AI measures. You still judge.

How leaders can use AI to monitor performance better — with a few vital indicators instead of dashboard sprawl, and self-regulation instead of surveillance. The final judgment remains yours.

Monitoring is the unloved leadership task. It gets confused with mistrust, occasionally with bureaucracy, often with surveillance. It is none of these. Monitoring is the disciplined assurance that results are actually being produced — without restricting the autonomy of those doing the work. What changes now that AI is embedded in reporting systems is that measuring becomes radically cheaper, while judging, intervening, and assuming responsibility do not.

Why performance monitoring stays a leadership task — even with AI

A leader who runs a function without knowing where it stands is flying blind. Refusing the task because it is uncomfortable means refusing part of your leadership responsibility. A note on terms: monitoring as a leadership task is not the same as controlling in the financial-function sense, not technical system monitoring, and not performance management as a process scaffold. Those functions provide material — they do not replace the leader's own monitoring work.

The foundation of this task is trust in two things: in the capability of your people, and in their willingness to deliver. If either is missing, you do not have a monitoring problem — you have a staffing problem. That distinction shifts the question from monitoring intensity to people-related decisions and relieves monitoring of the burden of compensating for missing prerequisites.

On that foundation, the task breaks into three steps that belong together and routinely get conflated:

Measure — collect objective data: revenue, quality, lead times, complaint rates, on-time delivery. This can be delegated — including to AI.
Judge — interpret the data: what does it mean, what consequences follow? Subjective in that no formula can do it — but not arbitrary. And it stays leadership work.
Act — intervene where intervention is warranted, and explicitly do not intervene where self-correction will carry. This calls for judgment, proportion, and experience.

The design goal of this task is self-regulation, not surveillance. Drucker laid this out early in Management by Objectives and Self-Control. The cybernetic management tradition (Stafford Beer's Viable System Model) builds organizational viability on self-regulation: each unit has its own steering authority within a clear frame, and higher-order systems intervene only where self-regulation fails. From that follows a different KPI architecture: visibility goes first to the people doing the work — not to the level above doing the watching. If data flows primarily upward, the same technology turns into a surveillance architecture. The difference doesn't lie in the data; it lies in who sees it first.

Balanced Scorecard, OKRs, or key result areas: which monitoring system holds up

Before any choice of framework, the older question: What actually needs to stay in view in this function for it to remain viable? Three lines meet in international leadership practice:

Approach	When it works best	Risk if misapplied
Balanced Scorecard (Kaplan/Norton: financial, customer, process, learning)	Structured oversight when strategy is clear and cause-effect hypotheses hold	In practice often a KPI grab-bag without the BSC logic — dashboards with thirty metrics no one reads
OKRs (Objectives and Key Results, Silicon-Valley tradition)	Tech-adjacent functions and startups with short cycles and high autonomy	In a traditional industrial setting, OKRs become bureaucracy when the self-control architecture is missing
Drucker's key result areas (market standing · innovation · productivity · people · liquidity · profitability)	Supports a holistic view of viability — no single indicator alone supports a sound conclusion about the state of the function	Low risk; the trap lies in adoption — the framing is sharp but rarely held with discipline

None of these defects is fixed by switching frameworks. Moving from Balanced Scorecard to OKRs typically delivers the same symptoms under a new label after three years. What carries is not the choice of method, but the discipline with which a few vital indicators — three to seven, with five as a working anchor — are identified, maintained, and kept clean of the act of judgment. And the discipline not to lose sight of what cannot be measured. Drucker put the filter in two sentences: „That we can quantify something is no reason for measuring it" — and: „Any organization has important results that are incapable of being measured."

Performance monitoring with AI: what speeds up, what stays human

AI does not change the character of the task — it shifts the distribution of work. What used to take a controlling function two weeks each month now runs continuously in the background — as long as the data foundation is solid. What does not shift is judgment, conversation, and responsibility.

