If your product uses AI and any of your users are in the EU, the transparency obligations of the EU AI Act apply to you from 2 August 2026, and synthetic-content marking follows on 2 December 2026. You do not need an EU entity to be in scope. Placing the product on the EU market is enough. For most SaaS companies the required work is real but bounded: disclosure, labelling, logging, and documentation.
This is an engineering-side guide to what actually has to exist in your product, written for founders and CTOs rather than lawyers. It covers scope, the 2026 and 2027 deadlines, penalties, and a build checklist. It is not legal advice, and anything touching high-risk classification should go past counsel before you rely on it.
Are you in scope?
Three questions. Answer them in order.
Do you place an AI system on the EU market?
If EU users can sign up and use an AI feature, yes. Territory is determined by where the output is used, not where you are incorporated. A US company with EU customers is in scope.
Are you a provider or a deployer?
You are a provider if you develop an AI system and put it on the market under your own name. You are a deployer if you use someone else's system in your own operations. Most SaaS companies that embed a model into a customer-facing feature are providers of that feature, even though they did not train the model. This surprises people and it is the most common scoping mistake.
What risk tier is it?
The Act classifies by use case, not by technical architecture. The same model is unregulated in one product and high-risk in another.
| Tier | Examples | What it means for you |
|---|---|---|
| Prohibited | Social scoring, manipulative techniques, untargeted facial scraping | Cannot be placed on the market at all |
| High-risk | Hiring and recruitment, credit scoring, education, healthcare, law enforcement, critical infrastructure | Conformity assessment, human oversight, technical documentation, logging, risk management |
| Limited risk (transparency) | Chatbots, AI assistants, generative content features, emotion recognition | The Article 50 obligations landing 2 August 2026 |
| Minimal risk | Spam filters, recommendation ranking, most internal analytics | No specific obligations |
The overwhelming majority of SaaS AI features are limited risk. That is the good news. The catch is that "limited" still means shipping specific things by a specific date.
The deadlines
| Date | What applies | Who it hits |
|---|---|---|
| 2 Aug 2026 | Article 50 transparency obligations | Anyone with a user-facing AI feature |
| Sept 2026 | CRA incident reporting begins | Software with digital elements sold in the EU |
| 2 Dec 2026 | Machine-readable marking of AI-generated content | Generative features |
| Dec 2026 | Product Liability Directive transposition | Software as a product |
| 2 Dec 2027 | Annex III high-risk obligations (recruitment, credit scoring, law enforcement) | High-risk providers |
| 2 Aug 2028 | Annex I high-risk, systems embedded in regulated products | High-risk providers |
The high-risk dates moved later under the Digital Omnibus amendments, which bought those teams time. The August 2026 transparency date did not move. If you have been reading that deadlines were delayed, check which deadline the article meant.
Note that 2026 stacks three separate regimes: the AI Act in August, the CRA in September, and the PLD in December. A single AI-powered SaaS product can fall under the AI Act, the CRA, the Data Act, and the PLD simultaneously. Treat it as one compliance programme rather than four projects.
What has to be in the product
For a limited-risk system, five things. All five are engineering work, not policy work.
1. Disclose that the user is interacting with AI
Clear, at the point of interaction, before or at first contact. A line in the terms of service does not count. In practice: a label on the chat surface, an "AI-generated" badge on outputs, and an explanation available without leaving the flow.
2. Mark AI-generated content, machine-readably
Due 2 December 2026. Synthetic images, audio, video, and text must carry a marking that machines can detect, which means embedded provenance metadata such as C2PA content credentials, not only a visible watermark. Visible labelling stays required for deepfakes and for AI-generated text published to inform the public.
3. Disclose emotion recognition and biometric categorisation
If you infer emotional state or categorise people biometrically, subjects must be informed. Sentiment analysis on a person's voice or face is in scope. Sentiment analysis on written product reviews generally is not.
4. Keep records
You need to be able to demonstrate what your system does, on what data, with what safeguards. In engineering terms: retain prompts and outputs at a defensible granularity, log model versions and configuration changes, and keep a written description of the system's purpose and limits.
5. Publish an accessible explanation
A page users can reach that says what the AI does, what data it uses, what it cannot do, and how to reach a human. This one is cheap and it doubles as trust-building content.
The engineering checklist
What we implement on client products to meet the August obligations. Roughly two to five engineer-weeks for a typical SaaS product.
- AI interaction labelling on every surface where a model produces output. Component-level, not page-level, so new features inherit it.
- Provenance metadata pipeline for generated media. C2PA credentials attached at generation, preserved through storage and CDN delivery. Test that your image pipeline is not stripping metadata, because most are.
