Guide GDPR AI Compliance

When AI processing needs a DPIA: how to screen AI tools and what the assessment must contain

Not every AI tool needs a DPIA, but many uses of AI do. UK GDPR Article 35 requires one before processing that is likely to result in a high risk to individuals. This guide shows how to screen AI tools against the legal triggers and the nine WP248 criteria, where Copilot and HR AI usually land, and what the assessment has to contain.

RD
Ryland Deakin
Lead Consultant, Cyvra
21 July 2026
10 min read
Updated 10 November 2026

What Article 35 actually requires

UK GDPR Article 35 requires a Data Protection Impact Assessment before processing that is likely to result in a high risk to the rights and freedoms of natural persons. The assessment comes first. If a tool that met the threshold went live without one, complete it now and record that it was done late; the gap is a compliance failure for as long as it lasted.

Three sources tell you where the threshold sits. Article 35(3) names three cases that always need a DPIA: systematic and extensive evaluation of people, including profiling, that leads to decisions with legal or similarly significant effects; large-scale processing of special category or criminal offence data; and systematic monitoring of publicly accessible areas on a large scale. The ICO publishes a list of further processing types that need a DPIA, including innovative technology combined with another risk factor, large-scale profiling, and tracking of individuals. And the Article 29 Working Party guidelines (WP248 rev.01, endorsed by the EDPB and still used by the ICO) set out nine criteria that help you judge whether high risk is likely.

The nine criteria as indicators

The WP248 criteria are indicators, not a scoring sheet. In most cases, processing that meets two of them needs a DPIA. One criterion can be enough when the risk is clearly high. If you meet two or more and still conclude that the processing is unlikely to be high risk, you need to document your reasons.

01
Evaluation or scoring, including profiling and predicting behaviour, performance or preferences
02
Automated decision-making with legal or similarly significant effects
03
Systematic monitoring of individuals
04
Special category or criminal offence data, or other highly personal data such as financial or location data
05
Processing on a large scale
06
Matching or combining datasets in ways people would not reasonably expect
07
Data about vulnerable people, including employees, children and patients, where there is a power imbalance
08
Innovative use of technology or new organisational solutions
09
Processing that in itself prevents people from exercising a right or using a service or contract

Criterion 8 needs care with AI. Generative AI and machine learning can count as innovative technology, especially when the use is new to your organisation or the risks are not yet well understood. That is an assessment of the specific use, not a label that attaches to every AI tool. A writing assistant that drafts emails from what the user types sits differently from a model that scores job applicants.

Criterion 4 also covers more than the Article 9 special categories. Health, biometric and religious data are special category data. Financial data is not, but WP248 treats it as highly personal because misuse can cause real harm, so it still counts as an indicator.

Four common AI scenarios

These scenarios show how the criteria tend to apply. The outcome always depends on how you configure and use the tool.

Scenario 1
HR and recruitment AI
AI that screens CVs, ranks candidates or analyses performance data usually meets criterion 1 (scoring) and criterion 7 (employees and applicants). Criterion 2 applies where the AI effectively makes or decides the outcome, for example by rejecting candidates without meaningful human review. A DPIA is very likely in this scenario.
Scenario 2
Customer profiling and scoring
Platforms that profile behaviour, assess creditworthiness or personalise prices meet criterion 1. Criterion 5 applies once the processing is large scale, judged by the number of people, volume of data, duration and geography. Financial data adds criterion 4, and combining sources in ways customers would not expect adds criterion 6.
Scenario 3
Employee activity and communication analytics
Tools that analyse work patterns, response times or document activity to produce insights about individual employees meet criterion 3 (systematic monitoring) and criterion 7 (employees). Criterion 8 may apply too. Using Copilot or similar tools to summarise what specific employees have been doing falls into this scenario.
Scenario 4
Automated access and eligibility decisions
AI that decides who gets a service, benefit or account without meaningful human review meets criterion 2. Criterion 9 applies when the outcome actually blocks someone from a right, service or contract. An opaque model makes the risks harder to manage, but opacity alone does not trigger criterion 9. Fraud detection and onboarding checks often sit here.

