Practical guide

Workforce monitoring policy for AI-enabled work

A workforce-monitoring policy should constrain a specific monitoring practice before data is collected. It should not provide a blanket right to inspect activity because a tool makes it possible. For AI-enabled work, the policy must address source data, derived inferences, prompt or content access, automated analysis, individual consequences, and secondary uses as well as ordinary logs. Use.

By Two Prune

A workforce-monitoring policy should constrain a specific monitoring practice before data is collected. It should not provide a blanket right to inspect activity because a tool makes it possible. For AI-enabled work, the policy must address source data, derived inferences, prompt or content access, automated analysis, individual consequences, and secondary uses as well as ordinary logs. Use GUARD to design the policy: define the goal and authority, use the least intrusive data, assign access and action limits, provide rights and redress, and date review and shutdown triggers. This is an educational drafting framework. Current jurisdiction-specific legal, employment, data-protection, security, and worker-representation review is still required.

G: Define one goal, authority, and prohibited purpose

State the monitoring purpose in one sentence and name the decision it supports. Separate security, service reliability, capacity planning, learning support, regulatory duties, and individual performance. If one system serves more than one purpose, create a separate policy entry and approval for each purpose.

Name the accountable owner and the authority relied on after specialist review. Record the population, systems, locations, start date, duration, and whether suppliers are involved. Add an explicit prohibited-purpose clause. For example, an aggregate AI-assistant adoption dashboard designed to allocate support must not be reused to rank individuals or infer work quality.

The UK Information Commissioner's Office says monitoring purposes should be clear and documented, and that policies should describe the nature, purpose, and extent of monitoring. It also warns that documenting excessive monitoring does not make it lawful. Treat the policy as a constraint, not a cure.

U: Use the least intrusive data and define every inference

List each collected field, derived field, content source, identifier, sampling rule, frequency, and retention period. For every field, explain why a less intrusive alternative cannot answer the stated question. Distinguish metadata such as feature activation from content such as prompts, documents, screenshots, messages, or recordings.

Define every inference the system produces. Prompt volume is not work quality. Active days are not capability. Time in an application is not productivity. An AI-generated risk flag is not a verified event. If the evidence does not support an inference, prohibit it rather than adding a disclaimer after deployment.

Set minimisation rules in executable terms: aggregate at team level, suppress groups below a threshold, exclude content, mask identifiers, sample only when necessary, and delete raw records after a defined interval. Document the source cut-off and accuracy limitations.

A: Assign access, action, and automation limits

Use a role matrix that states who can see raw data, aggregates, content, identifiers, exports, and audit logs. Separate system administration from managerial use. Require named approval for exceptional access and record the reason, scope, time, and reviewer.

Describe what actions may follow from the data and what actions are prohibited. If monitoring information can contribute to an individual decision, state the human review, corroborating evidence, opportunity to respond, and escalation route required before action. Do not let an AI-generated score or summary silently become an adverse conclusion.

Cover suppliers and integrated tools. The ICO notes that organisations remain responsible for deciding how and why monitoring occurs when third-party technology is used. Record processor responsibilities, configuration, data flows, locations, subprocessor changes, deletion, incident handling, and audit access.

R: Provide notice, consultation, correction, and redress

Tell affected workers what is collected, why, how it is used, who receives it, how long it is kept, and what choices or rights apply. Make the notice match actual configuration. A policy hidden in an employee handbook does not explain a new inference or content-capture feature.

Consult workers or representatives during design unless specialist advice identifies a justified exception. Acas recommends consultation before introducing monitoring and a clear policy explaining the arrangement and its reason. Record feedback, changes made, unresolved objections, and the reason for any departure.

Provide a usable route to ask questions, access relevant information, correct an error, challenge an inference, and report misuse. State response owners and time targets. Protect the integrity of security or investigation processes without turning that need into a general exception from transparency.

D: Date tests, reviews, changes, and shutdown triggers

The policy should not remain valid indefinitely. Record pre-launch tests for permissions, aggregation, suppression, notice, retention, export, correction, and deletion. Set an owner and review date based on risk and change.

Require reassessment before adding a data source, AI model, inference, purpose, population, supplier, location, access role, consequence, or retention period. Define incident and shutdown triggers. Examples include unauthorised content access, an inference used for a prohibited purpose, broken suppression, failed deletion, material notice mismatch, or inability to process a correction.

Preserve policy versions and approvals. A new dashboard configuration can invalidate an earlier policy even if the document title remains unchanged.

Copyable policy artifact

Policy ID and version: [identifier, version, approval date]

Purpose and decision: We monitor [system or activity] solely to support [bounded decision]. We do not use this information for [prohibited purposes].

Scope: This policy covers [population, systems, locations, period]. It excludes [systems, content, people, or contexts].

Data and inferences: We collect [fields] at [frequency] from [source]. We derive [inferences]. We do not collect [content or fields] or infer [unsupported conclusions].

Minimisation and retention: We apply [aggregation, suppression, masking, sampling]. Raw data is retained for [period], derived records for [period], and deletion is verified by [owner and test].

Access and actions: [roles] can access [views] for [permitted actions]. [prohibited actions] are not allowed. Exceptional access requires [approval, record, expiry, review].

Automation and human review: Automated processing may [bounded function]. It may not [decision or action]. Before an individual consequence, [human review and corroboration] are required.

Notice and consultation: Affected people receive [notice route and timing]. Consultation occurred with [groups] on [date], with [outcome].

Questions, correction, and challenge: Contact [route]. [owner] handles requests and challenges under [process and timing].

Suppliers and transfers: [supplier roles, data locations, subprocessors, controls, change notice].

Testing and review: Pre-launch tests are [list]. Next review is [date]. Reassessment triggers are [changes]. Shutdown triggers are [failures].

Approvals: [operational owner, privacy, legal or employment, security, worker-representation route, other required authorities].

Worked example: an AI-assistant support dashboard

A synthetic organisation wants to identify teams that need help adopting an approved AI assistant. The first proposal collects named prompt counts, prompt text, active minutes, and manager rankings. GUARD changes the design before launch.

The bounded purpose becomes allocating optional support materials by team. The policy excludes performance management, capability judgments, attendance, and disciplinary use. It retains weekly team-level feature activation and support requests, suppresses groups smaller than ten, excludes prompt content and active minutes, and deletes raw identifiers after aggregation. Managers receive only the aggregate view; named records remain inaccessible through their role.

Workers and representatives review the notice and identify that contractors are grouped separately, producing small cohorts. The team changes the grouping and records the limitation. Pre-launch testing then finds that an administrator export can bypass suppression. The launch is paused until the export role is removed, logs are reconciled, and the suppression test passes. The policy's shutdown trigger makes this a required response rather than an informal preference.

Verification and Two Prune boundary

A policy draft is ready for approval only when every data field, inference, recipient, permitted action, prohibited action, retention rule, rights route, supplier, test, change trigger, and owner has a disposition. Approval does not prove the control operates; operating evidence must later test practice against the policy.

Two Prune can place a participant in a synthetic policy-design task and capture evidence of source use, trade-off analysis, AI assistance, verification, judgment, and communication. It does not provide legal advice, approve monitoring, determine compliance, operate a monitoring system, or make employment decisions.

Primary sources

Information Commissioner's Office, Data protection and monitoring workers

  • Information Commissioner's Office, Guide to accountability and governance
  • Acas, Monitoring performance for employees who work from home

Sources: [1] [2] [3]