People and Change KPI Tree

COMPEL Specialization — AITE-WCT: AI Workforce Transformation Expert Article 33 of 35

A KPI tree for an AI workforce transformation is the artefact that lets the CEO, the CHRO, and the Head of AI Governance share a common picture of programme health without individual-project-level detail. It is the sensemaking instrument the programme uses weekly and the reporting instrument the board uses quarterly. A tree that is well-built answers the question “is the transformation actually working, and how do we know” in a way that survives scrutiny. A tree that is badly built produces either complacency (the numbers look fine because they measure the wrong things) or panic (the numbers look bad because the thresholds are mis-set).

The tree is a three-level hierarchy rooted in the transformation’s outcomes: what the organisation is trying to produce through the workforce transformation. Beneath the outcomes sit the drivers: the intermediate states that produce the outcomes. Beneath the drivers sit the metrics: the specific, measurable, wired-to-data signals that let the organisation know whether the drivers are moving.

The tree’s value is not in the metrics. The tree’s value is in the discipline it imposes on the argument that connects work to outcome. An organisation that cannot articulate its people-and-change KPI tree does not, in practice, know what it is trying to produce from its workforce transformation investment.

The three levels

Level 1 — outcomes

Outcomes are the aggregate states the transformation is designed to produce. They are organisation-level, multi-year, board-meaningful.

The standard outcomes for an AI workforce transformation are three: productivity, retention, and engagement. Each is a composite; each can be measured; each has standing organisational instrumentation the transformation can build on rather than reinvent.

• Productivity. Output per unit of input. The input is typically hours; the output is task-completion, decision-quality-adjusted throughput, revenue per employee, or domain-specific productivity measures. AI workforce transformation is typically justified in part by productivity improvement, and productivity is the outcome the board will ask about most specifically.
• Retention. Voluntary exit rates, filtered to exclude desired exits (low performers) and examined by segment (high performers, mid-career, newly hired, AI-role incumbents, and so on). Retention matters during transformation because the transformation itself increases voluntary-exit risk; a transformation that produces strong productivity gain while losing the high-performing workforce has not, on net, succeeded.
• Engagement. Employee engagement scores from the organisation’s standing instrument, decomposed by the dimensions the instrument measures (usually variations on enablement, meaning, growth, recognition, belonging). Engagement is both an outcome and a leading indicator — it correlates, with lag, with productivity and retention.

Some programmes add a fourth outcome: capability, the workforce’s aggregate ability to do the work the transformation is building toward. Capability is measurable as a composite of literacy coverage, skills-adjacency coverage, and observed proficiency in AI-integrated tasks. It is arguably a driver rather than an outcome; organisations that include it at Level 1 are signalling that capability itself is a strategic outcome.

Level 2 — drivers

Drivers are the intermediate states that produce the outcomes. They are programme-level, shorter-horizon, and connect directly to the work the transformation is doing.

For productivity, the drivers typically include: AI-tool adoption depth (how much of the target workflow is actually AI-integrated, not just whether the tool is available); AI-tool quality-of-use (are employees using the tools well, as measured by proxy quality measures); manager coaching cadence actual-versus-target (a strong predictor of sustained adoption); workflow-friction reduction (the changes to upstream and downstream processes that let AI-integrated work pay back).

For retention, the drivers include: psychological-safety score; clarity of role (do employees understand their AI-augmented role); growth-path visibility (do employees see a future in the organisation); recognition pattern (do employees receive recognition aligned with what the organisation claims to value); compensation competitiveness (particularly for high-adjacency-skill employees who are retention-risk).

For engagement, the drivers include: meaningfulness (do employees experience their AI-integrated work as purposeful); autonomy (do they have appropriate latitude); mastery (are they developing capability); inclusion (do they feel belonging and equitable opportunity).

Drivers are the level at which the programme acts. The interventions the transformation runs target the drivers, and the drivers move the outcomes. An intervention that cannot be placed on the tree — that does not connect to a specific driver that connects to a specific outcome — is a programme hygiene failure.

Level 3 — metrics

Metrics are the specific measurements that let the organisation know whether the drivers are moving. They are programme-operational, weekly to monthly, wired to data sources.

The metrics must be: specific (measured in a defined way, from a defined source), reliable (measurement consistent across time and across observers), valid (actually measuring what the driver claims to measure), and actionable (a change in the metric provides information the programme can act on).

For adoption-depth driver, example metrics: fraction of target-workflow invocations that use the AI tool; distribution of depth-of-use across the user population; month-on-month trajectory.

For manager-coaching-cadence driver, example metrics: fraction of manager-direct-report pairs with a weekly one-to-one in the past four weeks; fraction of one-to-ones that include AI-coaching content; direct-report-reported usefulness of the coaching.

For meaningfulness driver, example metrics: survey items on “my work feels meaningful and valuable”; trajectory over quarters; variance by segment.

The Level-3 metric set is where most KPI trees accumulate clutter. An organisation with 90 metrics at Level 3 does not have a KPI tree; it has a data dump. The disciplined set is 3–5 metrics per driver, chosen for signal-to-noise and for coherence.

Wiring to data sources

A KPI tree is only useful if the metrics are actually computed, with acceptable latency, from a source the organisation trusts. The wiring design is a serious undertaking that, if neglected, produces the common pathology where the tree is beautifully designed and nobody can populate it.

