Practical guide

How to report AI capability evidence to an organizational development leader

Report AI capability evidence to an organizational development leader as a work-system decision brief, not a leaderboard of people, a tool-usage dashboard, or a transformation success story. The report should show which work behavior was observed, under what conditions, what changed across roles and handoffs, where evidence is missing, and which organizational decision the evidence can support.

By Two Prune

Report AI capability evidence to an organizational development leader as a work-system decision brief, not a leaderboard of people, a tool-usage dashboard, or a transformation success story. The report should show which work behavior was observed, under what conditions, what changed across roles and handoffs, where evidence is missing, and which organizational decision the evidence can support next.

Classify the stakeholder and the decision

The organizational development leader is usually the owner of organization-design interpretation and the reporting process for a bounded change initiative. They may be an influencer on investment, technology, or workforce decisions rather than their final owner. Workflow leaders own operational quality; learning teams own programme delivery; technology teams own systems; privacy, security, legal, accessibility, and employee-representative functions review within their mandates. State the local allocation instead of assuming this role owns every budget or decision.

CIPD frames organization development and design around organizational effectiveness, health, systems thinking, and behavioral science. GOV.UK's performance guidance says metrics need a clear purpose, should combine quantitative evidence with user research, and should be used to improve a service. The UK employer guide to AI upskilling emphasizes practical, integrated learning and workplace transfer. These sources support a decision-oriented report, but they do not validate one reporting model or establish Two Prune outcomes. CIPD: Organisation development and design GOV.UK: Set performance metrics UK employer guide to AI upskilling

Sources: [1] [2] [3]

Use the ALIGN reporting brief

ALIGN keeps assessment evidence connected to work design: aim, limits, interactions, gaps, and next decision.

### A — Aim and decision

Open with one sentence naming the decision, owner, and date:

> Decide whether the proposal-review team needs revised decision rights, targeted verification practice, or another bounded pilot before the workflow expands.

State the workflow, population, assessment version, observation window, and alternatives. Do not lead with an average score. A report cannot be interpreted until the reader knows whether it informs task redesign, learning design, support allocation, pilot continuation, or a later evidence request.

### L — Limits and observation conditions

Describe what participants could demonstrate and what they could not. Record task materials, time, tools, AI access, instructions, accommodations, scoring rubric, reviewer process, missing submissions, and technical interruptions. Distinguish assessment performance from live workplace behavior.

Show evidence strength by source:

| Evidence source | What it can show | Important limit | | --- | --- | --- | | assessment deliverable | quality of the submitted work under stated conditions | may not reveal the full process | | workspace evidence | source use, AI interaction, revisions, verification, and timing when captured | logging coverage and interpretation vary | | reviewer judgment | performance against the rubric | requires calibration and can contain disagreement | | participant reflection | perceived reasoning and support needs | self-report is not direct behavior evidence | | operational data | quality, rework, workload, escalation, and service effects in the workflow | is not produced by an assessment alone |

### I — Interactions in the work system

Translate pillar or task evidence into role, handoff, and control questions. For example, weak verification may reflect an individual practice gap, missing source access, unclear decision rights, time pressure, or an interface that hides provenance. Show competing explanations rather than assigning every result to capability.

Use one row per work-system issue:

| Observed pattern | Evidence | Possible mechanism | Affected handoff or role | Contrary evidence | Decision implication | | --- | --- | --- | --- | --- | --- | | | | | | | |

Keep capability, opportunity, motivation, tooling, process, and governance as separate hypotheses. An organizational response should match the mechanism: practice for a skill gap, redesign for ambiguous ownership, access repair for missing sources, or control change for an unsafe default.

### G — Gaps, groups, and confidence

Report distributions and case types before averages. Show how many participants had an opportunity to demonstrate each behavior, how many observations are missing, and whether critical failures occurred. Do not rank teams whose tasks, opportunity, or support differ materially.

For every conclusion, add:

Keep personally identifiable results out of the organizational report unless they are genuinely needed, lawfully handled, and within the stated purpose. Aggregate reporting does not remove re-identification risk in small groups.

