# DPIA for employee use of ChatGPT under the GDPR

*A Sluis guide for privacy officers and data protection officers. Last checked 19 November 2026.*

When employees use ChatGPT with personal data, the GDPR in practice often calls for a DPIA before the use starts. This guide shows how to test that against the nine criteria of the European regulators, what the DPIA must contain, which risks and measures belong in it and which facts to collect from each AI supplier. It is information, not legal advice: the assessment remains your organisation's, together with your DPO.

## In short

- **Is a DPIA needed?** In practice usually yes: as soon as employees enter personal data at any scale, or as soon as you monitor employees through the tool. The rule of thumb of the Dutch Data Protection Authority (the AP) is a DPIA when processing meets **two or more of nine criteria**. Employees count as **vulnerable data subjects** (unequal power relationship) and generative AI counts as new technology, so two criteria are quickly met.
- **A pilot counts.** A pilot of generative AI can also process personal data, and then a DPIA is needed.
- **Who does it?** The controller, so your organisation. The DPO must be asked for advice and you record that advice. The processor must support you.
- **What goes in it?** At least four elements (Article 35(7) GDPR): a systematic description with purposes, an assessment of necessity and proportionality, a risk assessment and the measures envisaged.
- **High residual risk?** Then you consult the supervisory authority before you start (Article 36 GDPR).
- **How often, and how many?** A DPIA is a continuous process: the AP advises repeating it periodically, for example every three years, and again when things change. Write one per processing purpose, not per tool, and do it yourself: pointing to another organisation's DPIA is not enough.

## When is a DPIA needed?

Article 35(1) GDPR requires a DPIA before processing where a type of processing, in particular using new technologies, is likely to result in a high risk to people given its nature, scope, context and purposes. In its Hulpmiddel inzet generatieve AI & de AVG (a practical tool for deploying generative AI, 13 July 2026) the AP names generative AI as an example of such new technology. You decide for yourself whether a DPIA is mandatory and must be able to justify why you do or do not carry one out.

There are two aids. The first is the AP's list of mandatory DPIAs (Staatscourant 2019, no. 64418). Item 11, *Controle werknemers* (monitoring of employees), covers large-scale processing of personal data and/or systematic monitoring of employee activities, with e-mail and internet use as an example. The list is not exhaustive, and the AP says ad hoc or incidental processing is not systematic. Simply chatting with an AI assistant is not on the list. In our view, systematically logging and reading what employees enter is the kind of processing item 11 has in mind.

The second is the nine criteria of the European regulators (EDPB, WP248 rev.01). If your processing meets two or more, the AP's rule of thumb is that you carry out a DPIA. If it meets none or one, you still have to be able to justify why you do not, as part of your accountability. The EDPB does not rule out that a single criterion is sometimes enough.

**Table 1. The nine criteria applied to employees using a generative AI chatbot**

| Criterion | Does it apply? | Why |
|---|---|---|
| 1. Evaluation or scoring of people | Only for certain uses | When prompts are used to assess or rank colleagues or applicants, for example summarising performance reviews or scoring CVs. |
| 2. Automated decisions | Only if the output drives decisions | The test is legal or similarly significant effects without meaningful human involvement. |
| 3. Systematic monitoring | Yes, if you log and review prompts per employee | Chatting itself is not monitoring; systematic logging and checking is. The EDPB gives monitoring of a work station and internet use as an example where a DPIA is likely needed. |
| 4. Sensitive data | Yes, unless excluded | Without a filter and a policy, special category data ends up in prompts. In 2024 the AP received reports about patient data and a customer file with addresses entered into a chatbot. |
| 5. Large scale | Probably | The EDPB looks at the number of data subjects, volume of data, duration and geographic extent. A whole organisation plus its customers, on a lasting basis, is quickly large scale. A small pilot not necessarily. |
| 6. Matching or combining datasets | Only with connections | As soon as the AI can reach mailboxes, drives or CRM. The AP points out that some systems can reach every file the user can. |
| 7. Vulnerable data subjects | Yes | The AP names employees explicitly, because of the unequal power relationship. Patients and customers in prompts count too. |
| 8. New technology | Yes | The AP names generative AI as an example. |
| 9. Blocking a right, service or contract | Only for decisions on access | When the output decides who gets a service or contract. |

On this reading, criteria 7 and 8 almost always apply, and that already meets the rule of two. The AP also advises first checking whether you can use generative AI without personal data. Record this threshold analysis in writing, even if you decide not to do a DPIA.

