Privacyboard

The EU AI Act for a SaaS with AI features: chatbots and generated content

By the Privacyboard teamPublished 17 November 20269 min read
Short answer

Since 2 August 2026, Article 50 of the EU AI Act requires that people know when they talk to an AI chatbot, that AI-generated content from a provider's system is marked in a machine-readable way, and that deep fakes are disclosed. Generative systems already on the market before that date have until 2 December 2026 for the marking. High-risk rules were pushed to December 2027.

If your SaaS has a chatbot or generates text, images, audio or video for users, the EU AI Act's transparency rules (Article 50) have applied to you since 2 August 2026. People must know they are talking to an AI, and content your system generates must carry a machine-readable mark. Generative systems already on the market before 2 August 2026 have until 2 December 2026 for that mark.

Most small SaaS teams are not running "high-risk" AI, and those rules were postponed anyway. This guide covers what applies now, who counts as provider or deployer, and what to do this week, as of 17 November 2026. It is general information, not legal advice.

#What applies to a SaaS with AI features today?

The AI Act (Regulation (EU) 2024/1689) phases in by date. A first amending act, the Digital Omnibus on AI (Regulation (EU) 2026/1744), was adopted on 8 July 2026 and entered into force on 27 July 2026. It moved some dates but left the transparency rules in place.

DateWhat appliesDoes it touch a typical small SaaS?
2 February 2025Prohibited practices (Article 5) and AI literacy (Article 4)Yes for AI literacy, if your team uses or builds AI
2 August 2025Rules for providers of general-purpose AI modelsRarely: these are the model makers, not the apps built on them
2 August 2026Transparency rules of Article 50: chatbots, generated content, deep fakes, emotion recognitionYes, if you ship a chatbot or generative feature
2 December 2026End of the extra time for machine-readable marking (Article 50(2)) for generative systems already on the market before 2 August 2026Yes, if your generative feature launched before August
2 December 2027High-risk rules for uses listed in Annex III (recruitment, credit scoring, education, essential services)Only if your product does one of these
2 August 2028High-risk rules for AI in regulated products (Annex I, such as machinery or medical devices)Rarely

The Omnibus also reworded Article 4: providers and deployers must now "take measures to support" AI literacy, instead of ensuring a sufficient level of it. You still need to do something and be able to show it.

#Are you a provider or a deployer?

The AI Act gives different duties to two roles.

  • Provider: the company that develops an AI system, or has it developed, and offers it under its own name or brand.
  • Deployer: the company that uses an AI system in its own professional activity.

The trap for SaaS teams is that building on someone else's model does not make you a deployer. If Acme, a 6-person invoicing SaaS, calls a third-party model API and ships an "Ask Acme" assistant inside its product under its own brand, Acme is very likely the provider of that assistant. The model maker is the provider of the model.

The Commission's guidelines on Article 50, published in final form on 20 July 2026, say a provider may rely on upstream tools for marking, but responsibility for compliance stays with the provider.

By contrast, when Acme's support team uses a chatbot product from a vendor on Acme's help site, the vendor is usually the provider of that chatbot and Acme is the deployer. Check what the vendor's settings and contract say about disclosure, and keep it switched on.

#What does Article 50 ask of a chatbot?

Article 50(1) asks providers to design AI systems that interact directly with people so that those people are told they are dealing with an AI.

According to the Commission's guidelines:

  • The notice must come no later than the first interaction.
  • No format is imposed: text, audio or a visual cue all work if people actually notice it.
  • A line buried in the terms and conditions is not enough.
  • One clear notice is usually enough, but sensitive contexts, such as health advice, may need reminders.
  • The exception for cases where AI use is "obvious" to a reasonably well-informed user is read narrowly. A human name or avatar on the bot works against you.

For Acme, that means a short line at the top of the assistant panel ("You're chatting with an AI assistant") and the same in any email reply the bot sends on its own.

#What does it ask of generated text, images, audio and video?

Article 50(2) asks providers of systems that generate synthetic content to mark outputs in a machine-readable format so they can be detected as AI-generated, as far as technically feasible. The guidelines mention techniques such as watermarks, metadata and cryptographic methods.

Two exclusions matter for SaaS features:

  • Standard editing: spell-check, grammar fixes and basic translation that do not substantially change the input.
  • Very short outputs, such as single words or interface labels.

A feature that drafts a full email, a contract clause or a product description is likely in scope. A feature that only fixes typos is likely out.

On 10 June 2026 the Commission published a voluntary Code of Practice on marking and labelling AI-generated content. It sets out practical steps, and following it is the simplest way to show your approach. If you do not sign, be ready to explain your own measures.

#When does a deep fake or published text need a label?

