# EU AI Act transparency rules for SaaS and apps: chatbot disclosure, AI content marking and the 2 December 2026 deadline

> EU AI Act Article 50 already applies to chatbots and AI-generated content. Here is what SaaS and app teams must ship for disclosure, machine-readable marking and detection before the 2 December 2026 deadline.

Source: https://authect.com/insights/blog/eu-ai-act-transparency-rules-saas-apps-chatbot-disclosure-content-marking

*Guide by Pranas Mickevicius, Chief Executive Officer, published 2026-11-12*

> EU AI Act Article 50 already applies to chatbots and AI-generated content. Here is what SaaS and app teams must ship for disclosure, machine-readable marking and detection before the 2 December 2026 deadline.

If your product lets people chat with an AI, or shows them AI-generated text, images, audio or video, the EU AI Act's transparency rules already apply to you. [Article 50 has applied since 2 August 2026](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations). Generative systems that were already on the EU market before that date have [until 2 December 2026](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august) to mark their outputs in a machine-readable way. The heavier high-risk obligations have moved to December 2027 and August 2028.

For most SaaS and app teams, the next three weeks come down to four jobs:

- List every AI feature.
- Decide whether you are the provider or the deployer of each one.
- Ship clear disclosure in the interface.
- Agree a marking and detection plan with your model vendor.

**Key dates at a glance**

- **2 August 2026:** Article 50 transparency duties apply. These cover chatbot disclosure, labels for deepfakes and public-interest text, emotion-recognition notices, and content marking for systems launched from that date.
- **2 December 2026:** machine-readable marking deadline for generative systems placed on the EU market before 2 August 2026.
- **2 February 2027:** the Code of Practice commitment on watermark-detection interoperability.
- **2 December 2027:** stand-alone high-risk (Annex III) obligations apply.
- **2 August 2028:** high-risk obligations for AI built into regulated products (Annex I) apply.

This guide is for teams selling software into the EU, wherever they are based. It is not legal advice. Use it to brief your counsel and plan your sprints.

## What changed in 2026: the Digital Omnibus in brief

The AI Act entered into force on 1 August 2024 and phases in over several years. By late 2025, [implementation was visibly off track](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/). On 19 November 2025 the European Commission proposed a set of targeted amendments, the Digital Omnibus on AI.

![Antique brass hourglass on a slate slab with sand half fallen](https://r78zyqbigtq5h1ka.public.blob.vercel-storage.com/posts/eu-ai-act-transparency-rules-saas-apps-chatbot-disclosure-content-marking-image-1-QbEvYOJ4NapvH6aJuh5umanB9vOuJi.webp)
*The Omnibus moved the high-risk deadlines, but not the transparency ones.*

That package became law this summer. [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act) gives the dates:

- Signed on 8 July 2026.
- Published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744.
- In force from 27 July 2026, days before the original high-risk deadline of 2 August 2026.

The changes that matter most to software companies:

- **High-risk deadlines moved to fixed dates.** Stand-alone high-risk systems listed in Annex III now face obligations from 2 December 2027. AI used as a safety component of regulated products follows from 2 August 2028 ([White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act)).
- **Transparency was barely touched.** The deferral covers Chapter III (high-risk) only. Article 50 kept its 2 August 2026 start, with one short grace period for marking the outputs of systems already on the market.
- **A new ban.** AI systems used to generate child sexual abuse material, or non-consensual intimate content of identifiable people ("nudifier" apps), are now prohibited practices. [Gibson Dunn's summary of the agreed text](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) says that, for providers, the ban covers any system where such output is foreseeable and repeatable and adequate safeguards are missing. It has a transition period to 2 December 2026.
- **AI literacy softened.** Providers and deployers must now take measures to *support* sufficient AI literacy among staff, rather than *ensure* it.
- **More relief for smaller firms.** SMEs and small mid-caps keep or gain several reliefs ([White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act)): simpler technical documentation, proportionate quality management obligations, lower penalties and priority access to sandboxes.
- **A narrower "safety component".** AI that only assists users or optimises performance is not automatically a safety component, as long as its failure does not create health or safety risks.

The headlines focused on the delay. For product teams, the more important point is the opposite: the rules that touch user-facing features did not move.

## The AI Act timeline as it stands in November 2026

| Date | What applies | Status | Source |
|---|---|---|---|
| 1 Aug 2024 | AI Act enters into force | In force | [Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) |
| 2 Feb 2025 | Prohibited practices (Article 5) and AI literacy (Article 4) | In force; literacy duty softened by the Omnibus | [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act) |
| 2 Aug 2025 | General-purpose AI (GPAI) model obligations; AI Office operational | In force | [Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) |
| 10 Jun 2026 | Final Code of Practice on transparency of AI-generated content published | Voluntary, but a recognised route to compliance | [Freshfields](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx) |
| 27 Jul 2026 | Digital Omnibus (Regulation (EU) 2026/1744) in force | In force | [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act) |
| 2 Aug 2026 | Article 50 transparency obligations | In force | [European Commission](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations) |
| 2 Dec 2026 | Marking deadline for generative systems on the market before 2 Aug 2026; end of the nudifier-ban transition | **Three weeks away** | [Morgan Lewis](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august), [Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) |
| 2 Feb 2027 | Code of Practice: watermark detection must be interoperable | For Code signatories | [Freshfields](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx) |
| 2 Aug 2027 | Deadline for member states to set up AI regulatory sandboxes | Per the agreed Omnibus text | [Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) |
| 2 Dec 2027 | Stand-alone high-risk systems (Annex III) | Upcoming | [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act) |
| 2 Aug 2028 | High-risk AI in regulated products (Annex I) | Upcoming | [White & Case](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act) |