Two caveats. In June 2026 the AP put a draft list of processing operations without a DPIA duty out for consultation, aimed at SMEs among others. The consultation is closed and the list was not final on that AP page, so you cannot rely on it. And a DPIA is not a legal basis: your processing has to be lawful even if the DPIA comes out well. Employee consent is difficult, because according to the AP the power imbalance means employees cannot freely give or refuse consent.

## Roles and process

**Who signs and who advises.** The controller makes sure the DPIA gets done. You may have someone inside or outside the organisation carry it out, but you remain ultimately responsible. Have the owner of the processing, a board or management team member, adopt the DPIA. The DPO advises and monitors the performance. Seeking the DPO's advice is mandatory where one is designated (Article 35(2)), and you record in the report what the DPO advised and what you did with it. The AP recommends also consulting the DPO on whether to do a DPIA, which method to use and which safeguards are needed.

**The processor.** If the supplier is a processor, the data processing agreement must oblige it to assist you with your obligations under Articles 32 to 36 GDPR (Article 28(3)(f)), so also with the DPIA and a prior consultation. The AP puts it this way: the processor must support you and give you the information you need.

**Employees and the works council.** Where appropriate, seek the views of data subjects or their representatives (Article 35(9)), for example by consulting employee organisations, and record your reasons if your decision departs from their views. If you have a works council, look at Article 27(1) of the Dutch Works Councils Act: consent is needed for a scheme on processing and protecting the personal data of people working in the business (point k) and for facilities suited to observing or checking attendance, behaviour or performance (point l). An AI policy with logging may fall under these, so involve the council early.

**When to go to the AP.** If a high residual risk remains after your measures, you consult the supervisory authority before you start (Article 36(1)). In your DPIA, describe which high risks you cannot fully prevent, in which situations, how likely they are and what harm could result. With the request you supply, among other things, the responsibilities of the parties, the purposes and means, the measures, the DPO's contact details and the DPIA itself (Article 36(3)). If the authority finds that the processing would infringe the GDPR, it gives written advice within eight weeks, extendable by six (Article 36(2)).

**Publishing and review.** Publishing is not required, but the AP recommends it. A summary is enough if you do not want to disclose security or commercially sensitive information. After changes in the processing, its risks or its context a new DPIA is needed, and the AP advises a periodic DPIA, for example every three years, even if nothing changed. Appoint people responsible for that review and use your record of processing activities to see which processing the AI use touches.

## What the DPIA must contain

The AP leaves the method open as long as you meet the four elements of Article 35(7). The AP tool is not itself a DPIA, but you can attach it as an annex. The skeleton below maps the elements to the chapters of that tool: 1 getting the basics in order, 2 choosing the system, 3 deployment.

**Table 2. Skeleton of a DPIA for employees using generative AI**

| Section | What to write | Evidence to attach |
|---|---|---|
| 1. Processing and purposes (Art. 35(7)(a); AP tool 1.1, 1.2, 2.1) | A concrete purpose, for example "summarise customer e-mails for the support team" and not "support employees". Categories of personal data and data subjects, systems and connections, data flow, legal basis. | Extract of the record of processing (Art. 30), data flow diagram, legal basis assessment |
| 2. Necessity and proportionality (Art. 35(7)(b); AP tool 2.1 to 2.5) | Why AI is needed, whether it can be done without personal data, which data is strictly necessary, how long it is kept and which alternatives you weighed: ChatGPT, an EU alternative or no AI. | Comparison of alternatives, policy on permitted data categories, retention settings |
| 3. Risk assessment (Art. 35(7)(c); AP tool 2.2, 3.2, 3.3) | Per risk, likelihood and severity for the data subjects, not for the organisation. Then the high residual risks. | Risk table (Table 3), DPO advice, views of employees or the works council |
| 4. Measures and demonstrability (Art. 35(7)(d); AP tool 1.3, 1.4, 2.3, 2.6, 3.4) | Technical, organisational and contractual measures, who carries them out and when, and how you show they work. | Processing agreement, sub-processor list, authorisation policy, training evidence, audit log export, review schedule |
| 5. Advice and decision (Art. 35(2) and (9); AP tool 3.1) | The DPO's advice and your response, the views of data subjects, the owner's decision, and whether prior consultation is needed. | DPO advice, minutes of the works council discussion, adopted decision |

Keep the DPIA per processing purpose, not per tool. Article 35(1) allows a single assessment to cover a set of similar processing operations that present similar high risks. The choice between ChatGPT, Copilot or an alternative is then a judgement inside section 2. The AP tool also asks you to have the basics in order before you start: an up-to-date record of processing, an overview of personal data, authorisations that match employees' roles and insight into processing in the cloud.

