Practical guide

AI learning needs analysis verification checklist

A learning needs analysis is decision-ready only when a reviewer can trace a real work requirement to evidence of current performance, test alternative causes of the gap, and see why the proposed response is proportionate. A list of requested courses, confidence scores, or role labels is not enough. Use this VALID checklist to review an existing analysis: Verify the decision, Anchor the work,.

By Two Prune

A learning needs analysis is decision-ready only when a reviewer can trace a real work requirement to evidence of current performance, test alternative causes of the gap, and see why the proposed response is proportionate. A list of requested courses, confidence scores, or role labels is not enough. Use this VALID checklist to review an existing analysis: Verify the decision, Anchor the work, Locate evidence, Isolate causes, and Design the response. Copy the checklist into the analysis record and attach evidence links, versions, owners, and dates.

Review header

Freeze the reviewed version. If tasks, tools, policies, groups, or source evidence change materially, record the change and decide which checks must be repeated.

  • [ ] Analysis title and version:
  • [ ] Decision to be made:
  • [ ] Accountable decision owner:
  • [ ] Affected group and work context:
  • [ ] Evidence period and cutoff:
  • [ ] Reviewer and relevant expertise:
  • [ ] Analysis method and source versions:
  • [ ] Review outcome: pass / repair / hold / stop
  • [ ] Next review date:

V: Verify the decision and boundary

The U.S. GAO's 2026 strategic training guide is written for federal agencies. It supports checking whether learning is aligned to mission needs, includes stakeholder input, and examines organizational, occupational, and individual needs as well as causes and possible solutions. Use it as contextual guidance, not as a universal verdict on a private programme. U.S. GAO: Strategic training and development guide

  • [ ] The analysis names one concrete decision, not only “improve AI skills.”
  • [ ] The owner has authority to act on the result.
  • [ ] The population, team, work context, and period are explicit.
  • [ ] Included and excluded tasks are named.
  • [ ] The analysis distinguishes organizational, occupational, team, and individual questions.
  • [ ] Data use, participant communication, access, and retention are defined.
  • [ ] The intended use excludes automated employment decisions.
  • [ ] Assumptions and unresolved questions remain visible.

Sources: [1]

A: Anchor requirements in real work

For every claimed need, require a row:

| Task | Trigger and inputs | Critical judgment | Observable behavior | Evidence of a good result | Failure consequence | | --- | --- | --- | --- | --- | --- | | | | | | | |

NIST's AI RMF Core calls for defined operator proficiency, documented human oversight, clear roles, and context-specific mapping of intended use and risk. It is voluntary risk-management guidance and does not prescribe a curriculum. NIST AI RMF Core

  • [ ] Tasks are described through work, not role labels alone.
  • [ ] Current tools, sources, policies, time pressure, and handoffs are included.
  • [ ] The required behavior is observable and reviewable.
  • [ ] AI use is separated from domain, data, communication, and risk-control requirements.
  • [ ] High-consequence and frequent tasks are prioritized transparently.
  • [ ] Human oversight and escalation actions are specified where relevant.
  • [ ] A non-AI or simpler method is considered where it could meet the need.

Sources: [2]

L: Locate evidence of current capability

Use this evidence table:

| Requirement | Evidence source | Sample | Observation | Missingness | Confound | Strength | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | supported / partial / unsupported / insufficient |

The OECD reports complementary employer and worker AI surveys because implementation and first-person experience reveal different information. That design supports using multiple perspectives, while the survey's sectors, countries, and self-reported measures limit direct transfer to a local needs decision. OECD: AI surveys of employers and workers

  • [ ] At least two evidence types inform each material conclusion.
  • [ ] Sources distinguish work samples, observation, simulations, system activity, interviews, self-report, and manager judgment.
  • [ ] Sampling and recruitment are recorded.
  • [ ] Task, rubric, reviewer, tool, and source versions are frozen.
  • [ ] Missing evidence is reported separately from poor performance.
  • [ ] Reviewers use observable anchors and record disagreement.
  • [ ] Activity and confidence are not presented as capability by themselves.
  • [ ] Conclusions are limited to the represented task, group, and period.

Sources: [3]

I: Isolate learning from other causes

Rate each possible cause:

| Cause | Evidence for | Evidence against | Decision | Owner | | --- | --- | --- | --- | --- | | access | | | repair / rule out / investigate | | | clarity | | | | | | workflow | | | | | | incentive | | | | | | source or system | | | | | | capability | | | | |

Hold the learning recommendation when a material operating cause is untested. Training someone to verify an obsolete source more carefully does not repair the source.

  • [ ] Tool, source, permission, and time access are tested.
  • [ ] Expected behavior and decision rights are clear.
  • [ ] Workflow design does not make the correct action impractical.
  • [ ] Incentives and local norms do not reward the failure.
  • [ ] Source quality and system behavior are checked.
  • [ ] Manager support and escalation routes are available.
  • [ ] The analysis records evidence for and against a capability cause.
  • [ ] Non-learning repairs are assigned to owners.

D: Design a proportionate response and test

[ ] The response matches the diagnosed cause.

  • [ ] The smallest useful intervention is considered first.
  • [ ] Target behavior, task conditions, and expected evidence are explicit.
  • [ ] A job aid, workflow repair, access change, coaching, or source correction is considered alongside training.
  • [ ] Measures and decision rules are set before results are seen.
  • [ ] The test includes missingness and confound rules.
  • [ ] Continue, repair, hold, and stop conditions are defined.
  • [ ] Transfer to everyday work is not assumed.
  • [ ] The owner, review date, and next decision are recorded.

Test 1: broad prompt course request

An operations leader requests a prompt-writing course after customer advisers produce inconsistent policy answers. The analysis lists low confidence and sparse AI usage as the need. VALID returns repair.

The decision is vague, no task is anchored, and self-report is the only evidence. A reviewer samples 28 cases and observes six with conflicting policy versions. Four errors involve a missing exception, including two answers created without AI. Search ranks the older policy first and the escalation route is unclear.

The revised analysis defines source reconciliation as the target behavior, assigns a source-link repair and an escalation rule, and proposes a short practice case with frozen reviewer anchors. The broad course is not approved because the evidence does not isolate prompting as the cause.

Test 2: targeted verification need

A finance transformation team must review AI-assisted variance explanations. The analysis maps the task, uses work samples and a controlled simulation, records reviewer agreement, and separates missing calculation logs from unsupported reasoning. It also confirms that the current workflow exposes source data and that review time is sufficient.

VALID finds a bounded gap: several participants accept a plausible explanation without reproducing the calculation. The response is a worked-example clinic plus a calculation-check aid, tested on an alternate task form. The analysis passes with a limitation: one simulation does not establish transfer or population prevalence.

Hold conditions

Hold the analysis when the decision owner is absent, the task cannot be described, requirements are unobservable, critical evidence comes from one weak source, missingness is scored as failure, alternative causes are ignored, a proposed course precedes the diagnosis, or claims exceed the sampled task and group.

Two Prune can provide a realistic assessment simulation and evidence packet for one bounded work context. It does not diagnose a workforce by itself, establish mature norms, promise learning outcomes, or make employment decisions. The programme owner remains responsible for the analysis and response.

Sources

U.S. GAO: A Guide for Developing and Assessing Strategic Training and Development Efforts

  • NIST AI Risk Management Framework Core
  • OECD: AI surveys of employers and workers

Sources: [1] [2] [3]

Sources

  1. 1.U.S. GAO: Strategic training and development guide · U.S. Government Accountability Office
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology
  3. 3.OECD: AI surveys of employers and workers · www.oecd.org