If you sell to EU customers, you may also need to plan for product-security rules under the Cyber Resilience Act, which sits alongside the AI Act. We cover that timeline in [our CRA guide](/insights/blog/eu-cyber-resilience-act-software-companies-guide).

## Provider or deployer? The question that decides your obligations

Article 50 splits duties by role. Get the role wrong and you will either over-build or miss the obligations that are actually yours.

[Morgan Lewis summarises the definitions](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august) well:

- A **provider** develops an AI system and places it on the EEA market, or puts it into service, under its own name or trademark.
- A **deployer** uses someone else's AI system without substantial modification or white-labelling, other than for personal, non-professional use.

What triggers the rules is placing a system on the EU market or putting it into service there. Where your company is registered doesn't matter. A studio in Dubai or a start-up in London selling an AI feature to customers in Madrid is in scope for that feature.

### When a SaaS company is a provider

This is the case that catches most teams out. You call a model API from a large lab, wrap it in your own product and sell it under your brand. The lab is the provider of the general-purpose *model*. You are very likely the provider of the AI *system* your customers use, because you built it and put it on the market under your name.

That means the provider-side duties in Article 50 land on you, even though you did not train the model. Those duties are telling users they are talking to AI and marking synthetic outputs. The [Commission's guidelines](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations) address exactly this question, setting out which obligation falls on which party along the value chain.

Your model vendor will cover part of the marking work. You still need to know what they provide, and to make sure your own pipeline doesn't undo it.

### When your customers are deployers

Business customers who use your product to publish content are deployers. Say they create a deepfake video with your avatar tool, or auto-publish AI-written articles on matters of public interest. The duty to label that content sits with them, not you.

This still affects your roadmap, because deployers can only label what your product lets them label. Morgan Lewis advises deployers of third-party AI systems to [obtain information about the provider's marking arrangements](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august), including any reliance on the 2 December transition. Expect those questions to appear in procurement reviews.

### If you only use AI internally

If you use a third-party AI tool inside your company and publish nothing it generates, your Article 50 exposure is small. There are two exceptions:

- Emotion recognition or biometric categorisation of people, for example in call analytics.
- Publishing deepfakes or AI-written public-interest text.

Separately, the softened [AI literacy duty](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4) still asks you to support staff who use AI systems on your behalf, and to keep a record of what you did.

## Article 50 in plain terms: the four transparency duties

The [Commission's summary](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations) is a useful anchor. Providers have two duties:

- Design AI systems so that people are explicitly told whenever they interact with one directly.
- Add machine-readable marks so that AI-generated or manipulated content can be detected.

Deployers must tell people when they are exposed to any of these:

- Emotion recognition or biometric categorisation.
- Deepfakes.
- Text on matters of public interest published without human review or editorial control.

| Duty | Who | Applies from | What you ship |
|---|---|---|---|
| Disclose AI interaction (unless obvious) | Provider | 2 Aug 2026 | A clear notice in chat, voice and email interfaces at first contact |
| Machine-readable marking and detectability of synthetic text, images, audio and video | Provider | 2 Aug 2026 for new systems; 2 Dec 2026 for systems already on the market | Signed metadata, watermarks, a detection tool |
| Inform people exposed to emotion recognition or biometric categorisation | Deployer | 2 Aug 2026 | Notices at the point of capture, plus updated privacy notices |
| Label deepfakes and AI-generated public-interest text | Deployer | 2 Aug 2026 | Visible labels or icons; an editorial-review process if you rely on that exception |

*Sources: [European Commission](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations), [Morgan Lewis](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august), [AI Act Article 50](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50).*

### 1. Tell people they are talking to AI

The provider must make sure people are told they are interacting with an AI system, unless that is already obvious. For this duty, the Commission says providers [can choose adequate measures themselves](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations), taking its guidelines into account. There is no mandated wording.

Don't stretch "obvious". A support widget with a human name and a stock photo is not obviously AI. Nor is a voice agent that sounds human. Good patterns:

- A persistent "AI assistant" label in the chat header, not only in the first message.
- An audible line at the start of every voice call.
- A clear route to a human where your service offers one.
- An AI signature on any outbound email or message written and sent without a human in the loop.

### 2. Mark synthetic content in a machine-readable way

Providers of systems that generate synthetic text, images, audio or video must make sure the outputs are [marked in a machine-readable format and detectable](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august) as artificially generated or manipulated. This is the only duty with a grace period to 2 December 2026, and only for systems placed on the market before 2 August 2026. Anything launched on or after 2 August had to comply from day one.

