivia, wording, or assumptions not required by the job.

objective_id: OBJ-03
behavior: classify a suspected data incident
conditions: given one of six approved scenarios
threshold: correct severity and escalation route in 5 of 6 cases
sources:
  - id: POL-PRIV-2026-04
    passage: sections 4.2-4.5
content: lesson-2/severity-criteria
practice: simulation-2/branching-cases
assessment: [Q-07, Q-08, TASK-02]
feedback_owner: privacy-office
release_gates:
  - source_current
  - sme_approved
  - accessibility_passed
  - pilot_threshold_met

Verify claims at the smallest useful unit. A paragraph can contain one supported rule, one outdated number, and one invented exception. Split it into statements and record the supporting passage. Reperform calculations and inspect examples for hidden authority. An example that shows a manager overriding a privacy rule can teach the wrong behavior even when the rule slide is correct.

Keep the generation record: tool, model/version when available, prompt, source set, settings, output version, edits, and reviewers. The record is not for blaming a model. It makes the released package reproducible and tells reviewers what must be rechecked after a change.

A plausible quiz can still be invalid

An assessment item must test the objective, not merely resemble the lesson. Review the stem, intended answer, distractors, feedback, reading burden, source, and consequence. Check for accidental clues: the longest option, copied wording, absolute terms, grammatical mismatches, or one answer much more specific than the others.

Multiple-choice questions are strongest for knowledge, discrimination among options, and some scenario judgments. They are weak evidence for a learner's ability to navigate a system, perform a physical task, conduct a difficult conversation, write a defensible analysis, or exercise authority under uncertainty. Use observed practice, simulation, work samples, or supervised performance when the objective requires application.

Do not ask the same model to generate and certify the item as the only review. Independent model critique may find defects, but it is not validity evidence. Require human subject-matter approval and inspect pilot behavior. For consequential assessment, use psychometric expertise and qualified legal or professional review where applicable.

Worked example: a conduct-policy course teaches the wrong escalation

An HR team asks an AI authoring tool to turn a new conduct policy into a 15-minute course. The draft looks polished. It states that an employee should report concerns to their manager before using the confidential hotline. The real policy says employees may contact a manager, HR, compliance, or the hotline directly. The generated sequence quietly adds a gate that could deter reporting.

The traceability review catches the sentence because it has no supporting passage. The reviewer replaces it with the actual choice of routes. A generated quiz then asks, “What should you do first?” and marks “tell your manager” correct. That item is also blocked: the stem creates an order the policy does not require, and more than one option should be valid.

The corrected assessment uses a scenario: a manager is named in the concern. Learners must select two available routes and identify the prohibition on retaliation. The answer key cites the relevant sections. Feedback explains why no manager-first rule exists. A compliance owner approves the scenario, and an accessibility reviewer confirms the options and feedback work by keyboard and screen reader.

release_record:
  defect: invented manager-first escalation rule
  source_conflict: true
  affected_assets: [lesson-3, Q-06, narration-3]
  correction: direct routes listed without priority
  reviewers: [conduct-policy-owner, compliance, accessibility]
  challenge_case: manager-is-subject-of-report
  pilot_result: 18/20 select two valid routes; 2 require coaching
  decision: release-with-monitoring
  review_trigger: policy-route-change

The example shows why prose review is not enough. One unsupported phrase contaminated lesson content, narration, and the answer key. A trace identifies every affected asset and prevents a partial correction.

Accessibility, language, privacy, and employment impact are release gates

Generated learning can fail before anyone reaches the quiz. Test headings, labels, focus order, keyboard operation, contrast, captions, transcripts, alternative text, timing, motion, documents, error messages, and alternative formats. Automated accessibility checks catch only part of the problem; use assistive technology and human review for the real learner path.

Translation needs both language and subject-matter review. A fluent translation can change a defined term, authority, safety instruction, or legal meaning. Record the source language, target language, tool, translator or bilingual reviewer, changed terms, and approved version. Reapprove translations when the source changes.

Do not upload real investigation files, employee examples, performance notes, medical information, customer records, or confidential incidents just to make scenarios realistic. Use synthetic cases that preserve the decision structure. If real data is necessary, document purpose, approved tool, access, retention, deletion, logging, incident response, and prohibited downstream uses.

Keep learning data separate from employment decisions unless policy and law explicitly permit a designed use. A low score can reflect accessibility barriers, language, a bad item, technical failure, lack of prerequisite knowledge, or an invalid answer key. Do not let an AI-generated assessment become an unreviewed basis for discipline, promotion, scheduling, or termination.

Pilot the released behavior, not the authoring speed

A pilot should include representative learners, not only project sponsors and subject-matter experts. Sample roles, prior knowledge, language, location, device, work conditions, and access needs. Include people likely to encounter the edge cases the course is meant to address.

Measure at four levels. First, can learners access and complete the experience? Second, do they meet objective-aligned knowledge or skill thresholds? Third, can they apply the behavior in a realistic task? Fourth, does the workplace outcome improve without unacceptable side effects? Not every course needs a controlled study, but every release needs measures proportionate to consequence.

Set consequence-based stop rules. One typo can enter a correction queue. An invented safety instruction, wrong legal rule, inaccessible mandatory assessment, confidential-data exposure, discriminatory scenario, or unsupported answer key should stop the affected release immediately. Preserve learner records, correct scores, notify owners, replace the asset, and retest.

Reapproval is event-driven, not annual by default. Trigger it when a policy, process, law, system, learner population, language, source, model, prompt, authoring product, integration, scoring rule, accessibility requirement, or outcome changes. A fixed review date remains useful as a backstop.

Frequently asked questions

Can we use AI to generate mandatory compliance training?

Use it to draft only within an approved workflow. Mandatory or regulated content needs current authoritative sources, qualified SME and legal/compliance review where applicable, accessibility, valid assessment, release authority, monitoring, and a correction route.

Does author control inside the LMS solve the risk?

It creates an important control point. It does not define what the author must verify, who is qualified, which evidence is required, or how defects are detected after release. Those belong in the operating workflow.

Are AI-generated quiz questions good enough?

They can be efficient drafts. Current reviews show promising averages but variable quality and incomplete validity evidence. Human review and pilot evidence remain necessary, especially when scores have consequences.

What is the minimum source record?

Capture source title, owner, version, effective date, jurisdiction, audience, exact passage, access location, expiration or review date, and the content and assessment items it supports.

What if the course is correct but learners still fail?

Check whether training addressed the real cause, whether objectives and practice match the task, whether the interface and workplace support performance, and whether the assessment is valid. Do not assume learner motivation is the problem.

Can completion data go into performance management?

Only under an explicit, lawful, reviewed policy. Completion and quiz scores can be distorted by access, language, bad items, technical failures, and accommodation needs. Keep generated assessment from becoming an unreviewed employment signal.