Practical guide

Privacy impact assessment checklist for AI-assisted work

Use this checklist to test whether a privacy impact assessment is decision-ready. It converts the assessment into observable completion gates: trigger, purpose, people, data flow, alternatives, risk pathways, controls, consultation, sign-off, and review. A checked box needs evidence; it is not a substitute for professional judgment or applicable legal analysis. The checklist is educational..

By Two Prune

Use this checklist to test whether a privacy impact assessment is decision-ready. It converts the assessment into observable completion gates: trigger, purpose, people, data flow, alternatives, risk pathways, controls, consultation, sign-off, and review. A checked box needs evidence; it is not a substitute for professional judgment or applicable legal analysis. The checklist is educational. Adapt it to the jurisdiction, processing context, organization, and people affected. A completed copy does not establish legal compliance or prove that a control operates.

How to use the checklist

Freeze one proposal version and name an accountable owner. For each item, mark complete, incomplete, not applicable with reason, or unknown. Add an evidence link, owner, and due date. Unknowns that could change the decision remain open; they are not passed because a deadline is close.

The UK Information Commissioner's Office publishes DPIA screening, process, and quality checklists covering the need for an assessment, processing description, processor input, consultation, necessity and proportionality, risk to individuals, measures, disagreements, integration into the project plan, and review. Its guidance is currently under review following legal changes, so users should check the live source and applicable law. ICO: Data protection impact assessments

Sources: [1]

The CLEAR checklist

### C — Confirm the trigger, purpose, and authority

Evidence: proposal record, screening note, role map, decision-rights record, change rule.

Hold when: the purpose keeps expanding, authority is inferred from a title, or the team cannot identify the decision being assessed.

### L — List people, data, and the complete flow

Evidence: data inventory, system diagram, provider documentation, configuration capture, access list, retention and deletion test.

The NIST Privacy Framework is a voluntary enterprise-risk tool for identifying and managing privacy risk. Its functions and outcomes can help a reviewer check governance, data processing, communication, control, and protection coverage. Framework alignment does not answer a local legal question or demonstrate local control effectiveness. NIST Privacy Framework

### E — Examine necessity, alternatives, and risk pathways

Evidence: alternative analysis, necessity table, risk-pathway register, source citations, residual-risk record.

Hold when: a color matrix replaces causal explanation, the preferred option is compared only with an artificially weak baseline, or the assessment treats privacy solely as cybersecurity.

### A — Assign controls and test their evidence

Evidence: configuration, access test, deletion test, sampled review record, incident exercise, exception log, corrective action.

### R — Record consultation, disposition, and review

Evidence: consultation log, dissent record, signed decision, delivery actions, monitoring plan, reassessment schedule.

The ICO's detailed process guidance says views of individuals or representatives should be sought and documented unless there is a good reason not to, and that necessity and proportionality should consider whether another reasonable route can achieve the purpose. This supports explicit consultation and alternative-analysis gates; applicability still depends on the actual processing. ICO: How do we do a DPIA?