Wiring principles:

• Use existing sources where possible. HRIS for headcount and composition; engagement platform for survey metrics; LMS for literacy coverage; AI-tool telemetry for adoption depth; performance-management system for performance distributions. Pulling from existing sources is cheaper and more sustainable than building new collection.
• Instrument new sources where necessary. Some metrics require instrumentation that does not exist: AI-tool quality-of-use often needs sampling-based review; workflow friction often requires targeted interview-based measurement; voice-in-meetings requires observation design. New instrumentation is resourced and staffed; it does not emerge by itself.
• Document provenance. Every metric has a documented source, a documented computation method, a documented owner, and a documented refresh cadence. Metrics whose provenance cannot be described are not ready for board reporting.
• Automate where tractable. Metrics that update automatically from source systems are maintained; metrics that require manual periodic update tend to decay. Automation investment at Level 3 pays back over the programme’s multi-year horizon.

The wiring is typically under-resourced. A programme that allocates 10% of its measurement budget to wiring spends the other 90% designing dashboards that do not populate.

The two common failures

Vanity metrics at the top

Vanity metrics produce a dashboard that looks green regardless of what is actually happening. In people-and-change contexts, the common vanity metrics include: training hours delivered (a measure of programme activity, not of capability built); communication reach (a measure of broadcast, not of understanding); satisfaction scores on training (a measure of experience, not of behaviour change); participation in optional events (a measure of engagement by the already-engaged).

Each of these metrics has a legitimate use at Level 3 or as a sub-metric; the failure is when they are lifted to Level 1 or 2 and treated as if they measure outcome. A Level-1 dashboard dominated by vanity metrics produces complacency at the board level and cynicism among the workforce who can see through it.

The expert’s discipline is to resist vanity-metric elevation under pressure. When outcome metrics are moving slowly, the temptation to show activity metrics as progress is strong; the honest response is to show the actual trajectory of the outcome metrics and explain why the trajectory is consistent with the programme design at its current stage.

Proxy metrics at the bottom

Proxy metrics are measurements that stand in for what is actually being tracked but do not capture it well. In people-and-change contexts, common proxies include: calendar-meeting frequency as proxy for collaboration; email volume as proxy for productivity; tenure as proxy for capability; promotion rate as proxy for development.

Each of these proxies has correlation with the underlying construct, but each also has substantial independent variation that makes the proxy an imperfect stand-in. Over time, if the proxy is what the programme measures, the proxy becomes what the programme optimises, and the underlying construct drifts out of sight. This is Goodhart’s Law in KPI-tree form.

The corrective is to combine proxies with direct measurement where possible. Calendar-meeting frequency + observed meeting effectiveness (sampled review) is better than either alone. Promotion rate + demonstrated development evidence (performance-review content analysis) is better than promotion rate alone. The combination resists Goodhart drift.

The standing review rhythm

The KPI tree is a living document with a standing review rhythm.

• Weekly at programme level. The programme team reviews the Level-3 metrics. Anomalies are investigated; short-term interventions are decided; the tree is not modified.
• Monthly at steering level. The transformation steering group reviews Level-2 drivers and the aggregate view of Level-3 metrics. Intervention adjustments are decided; the tree is modified only on explicit decision.
• Quarterly at board level. The board sees Level-1 outcomes with Level-2 context. Strategic adjustments are made here; the tree’s design is reviewed annually at the year-end board meeting.
• Annually at full tree review. The full tree is re-examined: are the outcomes still the right outcomes; are the drivers still the right drivers; are the metrics still measuring what matters; has the organisation’s context changed enough that the tree should change. The annual review produces a new tree version.

The rhythm is protected. Organisations that skip weekly reviews lose early signal; organisations that skip annual tree redesign carry forward a tree that is progressively less useful.

Two real-world anchors

Kaplan & Norton Balanced Scorecard adapted to people and change

Robert Kaplan and David Norton’s Balanced Scorecard, introduced in the Harvard Business Review in 1992 and developed through subsequent books, provided the foundational grammar for multi-level performance-measurement frameworks. The people-and-change KPI tree in this article borrows its three-level structure from that tradition, adapted to the specific domain of workforce transformation. Source: Kaplan & Norton, Harvard Business Review 1992 and subsequent publications.

The lesson: the three-level tree is not a novel invention; it is a mature measurement-design discipline that has been adapted across domains. Organisations implementing it can reference the Balanced Scorecard literature for implementation patterns that have been tested for three decades.

MIT Sloan people-metric research

MIT Sloan Management Review has published extensive research on people-metrics including specific work on AI-era adaptations. The research documents both the challenges (the common failures above are consistent across organisations) and the patterns that work. Source: https://sloanreview.mit.edu/.

The lesson: the failures are predictable; the patterns that work are documented. An expert who reads the record before designing their tree can avoid the common failures; one who reinvents the measurement from scratch reproduces them.

Learning outcomes — confirm

A learner completing this article should be able to:

• Name the three levels of a people-and-change KPI tree and articulate the purpose of each.
• Identify the three (or four) standard Level-1 outcomes and defend the decomposition.
• List the drivers that typically sit below each outcome and the metrics that populate each driver.
• Wire metrics to data sources with documented provenance, owner, and refresh cadence.
• Recognise and resist vanity metrics at the top and proxy metrics at the bottom.
• Maintain the standing review rhythm (weekly programme / monthly steering / quarterly board / annual redesign).

Cross-references

• EATP-Level-2/M2.5-Art05-People-and-Change-Metrics.md — Core Stream people-and-change-metrics anchor.
• Article 15 of this credential — measurement beyond completion.
• Article 17 of this credential — sustainment rhythm (the tree’s review rhythm fits inside).
• Article 32 of this credential — equity as Level-1 or Level-2 (organisation choice).
• Article 34 of this credential — organisational readiness score (composite that can be a Level-2 driver).

Diagrams

• ConcentricRingsDiagram — outcome at centre (productivity / retention / engagement); drivers in inner ring; metrics in outer ring; visible hierarchy reveals the argument from metric to outcome.
• Matrix — metric × quality dimension (specific / reliable / valid / actionable); applied to a sample set to illustrate the discipline.

Quality rubric — self-assessment