A DPIA for an AI automation starts with a technical description of the system. Prepare the data inventory, data flows, model and provider details, access and retention settings, and logs showing what the automation actually did. These inputs let the controller and data protection officer judge necessity, risks and safeguards under Article 35 of the GDPR.
Key takeaways
- A DPIA can only assess risks that the technical documentation makes visible.
- Field-level data flow mapping catches personal data that open text fields quietly carry.
- Provider terms on training use, hosting region and sub-processors must be recorded with dates.
- Retention, access and logging should be written as settings you can check, not as intentions.
What does a DPIA need from the technical side?
The engineering team supplies the facts that a DPIA depends on; the controller carries out the assessment, and the data protection officer, where one is designated, advises on it and monitors its performance. Article 35(7) of the GDPR requires at least a description of the envisaged processing and its purposes, an assessment of necessity and proportionality, an assessment of risks to people's rights and freedoms, and the measures planned to address them. Regulation (EU) 2016/679, the GDPR, in its official text on EUR-Lex
This article is general information and not legal or tax advice.
Rui Codex offers technical implementation support for AI automation projects. It does not provide legal, tax or compliance advice, audits or certification, so the legal assessment stays with your data protection officer and counsel.
Which data flows should you document for an AI automation first?
Document every point where personal data enters, moves through or leaves the automation. Start from the business process, then trace each system it touches, so that no hidden copy is missed. The trigger, source fields, transformations and destinations are the minimum.
A minimal data flow inventory
- The trigger that starts the automation, such as a new email, form submission or record change.
- Source systems and the exact fields read, including open text.
- Transformations, including every prompt sent to a model and every enrichment step.
- Destinations and storage locations, and whether data leaves the EU or EEA.
- Processors and sub-processors that handle the data at each step.
The trade-off is granularity. A coarse diagram is quick to draw but hides the fields that create the risk, while field-level mapping takes longer and needs engineering time. Open text fields are the usual blind spot, because staff type personal details into fields never designed for them. Those fields are where an automation most often fails the data minimisation test.
What must you know about the model and provider?
You need to know where the model runs, who can read the inputs, whether inputs are used to train or improve models, and how long the provider keeps them. Answer each point from the provider's written terms and technical documentation, not from sales material. If the provider cannot answer in writing, record that gap as an open risk.
The deployment model changes where control sits. A hosted API, a self-hosted open model and a model embedded in a SaaS product each place patching, logging and access control with a different party. A self-hosted model keeps data inside your own infrastructure, but it moves those responsibilities onto your team. Our AI automation services cover how such pipelines are built and documented.
How should access, retention and logging be documented?
Document who can access the automation's inputs and outputs, how long each data category is kept, and which logs record decisions. Write each item as a setting that can be checked. A DPIA that says data is deleted when no longer needed gives a reviewer nothing to test.
- Access: named roles rather than shared credentials, with a stated review frequency.
- Retention: a deletion rule for each category, including raw attachments, embeddings and cached outputs.
- Logs: the decision made, the rule version applied and the reviewer involved, with a fixed retention period.
- Backups: whether deleted personal data persists in backups, and for how long.
Logging involves a trade-off. Detailed logs help you prove what the automation did, but they store personal data, often as a second copy. Log structured events such as the decision and the rule version, and keep full prompts and outputs only where you can justify them and delete them on a fixed schedule. Our security, GDPR and compliance engineering articles describe baseline controls for this kind of setup.
Where do AI automations fail a DPIA review?
They usually fail where the documentation describes the intended behaviour but not the actual behaviour. Reviewers look for the gap between what the process owner expects and what the system does with edge cases. The following gaps appear most often.
- Scope creep: a pilot reads one mailbox, then quietly gains access to the whole shared inbox.
- Unreviewed outputs: a generated summary about a person enters a record without a human check.
- Unclear automated decisions: the model ranks or routes people in ways that affect them, with no documented review path.
- Lingering copies: prompts, embeddings and logs keep personal data after the source record is deleted.
When a gap is found, the fix is usually technical: narrower scopes, filtering before the model call, or an approval step. Each fix costs build time and sometimes output quality, so record the trade-off as a decision rather than leaving it implicit.
Hypothetical examples of technical inputs
Here are two hypothetical scenarios:
In the first, a fictional distributor routes supplier invoices from a shared inbox into its accounts system through an extraction model. The input pack lists the mailbox, the invoice fields read, the hosting region of the provider, a fixed deletion rule for raw attachments and a log recording which rule version approved each posting.
In the second, a fictional recruitment agency uses a summarisation step to prepare candidate profiles. The input pack flags special category data as a risk, names the human reviewer for each shortlist decision and shows that the model never receives candidate contact details. Both examples are invented to show the shape of the documentation, not outcomes.
How do you assemble the input pack in practice?
Assemble the pack in sequence, so that each document feeds the next. Start with the process, then the data, then the systems and finally the controls. Hand the complete pack to the data protection officer, who decides what further advice is needed.
- Write the processing purpose in one paragraph, including what the automation must not do.
- Produce the field-level data flow diagram and inventory.
- Record the provider, hosting region, sub-processors and training terms, with dated sources.
- List access roles, retention rules and log contents, each with the setting that enforces it.
- Test the failure cases on a staging copy and record the results.
- List the open risks and name an owner for each one.
For an automation that moves data between systems, the pack also needs the interface contract. Our system integration services page explains how those interfaces are specified and tested.
Conclusion
A DPIA for an AI automation is only as good as its technical inputs. Map data flows at field level, document the model and provider with dated sources, express retention and logging as checkable settings, and test where the automation fails before the assessment is finalised. These steps give your data protection officer concrete material to review and make later changes easier to assess. If you want help preparing the technical side, request a project consultation through our contact page, or email info@ruicodex.com to discuss your project.
Frequently Asked Questions
Who prepares the technical inputs for a DPIA?
The engineering team that builds the automation prepares most technical inputs, while the controller carries out the assessment with the advice of the data protection officer, where one is designated. Engineers document data flows, providers, access and logs, and the data protection officer advises on the GDPR questions. Rui Codex provides technical implementation support for this work but does not give legal advice.
Do I need a DPIA before building an AI automation?
Whether a DPIA is required depends on the processing; the controller decides, and can ask the data protection officer, where one is designated, for advice. Article 35 of the GDPR, in the official text at https://eur-lex.europa.eu/eli/reg/2016/679/oj, requires one where processing is likely to result in a high risk to people's rights and freedoms. Start the screening from the data flow inventory, because the answer depends on the facts it records.
Which provider documents should I keep for the DPIA?
Keep the written terms, the data processing agreement, the sub-processor list, the hosting region documentation and the technical pages describing how inputs are used and retained. Save each document with the date you read it, because terms change. Where the documents are silent on training use or log retention, ask the provider in writing and record any unclear answer in the risk register.
Should a DPIA include prompts and model outputs?
Include a small set of representative prompts and outputs as samples, not every live record. Samples show the reviewer what the system reads and writes, and they reveal when open text carries personal data. Keep the samples access-controlled and delete them on the same schedule as the source data.
Sources
- Regulation (EU) 2016/679 (General Data Protection Regulation) - EUR-Lex, Publications Office of the European Union