AI Activity Is Not a Performance Score
AI tools create new traces of work. That does not make prompt counts, output counts, time in tool, or model-generated quality scores valid measures of employee contribution.
Do not approve individual AI activity as a routine input to pay, promotion, ranking, or discipline. Measure quality-adjusted workflow outcomes at team or process level. A named employer data owner or controller should permit identifiable telemetry only for a concrete, necessary exception after the applicable legal, privacy, labor, security, accommodation, and human-review conditions are settled.
The sign-off test
Measure the work
Use completed outcomes, quality, rework, service levels, exceptions, and durable completion. Activity volume is an operating diagnostic, not an outcome.
Map the permission
Name the purpose, authority, jurisdiction, data fields, necessity, proportionality, notice, access, retention, review, and challenge route before collection or repurposing.
Prove the refusal gate
If aggregate evidence can answer the decision, or a required permission is missing, the named decision owner refuses the individual field or stops the proposed use.
Owner, briefing, proof
Owner
The employer data owner or controller, with the legal, privacy, HR, labor, security, accommodation, and operational authorities required by the actual use case.
Briefing
The exact role, workflow, outcome, data fields, jurisdiction, affected groups, decision, aggregate alternative, exception path, and prohibited uses.
Proof
A baseline and comparison tied to work outcomes, plus permission records, access logs, human review, correction and appeal receipts, and deletion or de-identification at close.
What leaders should take from it
Two meta-analyses report no overall performance relationship from electronic monitoring, alongside small stress and satisfaction effects. This is broad-monitoring evidence, not an AI-telemetry trial.
Field evidence shows meaningful gains in a defined support workflow. Randomized access evidence also changes work-pattern counts. Neither result turns tool activity into a measure of employee value.
Collection, repurposing, automated decisions, impact assessment, discrimination, accommodation, notice, and consultation rules vary by jurisdiction and use. One compliant use does not authorize another.
When visible activity affects status, people have reason to optimize the metric, hide unfinished experimentation, and shift effort toward what is countable.
Individual evidence may be necessary for a defined safety, security, accommodation, regulated-quality, or employee-requested case. The case needs separate authority, minimum data, a time limit, named reviewers, and a close-out rule.
No reviewed source validates prompt counts, output counts, time in tool, or model-generated quality scores as durable measures of employee contribution across roles. No reviewed study directly compares individual AI telemetry with aggregate workflow measurement on performance, trust, gaming, fairness, and appeal outcomes.
The findings in full
Ravid and colleagues synthesized 79 sources and 94 independent samples. Siegel, König, and Lazar synthesized 63 studies and 70 independent samples. Both reject a general performance benefit, while the second reports small adverse satisfaction and stress relationships.
Brynjolfsson, Li, and Raymond found a 15% average gain in issues resolved per hour among 5,172 support agents. Microsoft randomized access across 58 organizations and 6,317 employees, but its work-pattern counts do not map directly to performance or business outcomes.
The GDPR constrains purpose, lawful basis, minimization, repurposing, automated decisions, and high-risk processing. The EU AI Act conditionally classifies some workplace monitoring and evaluation systems as high-risk. US disability and local automated-employment rules apply more narrowly.
Manager surveys, worker-monitoring cases, and the history of metric-driven management support the mechanism. Direct AI-specific causal comparison is still missing, so the claim remains bounded.
Official guidance supports purpose, necessity, least-intrusive means, human review, challenge, and retention limits. Comparative evidence does not yet show that these controls reliably prevent an exception from becoming routine surveillance.
First moves before using employee AI telemetry
The measurement owner uses team or process outcomes for ordinary management and treats activity counts only as operating diagnostics.
The named employer data owner or controller records purpose, authority, necessity, proportionality, jurisdiction, fields, retention, access, affected groups, and why aggregate evidence does not suffice. If necessity fails, that owner refuses collection.
The applicable legal or privacy owner requires notice, consultation where required, discrimination and accommodation review, meaningful human authority, challenge and correction, access logging, and prohibited uses. A missing condition stops the proposed use.
The experiment owner pre-registers one quality-adjusted outcome, a baseline, a comparison where feasible, rework, exceptions, review cost, and adoption. No employee field is added without a named unanswered decision.
The employer data owner and applicable legal, privacy, security, labor, or accommodation owner decide the minimum data, time limit, reviewers, appeal route, and deletion rule. Repurposing requires a new decision.
Choose one workflow already under measurement. Define the completed outcome and the aggregate evidence first. If the decision can be made without identifying an employee, recommend that the named owner refuse the individual field.
This is general decision-process guidance, not legal or employment advice. The applicable employer, controller, legal, privacy, labor, HR, security, accommodation, and operational authorities decide what may be collected and used for a specific entity, workforce, role, system, jurisdiction, and purpose.
Claim ledger
Do not assert that individual AI-use telemetry predicts valuable employee performance, or that aggregate measurement has already won a direct causal comparison. Revisit this conclusion when a peer-reviewed field study compares both approaches on quality, trust, gaming, fairness, and appeals; when a regulator or court rules directly on generative-AI employee telemetry; or when audited enterprise evidence links an employee-level metric to quality-adjusted value without disproportionate error or workforce harm.