The support ticket summariser has been live for six weeks. It sends every new ticket, including the customer's name, email, order number and whatever they typed about their problem, to a GPT model and posts a three line summary into the agent's queue. Agents love it. Then the data protection officer forwards a customer's access request asking "which third parties received my data", and nobody on the product team can answer with confidence.
That is where most teams meet GDPR and LLM APIs: after launch, under a deadline, with a feature that works. The good news is that the legal questions are the same ones that apply to any cloud processor. The bad news is that the defaults of a fast prototype tend to fail several of them at once. What follows is general information current as of November 2026, not legal advice; your DPO or counsel makes the final call.
Who is controller and who is processor when your app calls the API
Every GDPR analysis starts with roles, because they decide who carries which duty. In a typical B2B or B2C product that calls the OpenAI API, the roles split like this.
Controller: the organisation that decides why and how personal data is processed. If your company decides to summarise tickets with an LLM, you are the controller, and you answer to the customer and to the supervisory authority.
Processor: an organisation that processes personal data on the controller's behalf and on its documented instructions. When OpenAI processes API inputs to return an output for you, under its business terms and Data Processing Addendum, it acts as your processor.
Sub-processor: a processor engaged by your processor. Cloud hosting providers that OpenAI uses fall into this category, and Article 28 requires that you can find out who they are and object to changes.
Pseudonymisation: replacing direct identifiers, such as a name or email, with a token, while the key to reverse it is kept separately. Pseudonymised data is still personal data under GDPR.
If you are a SaaS vendor processing data for your own business customers, add one more layer: your customer is the controller, you are their processor, and OpenAI becomes your sub-processor. That means your customer contracts probably already require you to notify them before adding a new sub-processor. Many teams discover this clause only after shipping.
Lawful basis and purpose: the question the DPO will ask first
Sending data to an API does not need a special lawful basis of its own. It needs the same basis that covers the purpose you are serving. Article 6 offers six bases, and for LLM features in products two come up most often.
Contract performance fits when the processing is genuinely necessary to deliver what the customer asked for. Summarising a support ticket so an agent can resolve it faster arguably serves the support contract. Legitimate interests fits when the processing benefits you or the customer but is not strictly necessary, and it requires a documented balancing test: your interest, the necessity of the processing, and the impact on the person.
Consent is rarely the right basis for back office LLM processing. It has to be freely given and withdrawable at any time, which is awkward for a feature built into a workflow.
Purpose limitation is the trap. Data collected to handle a support case cannot automatically be reused to, say, train a classifier for marketing segmentation. If the LLM feature changes what you do with the data, check compatibility under Article 6(4) or find a new basis.
Two categories need extra care. Special category data under Article 9, such as health, religion or sexual orientation, may appear in free text tickets without anyone intending to collect it. And if an LLM output feeds a decision with legal or similarly significant effects on a person, such as refusing a refund automatically, Article 22 rules on automated decision making apply.
What the API does with your data, and what you must verify yourself
OpenAI's stated position for its API has been, since 1 March 2023, that data sent through the API is not used to train its models unless the customer opts in. That default is the single biggest difference from consumer chat tools, and it is the first thing a DPO will want to see in writing, in the current terms, not in a blog post.
Retention is the second question. By default, API inputs and outputs may be retained for up to 30 days for abuse and misuse monitoring, then deleted, unless the law requires longer. Eligible customers can apply for Zero Data Retention on supported endpoints, which removes that storage for those requests. Some features, such as stored conversation state or files uploaded for retrieval, keep data by design until you delete it, so ZDR does not cover everything.
The third is location. OpenAI has offered European data residency for eligible API projects since early 2025, which keeps stored data in the region. Check which endpoints and models are covered for your project, because coverage has changed over time.
None of these protections is automatic for every account. The practical rule is to confirm, for your own organisation and project: the signed DPA, the retention setting, whether ZDR or residency is active, and the sub-processor list. Screenshots of the settings and a copy of the DPA belong in your records of processing.
Regulators do look at this space. The Italian supervisory authority, the Garante, temporarily blocked ChatGPT in 2023 and in December 2024 fined OpenAI 15 million euros over its consumer service. That case concerned the consumer product, not API customers, but it shows that LLM processing gets scrutiny.
Transfers outside the EEA without guesswork
If personal data leaves the European Economic Area, Chapter V of the GDPR applies. For a US provider, three mechanisms matter in practice.
The EU-US Data Privacy Framework has had an adequacy decision since 10 July 2023. If the recipient is certified on the official DPF list, transfers to it can rely on that decision. Certification is per company and must be current, so check the list rather than assuming.
Standard Contractual Clauses, in the version the European Commission adopted in June 2021, are the common fallback, and provider DPAs usually incorporate them. SCCs require a transfer impact assessment: a documented look at whether the law of the destination country undermines the protection the clauses promise.
Regional processing, through EU data residency or through a cloud route such as Azure OpenAI deployed in an EU region, reduces the transfer surface. It may not remove it entirely, because support access, abuse monitoring or sub-processors can still involve access from outside the EEA. Read the provider's documentation on remote access before claiming "data never leaves the EU" in a privacy notice.
Keep the DPF in perspective. Its predecessor, Privacy Shield, was struck down by the Court of Justice in the Schrems II judgment of July 2020. Many legal teams keep SCCs in place as a second layer for that reason, so a future ruling does not leave transfers without a basis overnight.