What AI handles reliably

Aggregating and cleaning data — consolidation across source systems, consistency checks, anomaly detection in large data volumes.
Attention briefings instead of 60-page reports — one page per week: what's running, what's deviating, what needs your judgment. The data flood is condensed to the few things that warrant attention.
Sampling instead of full coverage — expenses, travel costs, supplier invoices, compliance reviews: AI samples on a risk-weighted basis, flags anomalies, and performs deeper audits. Effort drops sharply.
Indicators for the unmeasurable — sentiment patterns from customer conversations, early-warning signals in retention, reputation patterns. What Drucker called „most important and yet least measurable" doesn't become measurable — but indicators become more visible.
Machine and sensor data made usable — predictive maintenance, dynamic process-parameter optimization, quality deviations spotted earlier than classical statistical process control allows.
Open commitments tracked automatically — AI extracts open promises from emails, calls, and meeting transcripts, sets reminders, and surfaces recurring patterns. A discipline many leaders struggle to sustain without it.

What stays with the human

Choosing the few vital indicators — which metrics actually reflect the function's viability is a leadership decision. AI can suggest; the choice belongs to whoever knows the function, the strategy, and the responsibility.
Judging, not just measuring — AI delivers data; what it means and what consequences follow is a judgment. Subjective in the sense that no algorithm does it — but not arbitrary. Experienced people applying the same standards arrive at consistent (not identical) judgments.
Going to the source — reports contain only what can be described; what can be perceived is much larger. The more delicate, business-critical, or new a situation, the less reporting alone carries. Alfred Sloan stood behind sales counters several times a year — no report could replace the impression of an actual sales conversation.
Benevolent oversight — not every deviation needs to be named, not every weakness raised. Some things resolve with time. This is not the opposite of monitoring; it is its mature form — the deliberate decision not to react now, in clear knowledge that you will look once it becomes necessary.
Carrying the consequences — defending a performance assessment by pointing at a system is not an explanation. The leader who monitored stands behind the judgment.

The real risk is not faulty AI math. It is the distortions AI introduces into monitoring practice: Automation bias — flagged anomalies get over-weighted, missed situations slip out of view. Bias at scale — a judgment error potentially affects everyone evaluated by the same system, simultaneously and continuously. Goodhart effects — when a metric becomes a target, it ceases to be a good metric. Surveillance creep — what is technically measurable (keystrokes, screen time, break patterns) is not what should be monitored. On top of this, there are regulatory guardrails: the EU AI Act treats AI in personnel evaluation as a high-risk use case (most Annex III obligations apply from 2 August 2026); EU member states with strong co-determination regimes (Germany under §87 BetrVG, similar frameworks in Austria and the Netherlands) require employee representation when behavior or performance becomes individually traceable; non-discrimination law (Germany's AGG, EU-level equivalents) also covers indirect discrimination through proxy variables; GDPR governs purpose-limitation, minimization, and transparency for any employee data processing.

You want AI in your reporting — without your leaders becoming data jugglers or your people becoming surveilled?
Let's talk about your leadership team →

Better monitoring with AI: coaching, leadership development, and workshops

I work on this topic in three formats: in executive coaching, in leadership development programs, and in leadership workshops for intact teams — from mid-cap to enterprise, anchored in Drucker's management thinking, the cybernetic tradition of self-regulation, and the regulatory guardrails for AI use. Some of the underlying frameworks I share at international leadership events, including the ATD APC Taipei 2025. „Monitoring" is rarely booked as a training topic. It surfaces when you look behind reporting fatigue, performance-management redesigns, or AI pilot programs. Five questions sit at the center of every format:

What are the few vital indicators of your function? Three to seven, with five as a working anchor — derived from the key result areas, filtered by „What do I absolutely need to know to feel justifiably at ease?". Everything else is at best nice to know.
Where is measuring being conflated with judging? The separation is the most common source of unsound performance reviews. Data has its place, judgment has its — they cannot stand in for each other.
Is your reporting action-oriented or information-oriented? The distinction — what should people do (action-oriented) versus what do we want to know about them (information-oriented) — decides whether your metrics produce self-control or reporting overhead.
How do you use AI without slipping into surveillance? Which analyses are subject to co-determination, which fall under the EU AI Act, which are sensitive under GDPR — and which design choices protect the self-control architecture?
How do you monitor the AI agents in your processes? Once agents run inside workflows, they belong inside your monitoring scope: success rate, tone, boundary discipline, drift after model updates.