## Typical risks and measures

For the residual risk the AP asks for four things: which high risk you cannot fully prevent, in which situations, how likely it is and what harm follows. The wording in the last column is an example that you fill in with your own facts.

**Table 3. Risks, measures and residual risk for employee use of AI**

| Risk | Measure | Example residual risk wording |
|---|---|---|
| Prompts contain personal data of colleagues, customers or patients | A policy on which categories may and may not be entered, an input filter that warns or replaces (the AP names real-time input filters), training. See [personal data in prompts](/guides/personal-data-in-prompts). | "A filter does not recognise every name. Likelihood: medium. Impact limited for ordinary categories; special categories are excluded." |
| Supplier keeps the input and trains on it | Terms on further processing and training in the processing agreement, retention periods recorded, audit right (Art. 28(3)(h)). The AP notes you can agree that the provider does not store what is entered. | "Compliance can only be checked through contract and audit. Likelihood: low to medium." |
| Transfer outside the EU and access by non-EU authorities | Record the transfer basis, limit routing to the EU, send less. See [CLOUD Act exposure](/guides/cloud-act-risk-assessment). | "A non-EU authority's power remains possible for a supplier under non-EU control. Non-sensitive data only." |
| Wrong output about a person | AI literacy training (see [the AI Act for deployers](/guides/ai-act-for-deployers)), time and knowledge to check output, no use as a source of facts about people. The Copilot advice lists this as one of four high residual risks from the reference DPIA. | "A convincing but wrong answer remains possible. Likelihood: medium. Use limited to drafts a person checks." |
| Automated decision-making (Art. 22) | Human review stays decisive, employees may depart from the output and have time to test it. | "Under time pressure, human involvement can become superficial. Likelihood: low through working rules and spot checks." |
| Access and rectification when the supplier stores data | Assistance with requests in the processing agreement (Art. 28(3)(e)), an agreed export procedure, the shortest possible retention. | "Access is only possible as far as the supplier delivers data. If that stays unclear, the residual risk is high and prior consultation follows." |
| Logging prompts is itself monitoring (DPIA list item 11) | Log metadata only where possible, keep content briefly, limit read access to a named role, record the purpose, inform employees, works council consent (Art. 27(1)(k) and (l) WOR). | "Feeling watched may suppress use. Likelihood: medium. Limited by metadata, short retention and works council consent." |
| The AI gets access to mailboxes, drives or CRM | Limit access per purpose, align authorisations with roles, connections off by default. | "Documents that are not needed are reachable through the AI. Likelihood: low with role-based authorisations." |
| Employees use private accounts next to the approved tool | An equally easy approved alternative, clear rules, a reporting procedure. The AP points to the breach notification duty. See [rolling out without shadow AI](/guides/shadow-ai-rollout-guide). | "Use outside the approved tool remains possible. Likelihood: medium. Impact: a reportable data breach." |

## ChatGPT, Copilot and EU alternatives: what to put in the DPIA about the tool

The AP sees many people at work using digital assistants such as ChatGPT and Copilot. In your DPIA you do not choose a brand but a means for a purpose. The AP tool says that using a supplier's other services is not enough reason to choose that supplier's AI system as well. An existing Microsoft or Google contract therefore decides nothing. A system that shares input with the supplier and uses it for further training also has a greater impact than a system you developed and run yourself.

In 2025 the AP gave advice on a municipality's intended use of Microsoft 365 Copilot, following a prior consultation. The advice was published on 13 July 2026. The AP concluded that the intended processing, as described, would infringe the GDPR unless measures were taken. The reference DPIA the municipality relied on named four high residual risks: data subjects could not exercise their rights, incorrect generated personal data could have adverse effects, there was too little transparency about the supplier's service and telemetry data, and pseudonymised data could be re-identified because retention periods were unknown. According to the AP, a lack of transparency at the supplier does not relieve the organisation of its own obligations, and a supplier that does not cooperate cannot be used. The advice rests on that municipality's processing, and other organisations must judge for themselves whether theirs matches. Read it as an example of the questions the AP asks, not as a verdict on a product.

