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.
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.
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.
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.
- 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
- An assessment of necessity and proportionality: why the processing is needed for the purpose, and whether a less intrusive approach would achieve it
- 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
- 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
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
- Identify the processing. List the AI tools in use, including those departments bought on their own, and what personal data each one touches.
- Confirm the jurisdiction. UK GDPR, EU GDPR, or both if you serve or monitor people in the EU.
- Look for high-risk indicators. Run each use through the nine criteria.
- Check the mandatory triggers. Article 35(3) and the ICO list. If either applies, a DPIA is required.
- Decide on the rest. Where a DPIA is not mandatory, decide whether to run one anyway and record the decision.
- Assess and mitigate. Complete the DPIA, starting with HR, customer profiling and special category data.
- Consult where required. Seek DPO advice, and consult the ICO if high residual risk remains.
- 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.