Does Microsoft Copilot need a DPIA?

Not automatically. It depends on what you use it for.

Microsoft 365 Copilot works on email, Teams chats, documents and calendar data that the signed-in user can already access. Using it to summarise a document, draft a reply or search your own files is closer to ordinary productivity software. The picture changes when managers or HR use it to analyse what individual employees have written or done, when it feeds decisions about people, or when it gives easy access to special category data that permissions never properly restricted. Those uses can meet criteria 1, 3, 4 and 7.

Many organisations run a DPIA for a Copilot rollout anyway, because it processes large volumes of employee data and the risks are new. That is a reasonable choice. If you conclude one is not required, document the screening and the restrictions on use that support that conclusion. Either way, the permissions review, sensitivity labels, Purview controls and acceptable use policy from our Copilot governance article are the mitigations you would record.

Before vs after deployment

A DPIA completed after go-live is better than none, but it does not cover the period the tool ran without one. If the ICO investigates an incident involving that tool, a missing DPIA where one was required is an Article 35 infringement in its own right, and it weakens your accountability position.

What a DPIA must contain

Article 35(7) sets four minimum components, whatever the format.

  1. A systematic description of the processing: what data the AI tool processes, where it comes from, who can access it, how long it is kept, and why
  2. An assessment of necessity and proportionality: why the processing is needed for the purpose, and whether a less intrusive approach would achieve it
  3. An assessment of the risks to individuals: harms to the people whose data is processed, such as discrimination, loss of control over their data, inaccurate outputs about them, or unexpected disclosure, each rated for likelihood and severity
  4. The measures to address those risks: technical and organisational safeguards, mapped to each risk

Article 35(2) requires you to seek the advice of your Data Protection Officer if you have one, and to record it. If the DPIA shows a high risk that you cannot reduce to an acceptable level, Article 36 requires prior consultation with the ICO before processing starts.

What to document

Processing description. Name the tool, the vendor and its role, the data flows, the lawful basis, the categories of data and people involved, and the retention periods. For Copilot, record the retention policies you have configured in Microsoft Purview for Copilot prompts and responses, not a default you assume applies.
Screening record. Check the processing against Article 35(3), the ICO list and the nine criteria. Record which apply and why. If you decide no DPIA is needed, keep the reasoning on file.
Necessity and proportionality. State what the tool achieves and why it needs this level of data access. Record the alternatives you considered and why you rejected them.
Risk register. Rate each risk to individuals for likelihood and severity. Typical AI risks include overshared data surfacing in answers, biased or inaccurate outputs about people, use of data for purposes nobody was told about, and individuals being unable to understand or challenge a decision.
Mitigations mapped to each risk. Name the control that addresses each risk. Technical controls (permissions, sensitivity labels, DLP, audit logging) and organisational controls (use restrictions, human review, training, DPO review) both count.
DPO advice. Record the DPO's advice and what you did with it. The DPO advises; the controller decides. If you do not follow the advice, record why.
Review triggers. Set the events that reopen the DPIA: new features, new data sources, new uses, or changes to the vendor's terms. An annual review is a sensible default for tools in active use.

DPIA, FRIA and the EU AI Act

If you place AI systems on the EU market or use their output in the EU, the EU AI Act adds its own assessments. Providers of high-risk systems carry out the conformity assessment. Deployers have different duties. Under Article 27, public bodies, private entities providing public services, and deployers of systems that assess creditworthiness or price life and health insurance must complete a fundamental rights impact assessment (FRIA). Where a DPIA already covers some of the same ground, the FRIA complements it rather than replacing it.

After the Digital Omnibus, the main obligations for Annex III high-risk systems apply from 2 December 2027, and those for AI in products covered by Annex I from 2 August 2028. Deployers can reuse much of the DPIA evidence: the processing description, the data categories, the vendor information and the risk register.

