SaaS

How to audit an LLM wrapper for EU market access under the AI Act

Calling a large language model through an API does not keep your product outside the AI Act. In most cases it makes you the provider of an AI system. What you owe then depends on how the product is used, not on the model behind it. This is the seven-step audit Sila runs before a SaaS product is sold in the EU.

By Younes Lahrach, data protection specialist and external DPO, director of Sila International Advisory. First published on November 4, 2026, last reviewed against the texts on November 4, 2026.

Step 1: map what the model actually does

Start from features, not from the API agreement. For each feature that calls the model, record the input, the output, who reads it and what decision follows. A writing assistant and a CV-screening tool may use the same endpoint; they sit in different regimes.

Step 2: settle your role

Provider, Article 3(3)

You develop an AI system, or have one developed, and place it on the market or put it into service under your own name. This is the usual position of a SaaS vendor wrapping an LLM.

Deployer, Article 3(4)

You use an AI system under your authority in a professional context. Your customers are deployers, and so are you for internal tools.

Role shift, Article 25

Any operator becomes the provider of a high-risk system if it puts its name on it, substantially modifies it, or changes its intended purpose so that it becomes high-risk.

The underlying model

Obligations on general-purpose AI model providers (Articles 53 and 55) rest with the model vendor. Article 53(1)(b) requires them to give you the documentation you need to meet your own obligations.

Step 3: rule out prohibited practices

Article 5 has applied since 2 February 2025: subliminal manipulation, exploiting vulnerabilities, social scoring and emotion recognition in the workplace and in education, among others. Regulation (EU) 2026/1744 adds points (ba) and (bb) to Article 5(1), applicable from 2 December 2026. A feature that comes close must be removed or redesigned, not documented.

Step 4: classify each use case by risk tier

Obligations of a SaaS provider by risk tier of the use case
Tier	Example use	Main obligations	Applies from	Source
Prohibited	Social scoring, inferring employee emotions	Remove the feature	2 February 2025	Regulation (EU) 2024/1689, Article 5 and Article 113(a)
High risk (Annex III)	CV screening, creditworthiness, exam scoring	Risk management, data governance, technical documentation, human oversight, CE marking (Articles 9 to 17 and 43)	2 December 2027	Regulation (EU) 2026/1744, Article 113(c)(i) as amended
Transparency (Article 50)	Chatbots, generated text, images or audio	Tell people they are interacting with AI; mark generated content	2 August 2026	Regulation (EU) 2024/1689, Article 50 and Article 113, second paragraph
Minimal risk	Spell-check, internal summaries	No specific obligation beyond AI literacy (Article 4)	2 February 2025	Regulation (EU) 2024/1689, Article 4 and Article 95

Article 6(3) allows an Annex III use that poses no significant risk to fall outside the high-risk tier; the assessment must then be documented (Article 6(4)).

Step 5: implement Article 50 transparency

A conversational assistant must tell users they are interacting with an AI system unless that is obvious (Article 50(1)). Synthetic content must be marked in a machine-readable format (Article 50(2)). For systems placed on the market before 2 August 2026, marking is required from 2 December 2026 (Article 111(4) as amended).

Step 6: run the GDPR review in the same audit

The AI Act sits on top of the GDPR, not instead of it. A SaaS that sends customer data to an LLM API needs a lawful basis, an Article 28 processing agreement with the model vendor, a transfer mechanism for data leaving the EU (Chapter V) and, where the processing calls for it, a DPIA (Article 35).

The question customers ask first: is our data used to train the model? The answer lies in the vendor's contractual terms and belongs in your privacy notice.

Step 7: assemble the market-access file (checklist)

Inventory

AI features, model, version and vendor.

Qualification

Role and risk tier of each use case, with reasons.

Upstream documentation

Documentation obtained from the model vendor under Article 53(1)(b).

Transparency

Interface disclosures and marking of generated content.

GDPR

Records, processing agreement, transfers, DPIA where needed.