Five ways to route customer data, compared
The right architecture depends on how sensitive the data is and how much engineering you can spend. This table compares the common options for an EU product team.
| Option | Personal data reaching the provider | Retention at provider | Transfer exposure | Engineering effort |
|---|---|---|---|---|
| Raw records to standard API | All fields in the prompt | Up to 30 days for abuse monitoring | DPF or SCCs needed | Lowest |
| Redact or pseudonymise first, then API | Free text only, identifiers replaced with tokens | Up to 30 days unless ZDR | Reduced, still present | Medium: a redaction step and a token map |
| API with Zero Data Retention | All fields sent, nothing stored after processing | None on covered endpoints | DPF or SCCs still needed for processing | Low, but approval required |
| EU residency or EU cloud region | All fields, processed and stored in region | Per provider terms | Lower, check remote access | Low to medium |
| Self-hosted open weight model | None to an external AI provider | You decide | None beyond your own hosting | Highest: GPUs, updates, evals |
In practice the strongest everyday setup combines rows 2 and 3 or rows 2 and 4. Redaction cuts what you send; ZDR or residency cuts what is kept. The self-hosted route avoids the provider question entirely, but it moves security, model quality and patching onto your own team, which is a real cost for a product squad of 5 to 10 engineers.
Redaction has limits worth stating. Names in free text can be detected with a named entity model, but accuracy is never 100%, and context such as "my husband's surgery last Tuesday" can identify someone without any name at all.
Estimating exposure before the first request goes out
A DPIA asks you to assess likelihood and severity of risk. A rough number for how much personal data you actually send helps that conversation more than adjectives do. The heuristic below is not a legal test; it is a sizing tool for engineers and DPOs to argue from.
monthly identifiers sent = requests per month
x share of requests containing personal data
x average identifiers per request
exposure days = monthly identifiers sent x retention days at provider
An illustrative example: a support tool handles 40,000 tickets a month. Sampling 200 tickets shows that 90% contain personal data, with an average of 3 identifiers each (name, email, order number). Monthly identifiers sent: 40,000 x 0.9 x 3 = 108,000. With 30 days of abuse monitoring retention, exposure days come to 3,240,000.
Now add a redaction step that removes emails and order numbers and replaces names with tokens, leaving an estimated 0.4 identifiers per ticket in free text that the detector misses. Monthly identifiers sent: 40,000 x 0.9 x 0.4 = 14,400, a reduction of about 87%. If the project also gets Zero Data Retention, exposure days fall to zero on covered endpoints, while the 14,400 identifiers still pass through processing.
The numbers are crude, but they make the trade-off visible. A team can say "redaction costs us two weeks of work and cuts what we send by seven eighths" and the DPO can weigh that against the residual risk.
Launch checklist for an LLM feature that touches personal data
Use this before the feature goes live, or now, if it already has.
- Purpose written down in one sentence, and the lawful basis that covers it, with a legitimate interests assessment if you rely on that basis.
- Signed DPA with the provider in place, and the provider and its sub-processors added to your own sub-processor list.
- Customer contracts checked for sub-processor notice periods, commonly 14 to 30 days, and notices sent where required.
- Training on API data confirmed as off for your organisation, and the retention setting documented.
- Zero Data Retention or EU residency requested where the data justifies it, and coverage of the specific endpoints confirmed.
- Transfer mechanism recorded: DPF certification checked on the official list, or SCCs with a transfer impact assessment.
- Redaction or pseudonymisation step in front of the API, with a sampled test of what still gets through.
- Rules for special category data: either blocked from the prompt or covered by an Article 9 condition.
- DPIA completed if the processing is likely to be high risk, which is common when free text from many customers is processed with new technology.
- Privacy notice updated to mention AI service providers as recipients and any transfers outside the EEA.
- Records of processing under Article 30 updated with the new operation.
- Logging on your side reviewed, since your own logs of prompts and outputs are often the larger store of personal data.
- A process for access and erasure requests that covers prompts, outputs and any stored conversation state.
Questions product teams ask about GDPR and LLM APIs
Does pseudonymising data before sending it take it outside GDPR? No. Recital 26 of the GDPR treats pseudonymised data as personal data, because the controller can still re-identify it with the key it holds. Only data that is truly anonymous, where no one can reasonably re-identify the person, falls outside the regulation. Pseudonymisation still counts as a safeguard and lowers risk in a DPIA.
Do we have to tell customers that their data goes to OpenAI? Articles 13 and 14 require you to tell people about the recipients or categories of recipients of their data and about transfers outside the EEA. Most companies list AI service providers as a category in the privacy notice and name them in a sub-processor list, which customers in B2B contracts usually expect to see.
What about the EU AI Act? The AI Act is a separate regulation that adds obligations by risk category, such as transparency duties when people interact with a chatbot. It does not replace GDPR. A support summariser will usually sit in a low risk category under the AI Act while still being a full GDPR processing operation.
Can employees paste customer data into ChatGPT instead of using the API? Consumer ChatGPT accounts have different terms from the API, including different defaults for using content to improve models. Most companies block or restrict this in policy and offer a business plan or an internal tool built on the API instead, so the data flows under a signed DPA.
The recommendation, then, depends on what is in the prompt. If the data is ordinary contact and order information, a signed DPA, training off, a documented transfer basis and a redaction step are a defensible baseline. If free text routinely carries health or financial details, add Zero Data Retention or EU residency and run a DPIA before launch. If you cannot get either, keep that data out of the prompt.