A staged screening process

  1. Identify the processing. List the AI tools in use, including those departments bought on their own, and what personal data each one touches.
  2. Confirm the jurisdiction. UK GDPR, EU GDPR, or both if you serve or monitor people in the EU.
  3. Look for high-risk indicators. Run each use through the nine criteria.
  4. Check the mandatory triggers. Article 35(3) and the ICO list. If either applies, a DPIA is required.
  5. Decide on the rest. Where a DPIA is not mandatory, decide whether to run one anyway and record the decision.
  6. Assess and mitigate. Complete the DPIA, starting with HR, customer profiling and special category data.
  7. Consult where required. Seek DPO advice, and consult the ICO if high residual risk remains.
  8. Approve and monitor. Sign off, then review when the tool, data or use changes.

The length of a DPIA should match the risk. As a rough guide from our work, a focused assessment of one tool can take a few weeks when a data map already exists, and longer without one. A short document that answers the four Article 35(7) questions honestly is worth more than a long template filled in for the sake of it.

A DPIA sits within your wider GDPR compliance framework. Cyvra runs DPIAs for AI tools from screening through to the risk register, mitigation mapping and DPO liaison.

Frequently asked questions

Is a DPIA mandatory for every AI tool?

No. A DPIA is required when processing is likely to result in a high risk to individuals, which covers the cases in Article 35(3), the processing on the ICO's list, and other processing where the nine WP248 criteria point to high risk. In most cases, meeting two criteria means a DPIA is needed. HR AI, large-scale customer profiling and employee monitoring often meet that bar. A general writing assistant used on the user's own documents may not.

What happens if we deploy AI without a DPIA when one was required?

Failing to carry out a required DPIA infringes Article 35. Under UK GDPR, that falls in the lower fine tier: up to £8.7 million or 2% of global annual turnover, whichever is higher (under EU GDPR, €10 million or 2%). In practice, the ICO is likely to look at whether a DPIA exists when it investigates an incident involving the tool, and a documented assessment supports your accountability case.

Does Microsoft Copilot require a DPIA?

It depends on how you use it. Summarising documents or drafting emails from data the user can already see is different from managers using Copilot to analyse individual employees or support decisions about them. The second kind of use can meet several criteria and is likely to need a DPIA. Many organisations run one for the rollout anyway, given the volume of employee data involved. If you decide not to, document the screening and the use restrictions behind that decision.

How long does a DPIA take to complete?

It depends on the tool, the risk and the state of your records. As an illustration from our projects, a focused DPIA for one tool can take a few weeks when a processing inventory and data map already exist, and longer when they have to be built first. Length should follow risk: a concise document that answers the four Article 35(7) questions properly is better than a long template.

What is the difference between a DPIA and the EU AI Act assessments?

A DPIA is a data protection obligation of the controller. Under the EU AI Act, providers of high-risk systems carry out conformity assessments, and certain deployers (public bodies, private entities providing public services, and those using AI for credit scoring or life and health insurance pricing) must complete a fundamental rights impact assessment under Article 27. The FRIA complements an existing DPIA, and much of the evidence can be shared.

How does Cyvra help organisations complete a DPIA for AI tools?

We screen each AI use against Article 35(3), the ICO list and the nine criteria, then run the DPIA where one is needed: processing description, necessity and proportionality, a risk register focused on harm to individuals, and mitigations mapped to each risk. We work with your DPO where you have one and help prepare an ICO prior consultation if high residual risk remains.

Sources

Ryland Deakin
About the author
Lead Consultant, Cyvra · CISM · CompTIA Security+ · MCP

Ryland has delivered cybersecurity, compliance, and IT management programmes for regulated organisations across the UK and the Netherlands for over 20 years. View full profile

Work with Cyvra

Need a DPIA for your AI tools?

We screen your AI tools, run the DPIAs that are needed, and map each risk to a control, with your DPO's advice on record.

Disclaimer: This article is for informational purposes only and does not constitute legal, regulatory, or professional advice. Cyvra makes no representations as to the accuracy or completeness of this content. Readers should seek independent professional advice tailored to their specific circumstances.