Article 50(4) applies to deployers, the people who publish the content:

  • Deep fakes (images, audio or video that falsely appear authentic) must be disclosed as AI-generated or manipulated. For evidently artistic, satirical or fictional work, a lighter disclosure applies. A machine-readable mark alone is not enough here: people must see the disclosure.
  • AI-generated text published to inform the public on matters of public interest must be disclosed, unless a human reviewed it and a person or organisation holds editorial responsibility.

If your product lets customers generate images of real-looking people, your customers carry the deep-fake duty, but you can make it easy: a "made with AI" option on export and clear guidance in your help center.

Article 50(3) adds a duty for deployers of emotion recognition or biometric categorisation to inform the people exposed. Emotion recognition at work and in schools is prohibited outright under Article 5(1)(f), except for medical or safety reasons. A "sentiment score" on your own sales reps' calls deserves a careful look.

#Which duties match which SaaS feature?

FeatureYour likely roleWhat applies now
AI assistant inside your product, your brandProviderChatbot notice (50(1)), output marking (50(2))
Vendor chatbot on your help siteDeployerKeep the vendor's AI notice visible, AI literacy
"Write it for me" text generatorProviderOutput marking (50(2)), unless standard editing only
Image or video generator for usersProvider; your users are deployersOutput marking (50(2)); users disclose deep fakes (50(4))
Your team uses a chatbot for internal workDeployerAI literacy
CV screening or candidate ranking featureProvider of a high-risk systemHigh-risk rules from 2 December 2027

#Does the GDPR still apply to AI features?

Yes. The AI Act sits on top of the GDPR; it does not replace it. When your AI feature processes personal data:

  • Record it in your register with its purpose and legal basis.
  • Sign a processor agreement with the model vendor and check where prompts and outputs are stored, for how long, and whether they train the model.
  • Check transfers outside the EEA.
  • Run a DPIA when the use involves new technology at scale, profiling or decisions with real effects on people, and remember the Article 22 limits on fully automated decisions.

#How much are the fines?

Breaching Article 50 can cost up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two applies. National market surveillance authorities enforce the transparency rules.

#Checklist: AI features under the AI Act

  • List every AI feature in your product and every AI tool your team uses.
  • For each one, note your role: provider (your brand) or deployer (someone else's system).
  • Add a visible AI notice to every chatbot, at or before the first message.
  • Check that bot-sent emails and messages carry the same notice.
  • For each generative feature, decide how outputs are marked (metadata, watermark or the vendor's built-in marking) and record it.
  • Ask your model vendor in writing what marking its API provides.
  • Note the launch date of each generative feature: before 2 August 2026 means marking by 2 December 2026.
  • Give users a way to label exported images, audio or video as AI-generated.
  • Check that no feature infers emotions of employees or students.
  • Flag any feature touching recruitment, credit or education for the December 2027 high-risk rules.
  • Record your AI literacy measures: who was trained, on what, and when.
  • Update your register, DPAs and privacy policy for each AI feature that uses personal data.

#What to do this week

  1. Monday (1 hour): list your AI features and tools, with role and launch date.
  2. Tuesday (2 hours): add the AI notice to your chatbot and bot emails, and ship it.
  3. Wednesday (1 hour): email your model vendor about output marking; check the Code of Practice for the techniques it lists.
  4. Thursday (1 hour): write a one-page AI usage rule for your team and record it as an AI literacy measure.
  5. Friday (1 hour): update your register and privacy policy for each AI feature.

Privacyboard's AI Act tab, under Assessments, lets you mark which activities use AI, classify each use and record AI literacy measures next to your GDPR register.

#FAQ

#Does the AI Act apply if we only call the OpenAI or another model API?

Yes, if you put an AI feature in front of users. Building on a third-party model usually makes you the provider of your own AI system, with the transparency duties that come with it. The model maker has separate duties as a general-purpose model provider.

#Is "AI-powered" in our marketing enough to inform chatbot users?

No. The notice must reach the person when they interact with the bot, at the latest at the first interaction. A line on your homepage or in your terms does not do that.

#Did the Digital Omnibus postpone the chatbot rules?

No. The Omnibus postponed the high-risk rules to 2 December 2027 and 2 August 2028. Article 50 has applied since 2 August 2026. The only extra time is for machine-readable marking by generative systems already on the market before that date, until 2 December 2026.

#Do our AI-written blog posts need a label?

Only text published to inform the public on matters of public interest falls under Article 50(4), and even then a human review with someone taking editorial responsibility removes the duty. Your own marketing posts reviewed by a named team member are unlikely to need a label, but national consumer and advertising rules may still apply.

#Do we need a certificate for AI literacy?

No. The Act asks for measures suited to each role and risk, such as training, usage rules or a guide. Keep a record of what you did and when, so you can show it.

#Sources

Privacyboard is a guidance tool, not legal advice. This article explains general rules in plain language and was accurate on the date above; laws and guidance change. For an unusual situation, ask a qualified professional. Spotted a mistake? Write to hello@privacyboard.io.

Keep reading