
ByAhsan Nadeem@ashtech52Lead Full-Stack Engineer
Some developer guides say that calling an LLM API makes you a deployer. We read the EU AI Act differently: if you ship the feature under your own name, you are probably the provider of that AI system. So an EU AI Act compliance checklist for SaaS starts with one question per feature: provider or deployer?
The delay only helps high-risk features. As of 4 November 2026, the Digital Omnibus has moved the high-risk rules to December 2027 and later. Chatbot disclosure, the prohibited uses and AI literacy were not delayed, and the marking duty for existing generative systems falls due on 2 December 2026.
This is general information, not legal advice. We quote the consolidated text on the Commission's AI Act Service Desk, dated 27 July 2026, and haven't checked it against the Official Journal copy on EUR-Lex. Confirm anything you rely on against the official text, and take advice on your own product.
If you're adding AI to a product and want the design to carry these duties from the start, that's part of our LLM integration work.
The EU AI Act compliance checklist for SaaS, in short
- Today: give each AI feature a role, provider or deployer, in writing.
- Today: tell people when they are dealing with an AI system, unless that's obvious.
- Today: check nothing you ship falls under a prohibited practice, and take measures to support AI literacy.
- By 2 December 2026: if an existing feature generates content, have machine-readable marking and detection in place. Two new prohibitions also start that day.
- This quarter: screen each feature for high-risk uses. HR, credit and education are the usual ones.
- By 2 December 2027: if a feature is high-risk, meet the provider or deployer duties.
The Digital Omnibus delayed high-risk duties only: the dates as they stand
Regulation (EU) 2026/1744, the "AI Omnibus", was published on 24 July 2026 and took effect on 27 July, according to K&L Gates. It moved the high-risk dates but left the transparency start date where it was.
Dates as they stand on 4 November 2026
Date | What applies | Delayed by the Omnibus? |
|---|---|---|
2 February 2025 | Prohibited practices, AI literacy | No, though the literacy wording was softened |
2 August 2025 | General-purpose AI model duties, most penalty rules | No |
2 August 2026 | Article 50 transparency, and the rest of the Act unless excepted | No |
2 December 2026 | Article 50(2) marking for generative systems already on the market; two new prohibitions | New date |
2 August 2027 | General-purpose models on the market before 2 August 2025 must comply | No |
2 December 2027 | High-risk duties for Annex III uses | Yes, from 2 August 2026 |
2 August 2028 | High-risk duties for AI in regulated products (Annex I) | Yes |
The source for these dates is Article 113 and Article 111.
So does the delay let you wait? Only for high-risk work. A SaaS feature that chats, summarises or drafts sits under Article 50, which has applied since August.
A caution on older pages: the Service Desk's own FAQ still describes the Omnibus as a proposal, with a six-month marking transition. The final Article 111(4) gives four months.
Provider or deployer? A per-feature test for teams using LLM APIs
Published guidance conflicts. One vendor guide, sota.io, says using an LLM API for your SaaS features "makes you a deployer, not a provider". Another, LangWatch, tells API users: "You are the provider of that system." Both are vendor views, so go to the text.
Article 3 gives three definitions that matter:
- Provider: develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark, paid or free.
- Deployer: uses an AI system under its authority, other than for personal use.
- Downstream provider: a provider of an AI system that integrates an AI model, including a model supplied by another company under contract.
That last definition, in Article 3(68), covers the API case. Recital 97 adds that a model needs further components, such as a user interface, to become an AI system. Your interface, prompts and logic look like those components.
The Commission's Article 50 guidelines point the same way. Their example of a provider is a company offering a chatbot, image generator or AI agent under its own name, to consumers or businesses. They also refer to "downstream AI system providers" having their own Article 50 duties.
Our reading, not a ruling: neither the Act nor those guidelines mention API calls by name in their role definitions, and only the Court of Justice can settle the question. The guidelines are also non-binding. Still, a product team that ships an LLM feature under its brand should assume it is the provider of that feature, and that its customers are deployers.
Run this test on each feature, not on the company.
Feature | Likely role | Why |
|---|---|---|
Assistant in your product, built on an LLM API, under your brand | Provider; your customers are deployers | Article 3(3) and 3(68); guidelines para 11 |
AI summaries or drafts in your product | Provider, with marking duties | Summaries are a listed example that needs marking |
Chatbot you built for your own staff, under your name | Provider and deployer | Guidelines para 11 treats in-house systems put into service as a provider case |
Vendor's finished AI tool, used as supplied | Deployer | You decide its use; the vendor is the provider |
Staff using an AI tool to draft public content | Deployer | Article 50(4) applies if it's a deepfake or public-interest text |
Existing generative system you modify and run under your name | Provider of the new system | Guidelines para 11 |
White-labelling a vendor's system with no changes | Unclear unless high-risk | If high-risk, Article 25(1)(a) makes you the provider. Otherwise take advice |
This guide doesn't cover when fine-tuning makes you a model provider, which is a separate question. If you're still choosing an approach, see RAG vs fine-tuning.
