Skip to content

DPIA template
for an AI agent.

This template helps an Irish business assess an AI agent before it processes personal data. It follows the four elements of GDPR Article 35(7): a description of the processing, necessity and proportionality, risks to people and the measures that address them. Example rows cover an AI receptionist and a WhatsApp agent.

When does an AI agent need a DPIA?

GDPR requires a data protection impact assessment (DPIA) where processing is likely to result in a high risk to people’s rights and freedoms. Article 35(3) makes it mandatory for three cases: automated decisions based on profiling with significant effects, large-scale processing of special category data, and large-scale monitoring of public areas.

The Data Protection Commission (DPC) adds its own list of processing types that need a DPIA after a documented screening. The list includes systematic monitoring of behaviour, large-scale profiling, combining datasets for profiling, and the use of special category data to decide access to a service. The DPC also lists the factors that raise risk: new technologies, large scale, sensitive data, vulnerable people and automated decisions. The DPC also refers to the European (WP29) guidelines: in most cases, processing that meets two of their criteria needs a DPIA.

Most AI agents meet at least one criterion, because a language model is a new technology. A dental practice’s AI receptionist will hear about symptoms, which is health data. A retailer’s WhatsApp agent combines message history with order records. Both are good candidates for a DPIA. Even where a DPIA is not mandatory, the DPC says that it is good practice for any major new project that uses personal data. The screening in Part B of the template records your decision either way.

Get the editable file.

DOCX · editable

DPIA for an AI agent (Word)

We use your email to send updates to this template. We do not share it.

The DPIA template in full

Use this template to assess an AI agent before it processes personal data. It follows the elements in Article 35(7) of the GDPR and the guidance of the Data Protection Commission (DPC). Replace every field in [square brackets]. Delete the example rows that do not apply to you. Keep the completed DPIA with your records of processing.

Part A. About this assessment

Assessment details
FieldEntry
Organisation (controller)[Legal name, address, company number]
AI agent or project name[For example: AI receptionist for phone and WhatsApp bookings]
Assessment owner[Name, role]
Data protection officer (DPO)[Name and contact, or "No DPO appointed"]
Developer or vendor[Company name, contact]
Version and date[1.0, DD Month YYYY]
Planned go-live date[DD Month YYYY]
Next review date[DD Month YYYY, and after any major change to the agent]

Part B. Screening: do you need a full DPIA?

Answer each question. If you answer "Yes" to one of the first three, a DPIA is mandatory. If you answer "Yes" to two or more of the others, a DPIA is very likely to be necessary. If you decide that you do not need a DPIA, record the reasons here.

Screening questions
QuestionYes or NoNotes
Does the agent make automated decisions, based on profiling, with legal or similarly significant effects on people?[Yes / No][For example: the agent refuses a service or a refund without a person]
Does the agent process special category data (for example health data) or criminal offence data on a large scale?[Yes / No][For example: patients describe symptoms when they book]
Does the agent systematically monitor a publicly accessible area on a large scale?[Yes / No][Usually No for a booking or support agent]
Does the agent use new technology, such as a large language model?[Yes / No][Usually Yes]
Does the agent process data about vulnerable people (children, patients, older people, employees)?[Yes / No][ ]
Does the agent combine data from different sources or systems?[Yes / No][For example: CRM, calendar and message history]
Does the agent evaluate or score people (for example lead scoring, triage)?[Yes / No][ ]
Does the agent process data on a large scale (many people, many records or a long period)?[Yes / No][Give the expected monthly volume]

Screening result: [A DPIA is necessary / A DPIA is not necessary, for these reasons: ...]. Decided by [name] on [date].