- Model registry recording which model version, prompt version, and configuration produced each output. This is also the thing that saves you during an incident.
- Structured audit logging with a defined retention period and a documented rationale for it. Tie retention to your DPA and privacy policy so the numbers agree.
- Human oversight surface: a review queue, an override control, and a path to a human. Required for high-risk, and the cheapest insurance available for limited-risk.
- Technical documentation: purpose, data sources, known limits, evaluation results, and the risk assessment. Write it in the repo, not in a slide deck.
- Vendor chain mapping. If you build on a third-party model, know its terms, its data-processing behaviour, and its own compliance posture. You inherit its problems.
- Geo-aware behaviour if you intend to vary experience by region. Doable, but usually more expensive than complying globally, and it fragments your codebase.
Penalties
| Violation | Maximum penalty |
|---|---|
| Prohibited practices | EUR 35,000,000 or 7% of global annual turnover |
| High-risk non-compliance | EUR 15,000,000 or 3% of global turnover |
| Supplying incorrect information to authorities | EUR 7,500,000 or 1% of global turnover |
Realistically, enforcement in the first year will focus on prohibited practices and clear high-risk failures rather than on a startup missing a chat label. That is not a reason to skip it. The larger practical risk for most SaaS companies is commercial: enterprise procurement teams have started asking for AI Act attestation in the same questionnaire as SOC 2, and "we have not looked at it" loses deals well before a regulator ever calls.
What to do this quarter
- Inventory every AI feature and classify each by use case, not by model. One afternoon.
- Confirm nothing is high-risk. Hiring, credit, education, and health features are the ones to check carefully. If any are, get counsel now, because the 2027 obligations need a 12-month runway.
- Ship labelling and the explanation page before 2 August. This is the bulk of the immediate requirement.
- Plan the provenance work for December if you generate media.
- Fold the documentation into your security questionnaire pack, next to your DPA and subprocessor list. It is the same buyer asking.
Frequently asked questions
Does the EU AI Act apply to US companies?
Yes, if the AI system is placed on the EU market or its output is used in the EU. Incorporation location is irrelevant. A US SaaS company with EU customers using an AI feature is in scope, and non-EU providers may also need an authorised representative in the Union.
What happens on 2 August 2026?
The Article 50 transparency obligations start applying. Users must be told when they are interacting with an AI system, AI-generated content must be disclosed, and emotion recognition or biometric categorisation must be disclosed to the people subject to it. Machine-readable marking of synthetic content follows on 2 December 2026.
Is my SaaS chatbot high-risk under the AI Act?
Almost certainly not. Chatbots and assistants are limited-risk and carry transparency obligations only. High-risk is defined by use case: hiring, credit scoring, education, healthcare, law enforcement, and critical infrastructure. A support chatbot is limited risk; a chatbot that screens job applicants is high risk.
How much does EU AI Act compliance cost?
For a limited-risk SaaS product, typically two to five engineer-weeks, roughly $8,000 to $25,000, covering labelling, logging, documentation, and an explanation page. High-risk systems are a different order of magnitude because they require conformity assessment, a risk management system, and ongoing post-market monitoring.
Were the AI Act deadlines delayed?
Some were. Under the Digital Omnibus amendments, Annex III high-risk obligations moved to 2 December 2027 and Annex I to 2 August 2028. The Article 50 transparency obligations applying from 2 August 2026 were not delayed, which is the deadline that affects most SaaS products.
Do we need to label AI-generated text?
Yes, where text is published to inform the public on matters of public interest, and generally where a user could reasonably mistake AI output for human output. From 2 December 2026 the marking must also be machine-readable, which means embedded provenance metadata rather than a visible notice alone.
Key takeaways
- Transparency obligations apply from 2 August 2026. That date did not move, even though the high-risk dates did.
- Scope follows the EU market, not your incorporation. Embedding someone else's model usually makes you a provider.
- Classification follows the use case, not the technology. Most SaaS AI features are limited risk.
- Five deliverables: disclosure, machine-readable marking, biometric disclosure, records, and a public explanation.
- Budget two to five engineer-weeks for limited risk. Enterprise buyers will ask for this before regulators do.
If you are not sure where you land
The classification step is the one worth getting right, because everything downstream depends on it and the cost difference between limited and high risk is an order of magnitude. Describe your AI features in a few lines and we will send back a written classification per feature, the specific obligations that follow, and an implementation estimate. If your answer is "minimal risk, do nothing," that is a fine outcome and we will tell you. Related reading: the AI development services buyer's guide, SOC 2 for SaaS startups, and how retrieval systems make AI answers auditable. Primary sources: the consolidated AI Act text and the European Commission's regulatory framework page.