Customers

Instructions for use and contract terms restricting high-risk uses.

The method, in four steps

01Gap assessment
Establishing whether, and on what basis, the instrument applies to the activity as it is actually carried out, then comparing what the instrument requires against what already exists inside the organisation. The assessment ends with a ranked list of gaps, not an overall score. An organisation that falls outside the scope learns so at this stage, and the engagement stops there.
02Mapping
Recording what the measures will apply to: processing activities and data flows, systems, service providers, contracts in force, roles and responsibilities. The map is the working document for everything that follows, and it is the document produced to a supervisory authority in the event of an inspection.
03Technical and organisational measures
Implementing the measures the mapping has shown to be necessary: internal procedures, policies, contractual clauses to be renegotiated with providers, governance decisions, and training for the people concerned. Measures that sit inside the information system are specified by Sila and implemented by your own teams or your IT provider.
04Evidence file
Assembling the file that can be produced to a supervisory authority: records, impact assessments, dated and versioned policies, decision trails, proof of training. Compliance is not demonstrated by the absence of an incident but by the production of this file, which means it has to be kept current once the engagement ends.

A tailored quotation within 48 hours

The scope of an engagement cannot be guessed at: it depends on whether and how the instrument applies to you, on the number of processing activities, systems and contracts involved, and on what already exists inside your organisation. No amount is therefore published on this site.

The process starts with a scoping discussion, at no charge and without commitment, to define what the engagement will cover. A tailored quotation follows within 48 hours, setting out the scope, the deliverables and the amount.

How an assessment is conducted, and what it does not cover

What this means for you

You are probably a provider, even if you never train a model.
Risk tier is decided use case by use case, including your customers' uses.
Article 50 transparency already applies to your chatbots.
Your LLM vendor contract must cover both the AI Act and the GDPR.

Audit your product before the next funding round or enterprise deal

A first call qualifies your role and use cases and scopes the audit, quoted on request.

Book a call

Frequently asked questions

My SaaS calls the OpenAI API. Who is the provider?

OpenAI is the provider of the general-purpose AI model. You are, in principle, the provider of the AI system you sell under your own name (Article 3(3) of Regulation (EU) 2024/1689). Both roles exist at once.

Is a customer-support chatbot high risk?

Usually not: it falls under the Article 50 transparency obligations, applicable since 2 August 2026. It would become high risk if used for an Annex III purpose, such as deciding access to an essential service.

When do the high-risk obligations apply?

From 2 December 2027 for Annex III systems and 2 August 2028 for systems linked to Annex I products, under the timeline amended by Regulation (EU) 2026/1744.

What fines does the AI Act provide?

Article 99 provides up to EUR 35 million or 7% of worldwide annual turnover for prohibited practices, and up to EUR 15 million or 3% for most other obligations. For SMEs, the lower of the two amounts applies.

Do I need a new audit when I switch models?

At least revisit the qualification. Switching models does not change your role, but a new feature or intended purpose can change the risk tier, and the new vendor must provide its own documentation.

What this regime triggers elsewhere

Eight questions to place your company against the GDPR, the AI Act, NIS2 and DORA.

Exposure check: which EU rules apply to you

The full post-omnibus timeline and obligations, article by article.

The AI Act: timeline and obligations

Lawful basis, processors and transfers: the rules for data sent to the model.

GDPR obligations

A DPO mandate to carry the GDPR side of your product over time.

Outsourced DPO

About the author

Younes Lahrach is the founder and director of Sila International Advisory. A GDPR legal consultant and external data protection officer since July 2025, he holds a Master's degree (M2) in European Union law, specialising in data protection, from the University of Lille, and an MBA in International Finance from Middlesex University. He is a legal consultant, not a lawyer (avocat).

His background →

Reference texts

The information on this page is general in nature and does not constitute legal advice on a specific situation. Sila is not a law firm and does not provide representation before the courts.

All industries Request a quotation