Collect the following facts per supplier, preferably in writing. None of these points says a supplier complies: the answers decide that.

**Table 4. Facts to collect per AI supplier**

| Item | What to ask | Why it matters |
|---|---|---|
| Role | Processor, or (joint) controller, per type of processing | Decides whether you need a processing agreement (AP tool 2.6) |
| Processing agreement | Does it cover Article 28(3): written instructions incl. transfers, confidentiality, security, sub-processors, assistance with rights and with Articles 32 to 36, deletion, audit | Contractual basis of section 4 |
| Retention | Per category (input, output, metadata, diagnostic and telemetry data): where, how long, adjustable or not | Storage limitation (AP tool 2.5) |
| Training | May the supplier use input for its own models, and can you switch that off | Further processing (AP tool 2.6) |
| Sub-processors | List, locations, notice of changes, right to object (Art. 28(2) and (4)) | Chain of responsibility |
| Transfer basis | Which basis per country, and what access authorities outside the EU may have | See CLOUD Act exposure |
| Storage versus processing location | Where data sits at rest, where the model runs and who controls the supplier, as three separate questions | An EU region says nothing about control |
| Admin controls | Who has access, which models and connections, which retention period, sign-in through your own identity provider | Authorisation and data minimisation (AP tool 1.3, 2.4) |
| Audit log | Content or metadata, exportable, who reads it, how long | Demonstrability and monitoring risk |
| Filters and transparency | Insight into and influence over filters, and which service data the supplier collects | The AP asks about this explicitly in the Copilot advice |

## How a gateway or workspace layer such as Sluis maps to the measures

A gateway or workspace layer sits between employees and models and can enforce part of the measures in Table 3 technically instead of only describing them. Sluis, a managed service by 7Lab B.V. in Amsterdam with an ISO 27001 certificate, works like this: owners choose which models members may use, the full catalogue or a restricted set with a default. Residency defaults to EU only and is enforced at dispatch, so a request that would reach another jurisdiction is refused with a 403 before anything leaves. For Agent Harness model traffic there is a documented exception: the provider subscription chooses the region and Sluis records it as unverified. Detected personal data is replaced before sending by reversible, typed «TOKEN» placeholders and restored in the response. Each request gets one row in a hash-chained audit trail you can verify offline. MCP tools and sub-agents can be switched per organisation, and under an EU-only policy web search should be off, because web queries leave the boundary. Retaining request content for the audit log is a setting: on by default, 30 days by default, configurable from 1 to 36,500 days, or off for a metadata-only log. That choice belongs in your DPIA under monitoring.

The limits: detection is not perfect, Sluis does not itself block or detect shadow AI on other sites, and using Sluis does not by itself establish GDPR compliance for your processing. Assess Sluis with the same fact list as any other supplier (Table 4). It is a managed product from a young company, so weigh continuity too. More is at [the gateway](/gateway), [Sluis Workspace](/sluis-workspace) and [residency](/docs/residency).

## Checklist

- [ ] Threshold analysis recorded in writing, with the nine criteria applied (Table 1)
- [ ] Checked whether generative AI can be used without personal data
- [ ] Purposes described concretely and linked to the record of processing
- [ ] Legal basis set per purpose
- [ ] Fact list completed per supplier (Table 4), with written answers
- [ ] Processing agreement tested against Article 28(3), including assistance with the DPIA and rights
- [ ] Risks, measures and residual risks recorded (Table 3)
- [ ] DPO asked for advice; advice and response included
- [ ] Works council involved in logging and monitoring; consent recorded where required
- [ ] If a high residual risk remains: prior consultation requested before the start
- [ ] Employees trained; rules on what may and may not be entered
- [ ] Review scheduled, with an owner and triggers, at the latest after three years

## Questions

### Do I need a DPIA if employees only use a paid ChatGPT subscription?

The sources tie the DPIA duty not to the type of subscription but to the processing: nature, scope, context, purposes and the likelihood of a high risk (Article 35(1) GDPR). None of the AP sources we read says a paid subscription removes the duty. What counts: is personal data entered, how much and whose, and what does the supplier do with it (retain, train) in which role. What a specific subscription promises on that, you get from the supplier and record. We make no claims about it. In 2024 the AP also wrote that employees using a chatbot on their own initiative, against agreements, is a data breach when personal data goes in, and that use an organisation allows itself is often not lawful. So start with Table 1.

