Practical guide

How to create a privacy impact assessment for AI-assisted work

A useful privacy impact assessment is a decision record, not a generic risk list. It should let an accountable reader see the proposed purpose, the people and data involved, the full information flow, the necessity of the design, the risks to people, the controls selected, unresolved questions, and the person who accepted or stopped the proposal. This guide is educational and must be adapted to.

By Two Prune

A useful privacy impact assessment is a decision record, not a generic risk list. It should let an accountable reader see the proposed purpose, the people and data involved, the full information flow, the necessity of the design, the risks to people, the controls selected, unresolved questions, and the person who accepted or stopped the proposal. This guide is educational and must be adapted to the relevant jurisdiction, organization, and processing context. Privacy specialists and legal counsel should determine applicable obligations. The completed artifact does not itself establish legal compliance or control effectiveness.

Start with a decision boundary

Freeze the proposal before assessing it. Record the decision to be made, intended users, affected people, business purpose, allowed AI contribution, model and provider, deployment setting, information classes, source systems, recipients, retention, and the latest decision date. Name the product owner, privacy lead, security reviewer, data owner, operational reviewer, and final approver.

The UK Information Commissioner's Office describes a DPIA process that includes describing processing, consulting people where appropriate, assessing necessity and proportionality, identifying risks, specifying measures, and reaching a conclusion. Its current guidance is under review following the Data (Use and Access) Act, so teams should check the live guidance and applicable law rather than treating a copied template as permanent. ICO: Data Protection Impact Assessments

Use a clear change rule: if the purpose, population, data, provider, model, integrations, recipients, retention, or human decision changes materially, reopen the assessment. A label such as “internal tool” or “copilot” does not define the privacy boundary.

Sources: [1]

Use the MAPS assessment method

MAPS produces four connected artifacts: a mandate, a data-flow map, a risk-and-necessity register, and a signed disposition.

### M — Mandate, purpose, and people

Write the legitimate organizational purpose in plain language and identify how success will be judged without relying on extra personal data. Separate the intended purpose from attractive secondary uses. List the people whose information may be processed, including employees, customers, bystanders, correspondents, children or vulnerable people where relevant, and people represented inside attachments or free text.

Record what each person could reasonably expect, what notice or choice exists, how they can question or correct information, and what consequence could follow from an error, disclosure, inference, or inaccessible process. Do not collapse organizational harm and harm to individuals into one score.

### A — Account for the complete data flow

Draw the flow from collection to deletion. For every step, record the data fields, source, purpose, system, location, access roles, transformation, model input and output, logging, recipient, transfer, retention, and deletion route. Include prompts, retrieved documents, embeddings, temporary files, support access, feedback, telemetry, human notes, and derived classifications. Mark unknowns rather than assuming a vendor default.

The NIST Privacy Framework is a voluntary enterprise-risk tool intended to help organizations identify and manage privacy risk. Its Identify-P, Govern-P, Control-P, Communicate-P, and Protect-P functions can be used as a completeness check for the data-flow and governance record. Framework use does not resolve a specific legal question or prove that local controls operate. NIST Privacy Framework

For each field, ask whether the workflow can operate with less data, a shorter retention period, a smaller population, lower precision, local processing, stronger separation, or a non-AI method. Keep the rejected alternatives and reasons.

### P — Probe necessity, proportionality, and risk

Connect every data use to the stated purpose. Then test four layers:

Describe the pathway from event to effect: source condition, AI or human action, affected person, possible impact, existing control, residual uncertainty, and evidence needed. Rate likelihood and severity only after writing that causal chain. A colored matrix without the chain is difficult to review.

Test the proposal against credible alternatives: no change, process repair, a deterministic rule, a smaller dataset, an on-device or isolated design, a shorter-lived output, or a manual service. The assessment should explain why the selected approach is necessary and proportionate to the bounded purpose, not merely why it is convenient.

### S — Specify controls, consultation, and disposition

Every control needs an owner, implementation evidence, operating evidence, failure signal, review date, and response. Examples include field minimization, role-based access, encryption, tenant isolation, provider configuration, retention enforcement, deletion testing, output review, appeal or correction routes, logging, incident response, and change monitoring.