The Act reaches outside the EU. Providers are covered wherever they're based if they place systems on the EU market, and non-EU providers and deployers are covered where the output is used in the Union.
Duties that apply now: prohibited uses, chatbot disclosure, deepfake labelling, AI literacy
Prohibited uses
Article 5 has applied since February 2025. Two entries can catch product teams: social scoring, and inferring emotions at work or in education (with medical and safety exceptions). Check any "sentiment" feature aimed at employees or students.
Chatbot disclosure, Article 50(1)
This duty falls on the provider. People must be told they are dealing with an AI system unless that's obvious to a reasonably well-informed person. Under Article 50(5), the notice must be clear and given by the first interaction at the latest.
The guidelines read "obvious" narrowly and give useful examples:
- Disclosure needed: chatbots in platforms or helpdesks whose replies users might take for a human's.
- Probably not needed: coding assistants used only by professional developers, and internal assistants for trained, AI-literate staff.
- Good enough: a first-turn statement, or a label next to the input field.
- Not enough alone: a line in your terms, a URL or your documentation.
An AI agent that emails people is covered too: the guidelines suggest an AI label at the top of the message.
Deepfakes and public-interest text, Article 50(4)
This one falls on the deployer. Deepfakes must be disclosed. The same goes for AI-written text that a deployer publishes to tell people about public-interest matters, unless a person reviewed it and someone holds editorial responsibility. If your customers do this with your tool, the duty is theirs.
AI literacy, Article 4
The Service Desk FAQ still suggests the duty would move to the Commission and Member States. The final text doesn't do that. Article 4 as amended still says providers and deployers shall take measures to support the AI literacy of their staff and others operating AI on their behalf. It doesn't require a guaranteed level for any individual.
In practice: short, role-based guidance, with a record that you gave it. Article 4 isn't on the list in Article 99(4), though Member States set the penalty rules, so don't read that as "no consequence".
The 2 December 2026 marking grace period, and who it covers
Article 50(2) requires providers of generative systems to mark outputs in a machine-readable format and make them detectable. Article 111(4) gives systems placed on the market before 2 August 2026 until 2 December 2026.
Three limits:
- It covers marking only. Under guidelines para 153, a system that both chats and generates still had to meet the disclosure duty from 2 August.
- New systems get no grace. A generative system launched after 2 August 2026 falls outside the wording of Article 111(4), so check your release history.
- Internal tools are unclear. Article 111(4) says "placed on the market", while the guidelines add "or put into service". Take advice if you depend on that gap.
Which features need marking? The guidelines are specific.
Output | Marking? |
|---|---|
AI summaries; rewriting that changes style, structure or meaning | Yes |
Agent output that people read | Yes |
Grammar and spelling fixes, minor polish, translations, transcriptions | No |
Source code, SQL, JSON configuration | No |
Single words, captions, alt text, UI labels | No |
Machine-to-machine output no person sees | No |
Output from before 2 August 2026 | No retroactive marking, though older text published now needs its label |
There's a B2B exception, but it's narrow. Para 87 requires all of the following: strictly technical output, a limited set of professionals inside the provider and deployer, and no intent to share it outside. A summary your customer sends to a client won't qualify.
You may rely on marking from your model vendor or a third party if it's compliant, but para 74 leaves the burden of showing compliance with you. The Commission's voluntary Code of Practice of 10 June 2026 sets out practical steps. The Commission's opinion of 8 July 2026 found that it adequately covers Articles 50(2), (4) and (5). Signing is still voluntary, and Paul Weiss notes that following it hasn't been declared conclusive evidence of compliance.
One status point: the Commission's page says the guidelines were adopted, but the cover communication says formal adoption, and application, come once all translations exist. We found no confirmation of that step. Article 50 itself applies regardless.
Is any feature high-risk? Screening HR, credit and education uses
Annex III lists the uses. For SaaS, check these first:
- Employment: screening applications, evaluating candidates, targeted job ads, promotion or termination decisions, allocating tasks by behaviour or traits, monitoring performance.
- Credit and insurance: creditworthiness scoring (fraud detection is excluded), and life and health insurance risk and pricing.
- Education: admissions, evaluating learning outcomes, deciding a person's level, proctoring.
A route teams miss is Article 25(1)(c). If you take a general-purpose system and change its purpose so that it becomes high-risk, you become the provider of a high-risk system. A CV-ranking feature built on a general LLM is the obvious example. Remember that intended purpose includes what your sales materials say.
Article 6(3) offers a way out for narrow procedural or preparatory tasks that don't materially influence decisions. It never applies if the system profiles people. If you use it, document the assessment before launch and register under Article 49(2).
If a feature is high-risk, providers take on the Article 16 list: quality management, documentation, logs, conformity assessment, CE marking and registration. Non-EU providers also appoint an EU representative. Deployers take on Article 26: human oversight, at least six months of logs, and telling workers and affected people.
Don't plan around grandfathering. Article 111(2) spares existing high-risk systems only until a significant design change, which our reading says an actively developed product will soon reach.
