Practical guide
Privacy control policy for AI-enabled work
This template turns privacy intent into an operating policy for AI-enabled work. It defines permitted processing, accountable roles, control procedures, evidence, exceptions, tests, and review dates. Adapt it to the actual data, systems, people, purposes, contracts, and jurisdictions before approval.
This template turns privacy intent into an operating policy for AI-enabled work. It defines permitted processing, accountable roles, control procedures, evidence, exceptions, tests, and review dates. Adapt it to the actual data, systems, people, purposes, contracts, and jurisdictions before approval.
Legal and assurance boundary
The UK Information Commissioner's Office says data protection should be integrated from design through the processing lifecycle and that accountability requires appropriate technical and organizational measures, documented risk decisions, and evidence. Its AI guidance treats impact assessment as a way to identify and control risks rather than a box-ticking exercise. The NIST Privacy Framework provides a voluntary risk-management structure for identifying, governing, controlling, communicating, and protecting data processing. These sources support the policy structure; this template is not legal advice and does not establish compliance or control effectiveness. ICO: Data protection by design and default ICO: AI accountability and governance NIST: Privacy Framework
Sources: [1] [2] [3]
Policy record
Policy owner: [accountable role] Approved by: [authority] Effective date: [date] Version: [version] Applies to: [people, business units, systems, vendors, locations] Next review: [date or trigger]
1. Objective
The organization will use AI-enabled tools only for documented purposes and will apply proportionate controls to protect people and personal data throughout collection, preparation, transfer, use, output, review, retention, and deletion. Each use must have an accountable owner, approved data boundary, risk assessment, operating evidence, incident route, and review date.
This policy does not authorize a use merely because a tool is available. Stricter legal, contractual, security, records, sector, works-council, or customer requirements continue to apply.
2. Scope and inventory
The policy covers employees, contractors, vendors, integrations, models, agents, prompts, retrieval stores, evaluation datasets, logs, generated outputs, and human review where personal data may enter or be inferred.
The AI inventory must record the owner, purpose, people affected, data categories, source, legal or organizational basis, systems and vendors, locations, model and version, human role, recipients, retention, risk assessment, controls, incidents, and status. Uninventoried uses are not approved uses.
3. Permitted and prohibited data
Each use case must maintain an approved data schedule:
Operators must not paste personal data into a consumer or unapproved AI service, use real personal data for convenience testing, repurpose outputs without review, or assume redaction is complete because names are removed. Synthetic or properly anonymized data should be used when it can meet the task.
- permitted: the minimum named fields necessary for the documented purpose;
- restricted: fields allowed only under a recorded condition, approval, or protected route;
- prohibited: data that must not enter the AI tool or its logs;
- derived: classifications, scores, summaries, or inferences created by the system;
- output: who may receive, edit, rely on, or disclose the result.
4. Required operating procedure
Before activation, the use-case owner must document necessity and proportionality, alternatives, data flow, affected people, foreseeable harms, human roles, vendor terms, security review, retention, and applicable impact assessment. High or unresolved risk follows the organization's escalation and consultation route before processing begins.
At intake, the operator or automated control validates the source, approved purpose, field allowlist, consent or other required authority, and destination. Prohibited fields are blocked, not silently removed without a record. Transfer logs retain a reference, rule version, result, and exception identifier while minimizing copied personal data.
During use, access follows least privilege. Approved prompts, retrieval sources, connectors, and output destinations are versioned. Material model, vendor, purpose, population, data, or workflow changes trigger reassessment. Human reviewers receive instructions proportionate to the consequence and must be able to challenge, correct, defer, and escalate an output.
At output, reviewers check provenance, unsupported personal inferences, disclosure risk, accuracy appropriate to the use, and whether the result remains within purpose. Decisions with legal or similarly significant effects require the separately approved decision process and safeguards; this policy does not create that authority.
At closure, data and outputs follow the approved retention and deletion schedule. Legal hold, complaint, access, correction, objection, and deletion requests follow the organization's established rights process. Deletion claims require evidence from all relevant stores and vendors.