Consult people with relevant perspectives when feasible: affected users, worker representatives, accessibility specialists, frontline operators, security, privacy, data governance, and the owner of the underlying service. Record what changed because of consultation and what was not adopted.

Conclude with one disposition: proceed, proceed with conditions, redesign and reassess, hold for missing evidence, or stop. State the residual risks, accepting authority, monitoring plan, incident trigger, and reassessment trigger. Approval is not a blank cheque for future variants.

  • collection risk: excessive, inaccurate, unexpected, coerced, or inaccessible data;
  • processing risk: unintended inference, context loss, insecure transfer, provider reuse, memorization, or opaque transformation;
  • decision risk: people unable to understand, contest, correct, or escape a consequential output;
  • lifecycle risk: access drift, model or configuration change, stale data, retained logs, weak deletion, or unmonitored incidents.

Sources: [2]

Copyable assessment artifact

| Field | Required entry | | --- | --- | | Decision | Proposal, decision owner, deadline, and allowed outcome | | Mandate | Purpose, success measure, affected people, expected use, secondary uses excluded | | System boundary | Model/provider/version, integrations, human role, jurisdictions, and environments | | Data inventory | Field, source, person represented, sensitivity, necessity, accuracy issue | | Flow | Collection, transformation, storage, access, recipient, transfer, retention, deletion | | Alternatives | No change, non-AI option, narrower option, selected option, rationale | | Risk pathway | Trigger, action, affected person, possible effect, likelihood, severity | | Control evidence | Control, owner, implementation evidence, operating evidence, failure signal | | Consultation | People consulted, issue raised, change made, unresolved disagreement | | Disposition | Proceed/redesign/hold/stop, conditions, residual risk, accepting authority | | Review | Metric, incident trigger, change trigger, review date, record location |

Worked example: an internal meeting-note assistant

A project office proposes an AI assistant that turns internal meeting recordings into summaries and action lists. The initial description says it uses “ordinary business information,” but the MAPS review finds that meetings can include health-related absence explanations, customer identifiers, performance concerns, and comments about people who are not present.

The data-flow map shows audio uploaded to a third-party service, a transcript retained in the application, prompts and outputs stored in an audit log, summaries copied into the project system, and corrections made in email. The provider documentation does not yet resolve support access and deletion timing for every log. Those are recorded as evidence gaps, not inferred away.

The team compares five options: keep manual notes, prohibit recording and use a structured action form, process audio locally, restrict the assistant to meetings with a defined agenda and participant notice, or use the original broad design. The structured form meets most action-tracking needs. The narrower assistant remains useful for a limited set of project meetings.

Controls for the narrower design include an approved-meeting list, visible notice before recording, a host confirmation, exclusion of sensitive meeting types, participant access to the transcript, human review before sharing, automated deletion after the approved period, restricted support access, and a correction route. A monthly deletion test and sampled notice check provide operating evidence. A privacy specialist owns reassessment when the provider, retention, meeting population, or downstream integration changes.

The disposition is redesign and reassess rather than approve the original proposal. The record explains that the narrower design still needs provider evidence and a tested deletion route. It does not claim that meeting summaries are universally accurate or that the workflow satisfies every applicable requirement.

Acceptance and failure conditions

Accept the artifact only when the proposal and decision owner are explicit; the complete data flow is inspectable; data necessity and alternatives are documented; risks describe effects on people; controls have owners and evidence; consultation and disagreement are visible; residual risks have an authorized disposition; and reassessment triggers are operational.

Hold or reject the artifact when material flows are unknown, vendor claims substitute for evidence, the purpose keeps expanding, a generic risk matrix replaces causal analysis, affected people have no practical correction route, controls have no operating evidence, or the named approver lacks authority.

Two Prune boundary

Two Prune can create realistic work simulations that assess how participants frame a privacy problem, navigate evidence, use AI inspectably, verify claims, control risk, exercise judgment, and communicate a decision-ready artifact. It does not provide legal advice, conduct a customer's DPIA, operate privacy controls, certify compliance, or make employment decisions. Customers retain responsibility for their processing, tools, data, legal analysis, 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

Explore this topic