  • [ ] The exact proposal, version, environment, and decision deadline are named.
  • [ ] The business purpose is specific enough to test; attractive secondary uses are excluded.
  • [ ] The screening decision and reasons are recorded, including why a full assessment is or is not required.
  • [ ] The product owner, privacy lead, security reviewer, data owner, operational owner, and final decision maker are named.
  • [ ] The final approver has authority to proceed, narrow, redesign, hold, or stop.
  • [ ] The allowed AI contribution and the decisions that remain human are explicit.
  • [ ] Material changes that reopen the assessment are listed.
  • [ ] Every group of people represented in the data is listed, including people mentioned inside free text or attachments.
  • [ ] Reasonable expectations, notice, choice, access, correction, objection, and escalation routes are described where relevant.
  • [ ] Every input field and source is inventoried with purpose, sensitivity, accuracy concern, and necessity.
  • [ ] The flow covers collection, transfer, prompt, retrieval, model processing, output, logs, feedback, human notes, storage, access, support, recipients, retention, deletion, and backup behavior.
  • [ ] Provider, model, tenant, region, subprocessor, integration, and configuration are versioned.
  • [ ] Derived data and inferences are included, not only raw inputs.
  • [ ] Unknown vendor or system behavior is labelled unknown and assigned for resolution.
  • [ ] Each data element and transformation is connected to the stated purpose.
  • [ ] The assessment compares no change, process repair, a non-AI route, a narrower AI route, and the proposed design where credible.
  • [ ] Data minimization, lower precision, shorter retention, smaller population, local processing, and separation have been considered.
  • [ ] Each material risk identifies a trigger, action, affected person, possible effect, existing control, likelihood, severity, and uncertainty.
  • [ ] Risks cover collection, processing, inference, disclosure, access, decision, contestability, exclusion, lifecycle drift, and deletion.
  • [ ] Risks to people are not collapsed into organizational inconvenience.
  • [ ] Benefits and mitigations are supported by evidence rather than provider adjectives.
  • [ ] Residual risk after controls is recorded separately from inherent risk.
  • [ ] Every control has an owner, implementation evidence, operating evidence, failure signal, response, and review date.
  • [ ] Access, tenant isolation, encryption, minimization, retention, deletion, logging, human review, correction, incident response, and change monitoring are addressed where relevant.
  • [ ] The test method uses a representative case, not only the easiest path.
  • [ ] The control can fail visibly and has an escalation route.
  • [ ] A policy statement is not accepted as proof that the control operated.
  • [ ] Controls that depend on people include time, competence, authority, and access to source evidence.
  • [ ] Unresolved high-risk conditions are routed to the appropriate specialist or authority before processing.
  • [ ] Affected people or appropriate representatives were consulted where feasible, or a specific reason for not consulting is recorded.
  • [ ] Privacy, security, accessibility, frontline, data-governance, and operational perspectives are included as relevant.
  • [ ] Feedback, disagreements, and resulting design changes are visible.
  • [ ] The final disposition is proceed, proceed with conditions, redesign and reassess, hold, or stop.
  • [ ] Conditions, residual risks, accepting authority, and unresolved evidence are explicit.
  • [ ] Measures are integrated into the delivery plan with owners and dates.
  • [ ] Monitoring, incident, model, provider, data, purpose, population, integration, and retention changes trigger review.
  • [ ] The signed record and evidence are retained under an approved record rule.

Sources: [2] [3]

Copyable completion record

| Gate | Status | Evidence | Owner | Due date | Decision effect | | --- | --- | --- | --- | --- | --- | | Confirm | | | | | | | List | | | | | | | Examine | | | | | | | Assign | | | | | | | Record | | | | | |

Add a one-line conclusion: The proposal may proceed / proceed with conditions / must be redesigned / is held / is stopped because… Link the reason to the evidence and accepting authority.

Representative test 1: employee shift-swap assistant

A service organization proposes an assistant that recommends shift swaps using availability, qualifications, location, and absence records. The checklist exposes two missing flows: free-text absence reasons are sent to the provider, and generated recommendations are retained in support logs. The purpose can be met without absence reasons, and qualification matching can use approved codes.

The team removes free text, restricts the data fields, adds a human confirmation step, documents the provider setting, tests deletion, and gives employees a correction route. A sampled case reveals that an outdated qualification still blocks a valid swap, so data correction becomes a release condition. The disposition is proceed with conditions for a narrow test. The record does not claim that the assistant improves scheduling or satisfies every applicable obligation.

Representative test 2: customer-support call summarizer

A support team proposes broad recording and summarization of every call. The checklist finds that the population includes people discussing payment difficulty and health-related service needs; notice is inconsistent; the provider's support access and deletion behavior are unresolved; and the team has not compared a structured post-call form.

The structured form could meet the coaching purpose without retaining audio. The broad design is marked hold while the team narrows the purpose, consults relevant representatives, obtains provider evidence, and tests a lower-data alternative. The test demonstrates why “all boxes checked” is not the goal: unresolved evidence changes the disposition.

Quality assurance and failure conditions

Have a reviewer who did not write the assessment trace a risk from data field to person, effect, control, operating evidence, and disposition. Sample at least one ordinary case, one sensitive case, one access or deletion exception, and one material change. Reopen failed items rather than editing the conclusion alone.

Reject the checklist when entries are generic, links point only to policies, not-applicable decisions lack reasons, provider claims are copied without local verification, consultation is ceremonial, residual risk has no accepting authority, or the review rule cannot detect change.

Two Prune boundary

Two Prune can use realistic work simulations to assess how participants frame a privacy problem, navigate evidence, use AI inspectably, verify claims, control risk, exercise judgment, and communicate an assessment artifact. It does not conduct a customer's privacy impact assessment, provide legal advice, operate controls, certify compliance, or make employment decisions. Customers retain responsibility for processing, tools, data, specialist review, approvals, and use of assessment evidence.

Sources

  1. 1.ICO: Data protection impact assessments · Information Commissioner's Office
  2. 2.NIST Privacy Framework · National Institute of Standards and Technology
  3. 3.ICO: How do we do a DPIA? · Information Commissioner's Office