Part C. Description of the processing (Article 35(7)(a))

What the agent does with personal data
ItemEntry
Purpose[Why the agent processes personal data. One purpose per line.]
Lawful basis (Article 6)[Contract / Legitimate interests (attach the balancing test) / Consent / Legal obligation]
Special category condition (Article 9)[For example: Article 9(2)(h) health care, with the related condition in Irish law. Write "Not applicable" if none.]
People whose data is processed[Patients, customers, enquirers, staff]
Personal data the agent reads[Name, phone number, appointment time, order number, message text]
Personal data the agent writes or stores[Booking records, call transcripts, conversation logs, summaries]
Channels[Phone, WhatsApp, web chat, email]
Steps the agent performs[1. Receives a message. 2. Identifies the request. 3. Checks the calendar. 4. Books or hands over to a person.]
Actions the agent must never do[For example: give clinical advice, take card payments, change prices]
Systems the agent connects to[Practice software, calendar, CRM, e-commerce platform]
Expected volume[Number of conversations per month, number of people]
Retention[How long you keep transcripts, logs and summaries, and how you delete them]
Processors and sub-processors
ProcessorRoleData it receivesLocation of processingContract (DPA) in place
[AI model provider][Generates replies][Message text, booking details][EU / US with SCCs / other][Yes, reference]
[Messaging or telephony provider][Carries calls or messages][Phone number, message content][ ][ ]
[Hosting provider][Runs the agent and stores logs][Logs, transcripts][ ][ ]
[Developer or agency][Builds and supports the agent][Access to logs for support][ ][ ]

Part D. Necessity and proportionality (Article 35(7)(b))

Necessity and proportionality questions
QuestionYour answer
Why is an AI agent necessary for this purpose? What did you do before?[ ]
Can the agent do the job with less personal data? Which fields did you remove?[ ]
How do you keep the data accurate (for example, confirm the date and name back to the person)?[ ]
How do you tell people that they deal with an AI system, and how they can reach a person?[ ]
How do you update your privacy notice to describe the agent and its processors?[ ]
How do people use their rights (access, correction, erasure, objection), including on transcripts and logs?[ ]
Does personal data leave the European Economic Area? Which transfer mechanism applies?[ ]
Do the providers use your data to train their models? Where is this set?[ ]
Why is the retention period you chose the shortest one that works?[ ]

Part E. Risks to individuals (Article 35(7)(c))

Write each risk from the point of view of the person, not the business. Score likelihood and severity as Low, Medium or High. The rows below are examples for a dental practice AI receptionist (R1 to R4) and a retailer WhatsApp agent (W1 to W4). Keep the rows that apply and add your own.

Risk register
Risk to the personLikelihoodSeverityOverall risk
R1. A patient describes symptoms and the agent stores them in a general booking note that many staff can read.HighMediumHigh
R2. A caller does not know that they speak to an AI system and shares more than they intend.MediumMediumMedium
R3. The agent matches a caller to the wrong patient record by phone number and reveals appointment details.LowHighMedium
R4. Call transcripts are kept longer than necessary at the provider.MediumMediumMedium
W1. A customer sends card details or a photo of an ID document in the chat.MediumHighHigh
W2. A message contains instructions that make the agent reveal another customer’s order.LowHighMedium
W3. Message content is processed outside the EEA without a valid transfer mechanism.MediumMediumMedium
W4. The agent sends marketing messages to customers who did not opt in.MediumLowMedium
[Your risk][ ][ ][ ]

Part F. Measures to address the risks (Article 35(7)(d))

Measures and residual risk
RiskMeasureOwnerResidual risk
R1The agent records only the reason category (for example "check-up" or "pain"). It does not store free-text symptoms. Clinical notes stay in the practice software.[Practice manager]Low
R2The agent says at the start of each call that it is an AI assistant. The caller can ask for a person at any time.[Practice manager]Low
R3The agent confirms the name and date of birth before it gives any appointment details.[Developer]Low
R4Transcripts are deleted after [30] days. Provider data retention is set to the shortest option. Model training on your data is off.[Developer]Low
W1The agent does not ask for card or ID details. It masks card-like numbers in logs and sends a payment link instead.[Developer]Low
W2The agent can look up only the order linked to the sender’s phone number. Tool permissions stop all other lookups.[Developer]Low
W3Use an EU region where available. Otherwise use the provider’s DPA with standard contractual clauses and a transfer risk assessment.[Owner]Low
W4Marketing templates go only to contacts with a recorded opt-in. Opt-outs are applied at once.[Marketing lead]Low
[Your risk][ ][ ][ ]

Part G. Consultation

  • DPO advice (Article 35(2)): [Summary of the advice, date].
  • Views of the people affected or their representatives, where appropriate (Article 35(9)): [How you asked, what they said, or why you did not ask].
  • Staff consulted: [Front desk, IT, management].
  • Developer or vendor input: [Security documents received, test results].