5. Roles and accountability
| Role | Accountable activity | | --- | --- | | executive sponsor | approves objective, resources, and residual risk within authority | | use-case owner | purpose, inventory, operation, evidence, and corrective action | | privacy owner or DPO | independent advice, assessment route, rights and escalation | | security owner | access, integration, monitoring, incident coordination | | data owner | data schedule, quality, permitted use, retention | | vendor owner | due diligence, terms, changes, deletion and exit evidence | | operator | follows procedure and records exceptions | | reviewer | verifies outputs and escalates uncertainty | | assurance provider | tests within an agreed scope without owning the control |
One person may hold several roles in a small organization, but accountability, operation, and independent challenge should not be presented as separate when they are not.
6. Exceptions and incidents
An exception request must name the rule, reason, data and people affected, duration, compensating controls, risk owner, approver, evidence, and expiry. Emergency approval expires automatically and must be reviewed after use. Repeated exceptions trigger a control-design review.
Suspected prohibited disclosure, excessive collection, unauthorized access, harmful inference, rights failure, vendor change, missing log, or control bypass must enter the incident route. Preserve evidence, contain the route, assess affected processing, involve required specialists, make notification decisions under applicable rules, and track corrective action. Do not delete evidence to make the event disappear.
7. Testing and evidence
The control owner defines a population, period, procedure, evidence, and failure rule for each control. Evidence may include inventory reconciliations, configuration records, access reviews, blocked transfers, override logs, sampled transactions, deletion confirmations, vendor notices, rights requests, incidents, complaints, and retests.
Tests must distinguish design, implementation, operation, and outcome. A written policy shows design; a configured block may show implementation; population testing shows operation for the tested period; neither alone proves that people experienced no harm. Record missing evidence and failed events rather than excluding them.
8. Review and expiry
Review at least [frequency] and after changes to purpose, data, population, model, vendor, integration, law, contract, control, incident pattern, or risk appetite. The review records what changed, evidence examined, exceptions, complaints, test results, unresolved risk, actions, approver, and next date. Historical approval does not authorize the changed use.
Worked policy record: synthetic assessment workspace
An organization uses an approved AI assistant inside a synthetic work assessment. The documented purpose is to observe how a participant frames a task, uses supplied sources, checks AI output, and produces a business artifact. The approved dataset contains a participant identifier, assessment events, prompts, supplied synthetic materials, notes, submission, reviewer evidence, and scores. Real customer records, special-category data, unrelated workforce records, and data copied into external AI services are prohibited.
The assessment owner is accountable for operation. Privacy reviews the data flow and rights route; security controls access; the assessment team operates the workflow; trained reviewers evaluate evidence. The transfer control allowlists workspace fields and blocks attachments outside the synthetic case. Each event records assessment ID, field class, rule version, destination, result, and exception ID. Reviewers see only records required for their assignment.
The monthly test reconciles all attempted transfers to completed, blocked, failed, or unresolved states; inspects every override and incident; and reperforms a risk-weighted sample. One month contains 806 attempts, 791 completed transfers, 12 expected blocks, two approved test overrides, and one unresolved retry. The unresolved event makes the population incomplete. The owner contains it, determines no payload left the approved boundary, repairs the status handler, and retests all failed and override routes.
The conclusion is operating with an evidence exception for the tested month until reconciliation closes. It does not establish general legal compliance, absence of other privacy risks, or operation after a system change. Two Prune remains an assessment product: the record supports observed assessment performance, not automated employment decisions.
Sources
ICO: Data protection by design and by default
- ICO: Accountability and governance implications of AI
- NIST: Privacy Framework
Sources: [1] [2] [3]
Sources
- 1.ICO: Data protection by design and default · Information Commissioner's Office
- 2.ICO: AI accountability and governance · Information Commissioner's Office
- 3.NIST: Privacy Framework · National Institute of Standards and Technology