What you take away is not a new method — but a sharper view of what actually needs to be monitored in your function, and which reports and routines you should drop, keep, or rebuild under AI conditions.

7 practical routines for better monitoring with AI

Small habits you can start immediately.

1. The three-to-seven rule and an annual monitoring audit (half a day)

Once a year, two questions together. First: What are the three to seven indicators I genuinely need to steer my function? For each: Does it have a counter-metric that prevents one-sided optimization? (Goodhart protection: closing rate needs 90-day customer satisfaction). Second: Which reports, dashboards, and reviews am I retiring? AI makes it easy to keep everything alive — the leadership decision is to retire what no longer adds value.

2. Quarterly viability scorecard (60 minutes, alone or with a small team)

Once per quarter, walk explicitly through the key result areas: market standing, innovation, productivity (labor, capital, time, knowledge), attractiveness to good people, liquidity, profitability requirement. AI delivers the data; the judgment of the overall picture is yours. Never infer the state of the function from a single indicator — sound liquidity with eroding substance, healthy profit with slipping market position, high productivity with talent flight. Only the interplay yields a correct picture.

3. Weekly attention briefing with bias check (15 minutes, Mondays)

AI summarizes one page: green, attention recommended, intervention required. For each flagged alert, a reflex question: Would I see this if I looked at the raw data? For critical alerts: actually look at the raw data before reacting. Discipline matters — once the briefing grows to two pages, it reverts to a mini-report. Tracking over weeks how often a raw-data check relativized an alert calibrates trust in the system more accurately than any vendor specification.

4. Monthly Gemba walk (half a day, no dashboard)

Once a month, deliberately go into a function without dashboard or screen — production floor, sales, service, engineering, field. Gemba walk, genchi genbutsu in the Toyota tradition, MBWA in the US sense — the content is the same: go, look, talk to the people doing the work. What you perceive there — mood, tone, unspoken conflict, quiet exhaustion — fits no report and is exactly why it is monitoring-relevant. This routine is the antithesis to the reporting reflex and the strongest defense against automation bias.

5. Measure-vs-judge separation in every review

In every assessment, deliberately separate the facts — measured, documented, and verifiable from your judgment — your interpretation and conclusion. Both have a place; neither may stand in for the other. AI provides the factual basis; you carry the judgment. The conversation itself — human to human — is replaced by no AI, in no form, in no setup.

6. Open-loop hygiene with AI support (5-minute daily check)

A list of all open commitments, agreed tasks, deadlines. The daily check ensures: nothing agreed gets forgotten. What is not done is not done because a decision was made against it — never because it slipped through. AI extracts commitments from emails and meeting notes, sets reminders, and surfaces recurring patterns. Consistently lived, this routine teaches the environment that nothing falls through — a control effect no dashboard achieves.

7. Agent health check (per agent, frequency by usage intensity)

Per AI agent, four questions: Has it understood the tasks correctly? Are responses relevant, precise, on tone? Were boundaries respected? Where does the system prompt need updating? Plus drift control: prompt versioning, fixed baseline test suite, regression tests after model or tool updates. AI agents change over time even if no one changes the prompt — because of vendor model updates, new retrieval data, and prompt erosion from accumulated edge-case patches. The question „How is the team?" extends to „How is each agent running in our processes?".

Leadership development, coaching, keynote — how we work together

Leadership development & workshops

For leadership teams recalibrating their monitoring practice under AI conditions — anchored in your real reporting cycles, performance review formats, and KPI architectures, not generic cases.

