Practical guide

AI capability measurement agenda for an HR leader

An HR leader does not need another generic AI maturity score. They need a decision agenda that separates learning, work redesign, access, evidence quality, and human consequences. The useful artifact is a capability-measurement agenda: a compact record of the workforce decision, the people affected, the evidence needed, who owns each judgment, what a small pilot may establish, and what the.

By Two Prune

An HR leader does not need another generic AI maturity score. They need a decision agenda that separates learning, work redesign, access, evidence quality, and human consequences. The useful artifact is a capability-measurement agenda: a compact record of the workforce decision, the people affected, the evidence needed, who owns each judgment, what a small pilot may establish, and what the evidence must never be used to infer.

Start with the decision, not the metric

Use ALIGN: Assign the decision, Locate the work change, Identify evidence, Govern consequences, and Name the next gate. This keeps HR's mandate explicit without making HR the owner of technical, functional, or legal decisions that belong elsewhere.

The NIST AI Risk Management Framework Core calls for documented roles, differentiated human and AI responsibilities, training for people with risk-management duties, and engagement with relevant actors. Its actor-task appendix shows that AI work involves many roles rather than one universal owner. OECD's employer and worker surveys investigate how AI use relates to work organisation, skills, training, and worker experience. These are useful design inputs, not proof that a particular programme improves performance or that one governance model fits every employer. NIST AI RMF Core NIST: descriptions of AI actor tasks OECD AI surveys of employers and workers

### A: Assign the decision and HR's role

Write the exact decision before selecting evidence. Examples include whether to fund targeted practice, revise a learning pathway, repair access barriers, repeat a controlled trial, or ask a function owner to redesign a task. For that decision, label HR as owner, influencer, reviewer, or collaborator and state the authority attached to the label.

A responsibility map should name at least the functional work owner, learning owner, assessment owner, technology or security owner, participant representative, and accountable decision maker. “HR is involved” is not sufficient. “HR owns the learning portfolio and feedback route; the function owner defines task relevance; security governs access; the assessment owner governs scoring; the business sponsor decides whether to continue the pilot” is testable.

Do not assume that HR always owns the budget, system approval, work quality, or rollout decision. The agenda should expose the local mandate rather than import a universal buying process.

### L: Locate the work change and current alternatives

Describe the work before describing AI capability. Name the task, audience, inputs, output, judgment points, verification duties, time pressure, and material failure. Identify the affected population and whether the change alters access, workload, monitoring, review burden, or development opportunity.

Record the current alternatives: manager observation, course completion, tool-use analytics, work samples, structured simulations, quality audits, self-report, peer review, or no formal evidence. For each alternative, note what it can and cannot show. Course completion shows participation, not transfer. Tool activity shows use, not sound judgment. A work simulation shows bounded observed performance, not a universal trait or an assured workplace outcome.

### I: Identify a decision-evidence map

Use a table that ties each question to evidence rather than collecting whatever is easiest to count.

| Decision question | Useful evidence | Common misread | Evidence owner | | --- | --- | --- | --- | | Did people receive a fair opportunity to learn? | access, accommodations, practice exposure, attendance, feedback | attendance equals capability | learning owner | | Can participants perform a defined AI-enabled task? | deliverable, source use, corrections, workflow evidence, rubric | one task predicts all work | assessment owner | | Is the work design usable? | review load, exceptions, participant feedback, accessibility tests | satisfaction proves effectiveness | function and product owners | | What support is needed next? | pillar-level patterns, evidence gaps, repeated failure modes | low score justifies an employment action | HR and function owner |

Define the claims boundary in the agenda. A small pilot may reveal recurring verification failures in a bounded task. It cannot establish that a workforce is broadly “AI ready,” prove return on investment, or justify automated employment decisions.

### G: Govern consequences and objections

List predictable objections before launch. Participants may worry that development evidence will be repurposed for selection or discipline. Managers may fear extra review work. Security may object to uncontrolled tools or sensitive inputs. Functional leaders may reject generic tasks. Accessibility specialists may identify unequal evidence opportunities.

For each objection, name a control and an escalation route. Keep development, access, and performance-management purposes separate. State who may see participant-level evidence, how long it is retained, what human review is required, and how a participant can raise a material error. Missing evidence should become a confidence limitation, not an inferred weakness.

The agenda should also record prohibited uses. Two Prune is an assessment product, not an applicant tracker, resume screener, or automated decision system. Assessment evidence can inform a human-led learning or evaluation workflow within a stated boundary; it should not be presented as an employment recommendation or as evidence of performance in every future role.

### N: Name the pilot and next evidence gate

Design the smallest pilot that can answer the decision. Specify the participant group, task, materials, time, support, comparison or baseline where appropriate, evidence packet, reviewer calibration, success and stop conditions, and next decision date. Avoid choosing the sample size because it produces an impressive percentage. Choose it because the cases provide useful variation and because reviewers can inspect the evidence properly.

The next gate should be concrete: repair the task, repeat with a different work context, provide targeted practice, test an accessibility change, or stop because the evidence cannot support the intended use. Do not promise organization-wide outcomes from a pilot.

Sources: [1] [2] [3]

Worked agenda: a 20-person learning-transfer cohort

A synthetic professional-services team wants to know whether a source-verification workshop transfers to a policy-briefing task. Twenty volunteers who completed the workshop enter a controlled simulation containing an official release, an outdated secondary summary, a plausible unsupported claim, and a time constraint. The output is a one-page briefing with an evidence note.

The responsibility map makes HR the owner of the learning portfolio, participant communication, and use boundary. The policy function owner defines task relevance and reviews the business deliverable. Security approves the controlled materials and access. The assessment owner controls the rubric, scorer guidance, and evidence packet. The sponsor decides whether to fund another learning cycle. No participant-level result is used for selection, promotion, or discipline.

The evidence map separates four measures. Workshop completion confirms exposure. Tool logs show which materials were opened but not whether they were understood. The assessment records source selection, corrections, unsupported claims, final communication, and evidence strength. A short feedback form identifies access and workload issues. Reviewers receive the final deliverable and traceable evidence rather than a single opaque score.

The cohort shows that 14 participants find the official release, while nine still repeat the outdated statistic and only six explicitly preserve uncertainty. The agenda does not label the group “not AI ready.” It records a repeated source-hierarchy failure, asks the learning owner to add contrastive practice using dated releases, and asks the function owner to make source freshness an explicit task requirement. Two participants report difficulty navigating the source viewer, so the product owner must test the accessibility issue before the next cohort.

The next gate is a targeted repeat with repaired learning materials and the access fix. The sponsor will review whether the specific verification behavior improves and whether reviewer burden remains acceptable. No claim is made about organization-wide productivity, return on investment, or workplace performance.

Two Prune boundary

Two Prune can provide task-bounded evidence of how participants frame work, navigate sources, use AI, verify outputs, exercise judgment, and communicate in a realistic simulation. It explains observed assessment performance and evidence strength. It does not run HR processes, infer a universal capability trait, predict workplace outcomes, or make employment decisions.

Sources

NIST AI Risk Management Framework Core

  • NIST: Descriptions of AI actor tasks
  • OECD: AI surveys of employers and workers

Sources: [1] [2] [3]

Sources

  1. 1.NIST AI RMF Core · National Institute of Standards and Technology
  2. 2.NIST: descriptions of AI actor tasks · National Institute of Standards and Technology
  3. 3.OECD AI surveys of employers and workers · www.oecd.org