When did the EU Data Act start applying?

Regulation (EU) 2023/2854 entered into force on 11 January 2024 and became applicable on 12 September 2025. Article 29 phases out switching charges: reduced charges capped at directly linked costs apply until 12 January 2027, after which providers of data processing services may impose no switching charges at all.

Does the Data Act apply to a hosted LLM API?

The Act regulates data processing services, defined to cover digital services providing on-demand access to shared configurable computing resources across infrastructure, platform, and software models. A hosted inference API offered on those terms sits inside that framing, which brings the switching, interoperability, and Article 32 obligations into the relationship. Classification for a specific service belongs with counsel, and the evidence work is the same either way.

What does Article 32 require in practice for AI traffic?

Adequate technical, organisational, and legal measures to prevent third-country governmental access to and transfer of non-personal data held in the Union where that conflicts with Union or member state law. For AI traffic the technical half of that comes down to controlling and recording which endpoint in which region served each request, since a routing decision made automatically is still a transfer.

How does this overlap with GDPR?

Article 32 of the Data Act addresses non-personal data, which is the gap the GDPR transfer rules leave open. An organization running GDPR AI compliance work already produces some of the artifacts, and the Data Act extends the same destination and classification questions to commercial data that carries no personal element. One per-request record covers both categories.

What should we test before an exit clause is exercised?

The export itself, on a clock. Pick a date range, request every decision record for it in the documented schema, and time the retrieval. That rehearsal surfaces the two failures that matter, a schema nobody can read and a retention window shorter than the contract implies, at a point where both are still fixable. AI audit log retention requirements covers the periods to compare against.

EU Data Act AI Audit Evidence: Proving Where Your Prompts Went and How They Leave

The EU Data Act became applicable on 12 September 2025, and Article 29 sets a second date that changes commercial behaviour: from 12 January 2027, providers of data processing services may impose no switching charges on a customer at all. Between those two dates the charges are capped at the costs directly linked to the switching process. That is a pricing rule with an evidence problem attached, because a provider claiming a cost has to show what the switch involved, and a customer exercising the right has to show what it is entitled to take.

Regulation (EU) 2023/2854 covers data processing services, which the Act defines broadly enough to reach infrastructure, platform, and software services alike. An organization consuming hosted model APIs sits on the customer side of that relationship and, if it offers an AI feature to its own customers, on the provider side of another one. I want to walk the obligations that reach AI traffic and name the artifact each one asks for.

Article 30: switching, and what "all exportable data" means for AI

Article 30 requires source providers to facilitate switching by providing capabilities, information, documentation, and technical support, with open interfaces made available free of charge and documented well enough for interoperable software to be written against them. For services limited to infrastructural elements, the provider takes reasonable measures so the customer achieves functional equivalence after moving to a service of the same type.

For an AI deployment the awkward question is what travels. Prompts and completions are usually straightforward. The records describing which principal sent which classification to which endpoint under which policy revision are less so, because those frequently exist only inside a vendor console with no export path. The artifact is a decision-record export in a documented schema, produced on demand rather than promised in a contract clause.

Article 25: contractual terms that have to be true

Article 25 governs the contractual terms concerning switching, including the maximum notice period for initiating a switch and the mandatory transitional period. A contract that commits to returning customer data inside a fixed window commits the engineering team to a retrieval path that works under time pressure.

The evidence a customer asks for during a security review is a demonstrated export, timed. My candid view: the gap between a contractual exit clause and a rehearsed export is the single most under-tested claim in enterprise AI procurement right now, and it stays untested until the relationship is already ending.

Article 32: international governmental access to non-personal data

Article 32 requires providers of data processing services to take adequate technical, organisational, and legal measures, including contracts, to prevent international and third-country governmental access to and transfer of non-personal data held in the Union where that would conflict with Union law or the national law of a member state. Judgments and decisions from third-country authorities are enforceable where they rest on an international agreement in force, such as a mutual legal assistance treaty.

AI traffic makes this concrete in a way that stored data does not. A prompt carrying non-personal commercial data can reach a model endpoint in a third country because a routing library selected the nearest healthy region at 03:00, and the organization would have no record of it. The artifact is a per-request destination and region record, retained, which is what turns a data residency claim into something checkable. AI gateway data residency covers the routing side and AI data residency controls covers the enforcement side.

Articles 4 and 5: trade secrets travelling in the payload

Article 4 sets the rights and obligations of users and data holders regarding access to product data and related service data, and Article 5 covers the user's right to share that data with third parties. Both carry protections for trade secrets, with disclosure conditioned on agreed measures to preserve confidentiality.

A context window assembled from internal documents is a trade-secret carrier that nobody classified as one. The artifact is a classification decision applied to prompt content before transmission, showing which categories were permitted for that role on that route, produced by prompt-level DLP.

Articles 33 and 35: interoperability requirements

Article 33 sets essential requirements for the interoperability of data, data sharing mechanisms, and services, and Article 35 addresses interoperability of data processing services specifically, including open interfaces and published specifications. For AI this reaches the shape of the evidence itself: records held in a proprietary structure with no published schema satisfy an internal audit and fail a portability question.

The artifact is a documented record schema. AI audit log schema JSON covers what the fields need to carry for the record to answer more than one framework.

The evidence map

Where the AI-specific difficulty sits

Cloud switching debates usually centre on volume: how many terabytes move, how long the copy takes, whether the egress bill is capped. AI traffic inverts that. The payloads are small and the hard part is provenance, because the valuable artifact is the record of what was sent where, under which policy, by whom, and that record is generated continuously rather than stored once.

An organization that kept two years of prompt logs inside an orchestration vendor's platform holds a large volume of text and no independent account of the decisions around it. Rebuilding that account after the fact is not a data transfer problem. It is an archaeology problem, and the answer arrives as a description of the architecture instead of a timeline.

DeepInspect

Five of the six rows above resolve to records produced at one place: the boundary between your users or agents and the LLM APIs they call. DeepInspect sits inline at that boundary as a stateless proxy. It binds identity to each request, classifies prompt content before transmission, enforces per-role and per-route policy with a fail-closed default, records the destination endpoint and region, and writes a signed decision record before the response returns.

Because that record is written outside the application and held in a documented schema, an Article 30 export is a query rather than a negotiation, and an Article 32 question about where non-personal data went has an answer with timestamps on it. Book a technical deep dive at deepinspect.ai.