See programs →

Executive coaching

For individual leaders looking to slim down reporting, sharpen performance reviews, or redraw the surveillance threshold in their function — confidential, with a sparring partner who has no stake in the outcome.

Schedule a conversation →

Keynote

For leadership conferences opening the discourse on monitoring, trust, and AI use. No hype show — an honest look at what leadership requires in the AI era.

Request a keynote →

Common questions on monitoring, KPIs, and AI in reporting

How many KPIs does a function really need?

Three to seven, with five as a working anchor. Drucker put it early: an experienced executive pulls a handful of indicators out of every voluminous controlling report and runs the function with them — the rest is data, not information. Dashboards with twenty or thirty metrics are not a sign of strong monitoring; they signal that the selection was never narrowed down. AI lowers the barrier to capturing arbitrarily many indicators — the leadership decision is not to be tempted by it.

When does monitoring become surveillance — and when is it still leadership?

The threshold isn't the volume of data, it's the question of who sees it first. Action-oriented monitoring — your people see their own indicators first and steer themselves — is leadership. Information-oriented monitoring — data flows primarily upward, the level above watches — is surveillance with a steering label. Add the distinction outcome-based versus behavior-based monitoring: outcome-based is the default, behavior-based (keystrokes, break patterns, processing steps) is the justified exception for safety-critical and heavily regulated areas, not a general approach.

My dashboard has thirty metrics, no one looks — what now?

Don't accelerate it, retire it. Before any optimization, run the monitoring audit: which of these do I genuinely need to feel justifiably at ease? Three to seven survive — the rest gets cut, not digitized. AI makes it easy to keep everything alive; the leadership decision is to enforce the expiration date. For each surviving metric, two Goodhart questions: Does it have a counter-metric to prevent one-sided optimization? Can I explain it to a new colleague in five minutes?

How do I use AI in performance reviews without violating the EU AI Act?

This is general information, not legal advice. The EU AI Act classifies AI systems used in personnel management (selection, evaluation, promotion, termination) as high-risk under Annex III; most obligations apply from 2 August 2026. In practice: AI-assisted recommendations are permissible; the assessment must remain human — and must be defensible without invoking the system. The judgment goes into your own words, not copy-paste from AI output. On top of that, high-risk obligations include risk management, documented data quality, logging, human oversight, and discrimination testing — most of which align with EU member-state co-determination and equal-treatment frameworks anyway.

Do I need employee representation involvement for AI-supported analytics?

In most EU jurisdictions with co-determination: yes. Germany's §87 BetrVG requires the works council to be involved when technical systems are objectively suited to monitor employee behavior or performance. Once data is individually traceable or behavior/performance becomes personally evaluable, the threshold is typically crossed — AI-based analysis of tickets, customer communication, code contributions, or processing times usually falls under it. Aggregated, non-traceable analyses can sit differently. Similar frameworks apply in Austria, the Netherlands, and other co-determination regimes. Introducing this without employee representation is hard to repair later — clarify it early to save friction.

How do I monitor experienced people without demotivating them?

Monitoring intensity is a question of knowledge, not trust. With proven performers, monitoring is lighter — close oversight here is unnecessary and demotivating. With newer people or unclear performance patterns, it's tighter — here it acts as orientation, not suspicion. The right intensity follows the situation, not an attitude. Hersey and Blanchard's situational leadership operationalizes this: assess task readiness, derive style and monitoring density. High readiness means more delegation, low readiness more accompaniment. The full operationalization belongs in the people-development task.

How do I monitor AI agents running inside our processes?

Like human team members — applying the same Drucker principles (appropriate, no false precision, focused on a few vital indicators, structurally valid). Concretely: success rate, escalations, tone, boundary discipline. Plus a domain-specific dimension: drift. AI agents change over time without anyone changing the prompt — vendor model updates, new retrieval data, prompt erosion from accumulated edge-case patches. The operative answer: prompt versioning, a fixed baseline test suite, regression tests after model or tool updates. For each agent, a single human accountable owner must be visible — black-box agents without owners do not belong in production processes.