The checklist: action, owner, evidence, date
Action | Owner | Evidence to keep | By |
|---|---|---|---|
List every AI feature, its model and its vendor | Product | Feature register | Now |
Record provider or deployer for each | Product, legal | Role note per feature | Now |
Check each against Article 5 | Legal | Dated review | Now |
Add AI disclosure at first interaction | Engineering, design | Screenshots, release note | Now |
Give role-based AI literacy guidance | People, engineering lead | Materials, attendance | Now |
Classify generated outputs against the marking table | Product | Output list | Now |
Implement marking and detection | Engineering | Design note, test results | 2 December 2026 |
Confirm no feature can produce the newly prohibited content | Engineering, trust and safety | Safeguard tests | 2 December 2026 |
Screen features against Annex III, including sales copy | Product, legal | Screening note | This quarter |
Document any Article 6(3) exemption and register | Legal | Assessment, registration | Before launch |
Collect vendor documentation and terms | Engineering, legal | Vendor file | This quarter |
Plan any high-risk work | Leadership | Budgeted plan | Well before 2 December 2027 |
The new prohibitions cover non-consensual intimate imagery of identifiable people and child sexual abuse material. Under Article 5(1a), they reach a system where that output is reasonably foreseeable and reproducible and adequate safeguards are missing. Any image or video feature should test for this.
What to ask your model vendor
General-purpose model providers owe documentation to the companies building on them, under Article 53(1)(b) and Annex XII. Models on the market before August 2025 have until 2 August 2027 to comply, so answers may vary. Ask:
- Where's your Annex XII documentation, covering intended tasks, acceptable use, integration and input and output limits?
- Do your outputs carry a machine-readable mark, and does it survive our processing? Is there a detection method we can point to? The guidelines encourage marking at model level but don't require it.
- Does your acceptable use policy ban HR, credit or education uses? This matters because Article 25(2) releases an original provider from its cooperation duty if it clearly said its system mustn't be turned into a high-risk one.
- If we're high-risk, will you sign the written agreement that Article 25(4) requires?
Our guides to evaluating a RAG chatbot and prompt injection cover more of the engineering. If you also ship installed software, run the separate Cyber Resilience Act scope test.
Frequently asked questions
Did the Digital Omnibus delay the EU AI Act?
Only in part. High-risk duties moved to 2 December 2027 for Annex III uses and 2 August 2028 for AI in regulated products. Article 50 transparency still started on 2 August 2026, and general-purpose model duties weren't delayed. Generative systems already on the market got until 2 December 2026 for marking only.
Am I a provider or deployer if I use the OpenAI or Anthropic API?
For a feature you ship under your own name, our reading is that you are probably the provider of that AI system, and a downstream provider under Article 3(68). Your customers are deployers. The model vendor keeps its duties for the model. Published guides conflict and no court has ruled, so take advice.
Does my chatbot have to say it is AI?
Yes, unless that's obvious to a reasonably well-informed person. The Commission's guidelines read that exception narrowly, and they list helpdesk chatbots as needing disclosure. A first-turn statement or a label by the input field works. A line in your terms doesn't work alone. The notice must come by the first interaction.
Does the Act apply to companies outside the EU?
Yes, in many cases. Article 2 covers providers placing AI systems on the EU market wherever they are based, and non-EU providers and deployers where the output is used in the Union. Incidental or unforeseeable EU use shouldn't alone trigger Article 50, according to the guidelines.
Is AI literacy training still mandatory?
There's still a duty, though it's softer. Article 4 as amended says providers and deployers shall take measures to support AI literacy among staff and others operating AI for them, without guaranteeing any individual's level. Short role-based guidance with a record is a sensible way to show you've acted.
What are the fines?
Up to €35m or 7% of worldwide annual turnover for prohibited practices. Up to €15m or 3% for Article 50 and listed operator duties. Up to €7.5m or 1% for misleading information to authorities. SMEs face the lower of the amount or percentage, and small mid-caps get the same for the two lower tiers.
The nearest date is 2 December 2026, and marking is an engineering job. If you're shipping LLM features into the EU and want help deciding what to build first across disclosure, marking and logs, book a free consultation. Your lawyer settles the legal position.
Sources
- AI Act Service Desk consolidated text, Article 3
- Article 2
- Article 4
- Article 5
- Article 6
- Article 16
- Article 22
- Article 25
- Article 26
- Article 49
- Article 50
- Article 53
- Article 99
- Article 111
- Article 113
- Annex III
- Annex XII
- Recital 97
- Service Desk FAQ (partly out of date)
- Commission, Article 50 guidelines (20 July 2026)
- Commission, cover communication
- Commission, guidelines page
- Commission, Code of Practice (10 June 2026)
- EUR-Lex, Regulation 2026/1744 (not checked)
- EUR-Lex, Regulation 2024/1689 (not checked)
- K&L Gates via National Law Review (31 July 2026)
- Commission, opinion on the Code of Practice (8 July 2026)
- sota.io (vendor, 4 May 2026)
- LangWatch (vendor, 5 July 2026)