### May I reuse another organisation's DPIA?

Not as a replacement. According to the AP tool, you must carry out a mandatory DPIA yourself: referring to a generic DPIA (reference DPIA) of another organisation is not sufficient. In the Copilot advice the municipality had its own DPIA next to the reference DPIA, and the AP advised additional measures on top of those in the reference DPIA. Such a document can therefore be input for your risk list. Within your own organisation, one DPIA can cover a set of similar processing operations with similar risks (Article 35(1)).

### Is a DPIA needed for a pilot?

Yes, as soon as the pilot processes personal data. The AP tool says a pilot of generative AI can also process personal data and a DPIA is then needed. In the Copilot advice the AP writes that a pilot must also comply with the GDPR. A small pilot can help verify that measures work, for example on use cases with the smallest expected risk, but scaling up is not an aim in itself according to the AP. Start the DPIA in the design phase, even if not all details are known.

### How often must I repeat the DPIA?

There is no fixed legal interval. According to the AP a DPIA is a continuous process: you keep monitoring whether your processing, its risks or its context change, and a change calls for a new DPIA. The AP also advises a periodic DPIA, for example every three years, even if the processing has not changed. Article 35(11) requires a review at least when the risk changes, and the AP tool wants periodic reassessment built into your business processes. Because models, features and supplier terms change quickly, we also tie the review to triggers such as a new feature, purpose or data category. That is our practical addition, not a requirement from the sources.

## Sources

- [Autoriteit Persoonsgegevens, Data protection impact assessment (DPIA)](https://www.autoriteitpersoonsgegevens.nl/themas/basis-avg/praktisch-avg/data-protection-impact-assessment-dpia)
- [Autoriteit Persoonsgegevens, Lijst verplichte DPIA (Staatscourant 2019, no. 64418)](https://www.autoriteitpersoonsgegevens.nl/documenten/lijst-verplichte-dpia)
- [Autoriteit Persoonsgegevens, AP vraagt reacties op lijst DPIA-uitzonderingen, 24 June 2026](https://www.autoriteitpersoonsgegevens.nl/actueel/ap-vraagt-reacties-op-lijst-dpia-uitzonderingen)
- [Autoriteit Persoonsgegevens, Hulpmiddel generatieve AI en de AVG + VR Microsoft Copilot, 13 July 2026](https://www.autoriteitpersoonsgegevens.nl/documenten/hulpmiddel-generatieve-ai-en-de-avg-vr-microsoft-copilot)
- [Autoriteit Persoonsgegevens, Hulpmiddel inzet generatieve AI & de AVG (pdf, Dutch)](https://www.autoriteitpersoonsgegevens.nl/system/files?file=2026-07%2FHulpmiddel%20generatieve%20AI%20en%20de%20AVG_Autoriteit%20Persoonsgegevens.pdf)
- [Autoriteit Persoonsgegevens, advice on prior consultation about Microsoft 365 Copilot at a municipality (pdf, Dutch)](https://www.autoriteitpersoonsgegevens.nl/system/files?file=2026-07%2FVoorafgaande%20raadpleging%20advies%20Copilot%20gemeente%20Haarlemmermeer.pdf)
- [Autoriteit Persoonsgegevens, AP helpt ontwikkelaars en organisaties met nieuwe AVG-richtlijnen voor generatieve AI, 13 July 2026](https://www.autoriteitpersoonsgegevens.nl/actueel/ap-helpt-ontwikkelaars-en-organisaties-met-nieuwe-avg-richtlijnen-voor-generatieve-ai)
- [Autoriteit Persoonsgegevens, Let op: gebruik AI-chatbot kan leiden tot datalekken, 6 August 2024](https://www.autoriteitpersoonsgegevens.nl/actueel/let-op-gebruik-ai-chatbot-kan-leiden-tot-datalekken)
- [Regulation (EU) 2016/679 (GDPR), Articles 22, 28, 35 and 36, EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32016R0679)
- [EDPB (Article 29 Working Party), Guidelines on Data Protection Impact Assessment, wp248rev.01](https://ec.europa.eu/newsroom/article29/items/611236/en)
- [Wet op de ondernemingsraden (Dutch Works Councils Act), Article 27, wetten.overheid.nl](https://wetten.overheid.nl/BWBR0002747)