Part H. Sign-off

Sign-off
ItemName and decisionDate
Measures approved by[Name, role][ ]
Residual risks accepted by[Name, role][ ]
DPO advice[Summary][ ]
DPO advice followed?[Yes / No. If No, give the reasons.][ ]
Residual risk still high? (If Yes, consult the DPC before processing, Article 36.)[Yes / No][ ]
Next review[Date, or trigger: new channel, new data, new provider, new model][ ]

How to fill it in: a dental practice AI receptionist

Take a three-chair dental practice in Galway. It wants an AI receptionist to answer calls and WhatsApp messages, book check-ups and move appointments in its practice software. This is how the practice manager completes each part.

  1. Part B, screening. Patients will mention pain and treatments, so the agent processes health data. The agent uses a language model. The manager answers "Yes" to both and decides that a DPIA is necessary.
  2. Part C, description. The purpose is "book, move and cancel appointments, and answer questions about opening hours and published prices". The lawful basis is contract for bookings. For health data the manager writes the Article 9 condition that the practice already uses for its records, and asks the DPO to confirm it.
  3. Part C, processors. The table lists the telephony provider, the AI model provider, the hosting provider and the developer. For each one the manager records where the data is processed and the DPA reference.
  4. Part D, necessity. The manager removes fields that the agent does not need. The agent sees name, date of birth, phone number and free slots. It does not see treatment history.
  5. Parts E and F, risks and measures. The example rows R1 to R4 apply almost as written. The most important measure is R1: the agent stores a reason category, not free-text symptoms.
  6. Part H, sign-off. The principal dentist accepts the residual risks. The DPO confirms that no residual risk is high, so prior consultation with the DPC is not needed.

For a retailer’s WhatsApp agent the same steps apply. The rows W1 to W4 cover the usual risks: card details sent in chat, prompt injection, transfers outside the EEA and marketing without opt-in. Our security guide explains the controls behind W2.

Common mistakes in AI agent DPIAs

  • Doing it after launch. The DPC says that a DPIA must be done before the processing starts. Start it when you write the agent spec, because the measures often change the design.
  • Describing the technology, not the person. "The model may hallucinate" is not a risk to a person. "A patient gets a wrong appointment time and misses treatment" is.
  • Forgetting logs. Transcripts, debug logs and provider logs hold the same personal data as the conversation. Give each a retention period.
  • Copying the vendor’s security page. You are the controller. Ask the vendor the questions in our vendor questionnaire and record the answers.
  • Never reviewing it. The DPC refers to the European recommendation to reassess a DPIA at least every three years, and sooner when the processing changes. A new channel, model or data source is a change.
  • Ignoring the DPO’s advice without a record. If you do not follow the advice, Part H must say why.

What happens after sign-off?

If a high residual risk remains after your measures, Article 36 of the GDPR requires you to consult the DPC before the processing starts. For most well-scoped agents the measures bring the residual risk down to low or medium, and no consultation is necessary.

Keep the DPIA with your records of processing. Link it from your go-live checklist: our test plan template has a sign-off line for it. When you change the agent, update Part C first, then check whether the risks in Part E change.

Common questions

Is a DPIA mandatory for every AI agent?

No. It is mandatory where processing is likely to result in a high risk to people, and in the cases listed in Article 35(3) and in the DPC’s list. Many agents meet two or more risk criteria, for example new technology and sensitive data, so a DPIA is often necessary. If you decide that you do not need one, record the reasons.

Who should complete the DPIA: us or the vendor?

You, as the controller, are responsible for the DPIA. The vendor or developer must give you the facts: data flows, locations, sub-processors and security controls. If you have a DPO, you must ask for their advice. A good vendor will help you complete Parts C and F.

Do we have to send the DPIA to the Data Protection Commission?

Not usually. You keep it on file and show it to the DPC if it asks. You must consult the DPC before processing only when a high residual risk remains after your measures, under Article 36 of the GDPR.

How long does a DPIA for a small AI agent take?

For a small, well-scoped agent, expect a few working sessions: one to collect the facts from the vendor, one to assess risks with the people who do the work, and one for review and sign-off. It takes longer if the scope is not clear, which is a good reason to write the agent spec first.

Sources

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.