### N — Next decision and evidence plan

Finish with no more than three proposed decisions. For each, name the owner, action, evidence needed, stop condition, review date, and what remains unknown. Suitable dispositions are stop, repair the assessment or conditions, run targeted practice, redesign a handoff, or continue a bounded pilot.

Do not prescribe organization-wide rollout from a small simulation cohort. Connect assessment evidence to a later operational measure—such as correction type, review effort, exception handling, workload, or affected-user experience—without claiming the assessment caused that outcome.

  • evidence strength and reviewer agreement;
  • denominator and missingness;
  • alternative explanations;
  • accessibility or participation limitations;
  • whether comparison is directional or supported by a mature benchmark; and
  • what new evidence could change the interpretation.

Worked ALIGN brief: proposal compliance review

A synthetic proposal team of 20 completes a controlled exercise using a 26-page request for proposal, four internal source documents, and an approved AI assistant. The organizational development leader must decide whether to change the proposal manager–specialist handoff before a shadow pilot.

Eighteen participants submit. Two encounter a desktop access issue and are recorded as not observed, not failed. Fourteen locate the conflicting delivery dates. Nine preserve the conflict in the compliance matrix; five silently choose one date. Twelve verify the AI-generated source locations, while six accept at least one incorrect page reference. Reviewers agree on 16 of 18 dispositions and reconcile two cases where the rubric did not define whether a general escalation note was sufficient.

The weak report says “78% of the team is AI ready” and recommends more prompt training. ALIGN produces a different interpretation. The issue is concentrated at a handoff: specialists find the conflict, but the template does not name who owns resolution before the proposal manager approves the matrix. The incorrect page references suggest a verification practice need, while the two inaccessible sessions reveal an assessment-condition problem.

The brief shows three competing mechanisms: insufficient checking practice, ambiguous resolution ownership, and a template that visually treats AI suggestions like verified evidence. Contrary evidence matters: several participants with concise prompts still verify well, and several elaborate prompts do not. Prompt sophistication is therefore not used as the explanation.

The next decisions are: product operations adds a required conflict owner and evidence status to the matrix; learning runs a source-location verification practice for the 18 observed participants; and the assessment owner repairs the access condition before repeating the task. A six-case shadow pilot later records correction types, review minutes, unresolved conflicts, and escalations. No organization-wide capability percentage is reported.

Objections the report should answer

“Why not show one score?” A composite can summarize assessment performance, but organizational action depends on the underlying behavior, opportunity, evidence strength, and work-system mechanism.

“Can we compare departments?” Only when task, materials, support, opportunity, scoring, and relevant context are sufficiently comparable. Otherwise show each group's evidence against its own decision.

“Does this prove transfer?” No. A realistic assessment can provide task-bounded evidence. Transfer requires later observation in a new or workplace context with an appropriate design.

“Should managers receive individual results?” The report's purpose, privacy basis, access rules, and product boundary govern that choice. Do not repurpose assessment evidence for automated employment decisions.

Reporting QA

[ ] The first line names a decision, owner, workflow, and date.

  • [ ] Stakeholder roles distinguish owner, influencer, reviewer, and collaborator.
  • [ ] Assessment version, conditions, opportunity, support, and missingness are visible.
  • [ ] Observations are separated from inference and hypothesis.
  • [ ] Capability evidence is not confused with tool usage, confidence, or business impact.
  • [ ] Role, handoff, process, access, and control explanations are considered.
  • [ ] Critical failures remain visible and are not averaged away.
  • [ ] Comparisons state their denominator, benchmark maturity, and limits.
  • [ ] Privacy, accessibility, reviewer disagreement, and re-identification risks are addressed.
  • [ ] Every recommendation has an owner, evidence gate, stop condition, and review date.

Two Prune boundary

Two Prune can report evidence from a controlled assessment: the task, final deliverable, observed source use, AI interactions, verification, revisions, rubric results, and confidence limitations when available. It does not diagnose an organization, prove workplace transfer or productivity, determine organization design, validate mature benchmarks without evidence, or make employment decisions.

Sources

CIPD: Organisation development and design

  • GOV.UK Service Manual: How to set performance metrics for your service
  • UK Government: Employer guide—what works for AI upskilling

Sources: [1] [2] [3]