The [Article itself](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50) contains narrow exceptions. The Commission's guidelines give examples such as [standard editing](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations). A grammar fixer is not in the same position as a tool that writes a 1,000-word proposal from a three-line prompt. Document where each feature falls and why.

### 3. Disclose emotion recognition and biometric categorisation

If your product infers emotions, or sorts people into categories from biometric data, deployers must tell the people affected. Check [Article 5](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-5) before you get to disclosure. Some emotion-recognition uses are banned outright, and a notice does not make a prohibited practice legal. Teams building call-centre analytics, proctoring or workplace tools should get a legal view first.

### 4. Label deepfakes and public-interest text

Deployers who publish deepfakes must disclose them. The same applies to AI-generated text meant to inform the public on matters of public interest. For text, the exception is human review or editorial control. Under the Code of Practice, a deployer relying on that exception [must name the person holding editorial responsibility](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx). It must also document how its organisation ensures adequate review, but it does not need to log each individual review.

## What the Code of Practice asks for in marking and detection

On 10 June 2026 the Commission [published the final Code of Practice](https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_26_1328/IP_26_1328_EN.pdf) on marking and labelling AI-generated content, after drafts in December 2025 and March 2026. Signing it is voluntary. But generative AI providers and deployers who don't sign up [must show compliance through alternative, equally adequate means](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations), the Commission says. For most small and mid-sized companies, the Code is the practical specification to build against.

[Freshfields' analysis of the final text](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx) is the clearest summary we have found. The points that matter for engineering:

| Requirement | What the final Code says | What it means for your build |
|---|---|---|
| Layers of marking | Use several layers where one technique can't meet all four tests: effectiveness, interoperability, robustness and reliability. In practice, at least two machine-readable layers | Images, audio and video usually need signed metadata *and* an embedded watermark |
| Single-layer exceptions | Generative AI inside a closed, technically controlled physical product; free-form text, which cannot carry metadata | Text features can rely on one technique, typically a statistical watermark |
| Text threshold | Watermark free-form text longer than 200 tokens; text under 200 tokens is exempt | Short chat replies are out; long drafts, reports and summaries are in |
| Metadata quality | Digitally signed and time-stamped in a secure, tamper-evident way | You need key management and signing in the generation pipeline |
| Visible watermark | Watermarks people can see on audio, images and video are one interoperability option, not a stand-alone requirement | Visible badges are a design choice |
| Fingerprinting and logging | Optional extra measures | Useful for abuse investigations; weigh against data minimisation |
| Detection tool | Providers must offer a way to detect their marks. It must be hosted in the EU or run locally, comply with EU data protection law and, in principle, be free | Set aside engineering time for a detection endpoint or SDK |
| Fee carve-out | Providers with fewer than 1,000,000 monthly users may charge heavy individual users a proportionate fee. Access stays free for regulators, researchers, media and others | Smaller products can protect themselves from abusive detection volumes |
| Interoperability | Watermark detection must be interoperable by 2 February 2027. Options: a standard API, an embedded "signpost", a shared consortium detector or a comparable solution | Pick an approach now; retrofitting later costs more |

### Signed metadata: use the existing standards

Freshfields notes that [digitally signed metadata is already covered by existing standards](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx). For images and video, the [C2PA Content Credentials specification](https://c2pa.org/specifications/) is the best-known example. Most of the work is operational rather than code:

- Certificates.
- Key storage and rotation.
- Making sure every path that writes a generated file goes through the signer.

### The pipeline problem nobody budgets for

The most common failure is not a missing watermark. It is a pipeline that removes one. Image CDNs, resizers, thumbnail generators and "strip metadata to save bytes" settings routinely throw away embedded metadata. Re-encoding video for streaming and normalising audio do the same. Test the file your user actually downloads, not the one your model returned.

![Aerial view of a braided river delta with channels splitting across sandbars](https://r78zyqbigtq5h1ka.public.blob.vercel-storage.com/posts/eu-ai-act-transparency-rules-saas-apps-chatbot-disclosure-content-marking-image-2-HIJTxQ4suf0RQVBRoyBNF90C6cHMdp.webp)
*Every path a generated file takes is a chance for its marks to be stripped.*

### Labels and the EU icons

For deployers, the Code drops the earlier split between "fully AI-generated" and "AI-assisted" content. Instead it [sets design and placement rules](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx). It also provides three ready-to-use EU icons:

- One for fully AI-generated content.
- One for AI-modified content.
- A basic icon with an interactive second layer.

User testing across several member states found that icons with an explicit text label worked better. Labels must be:

- Recognisable without the user having to do anything.
- Visible for long enough.
- Embedded in the content where possible.
- Noticeable no later than first exposure.

Audio-only content needs a spoken disclaimer. The EU icon is optional, but any alternative must meet the same rules.

If customers publish content made in your product, give them these options in the UI. It is a small feature that removes a real obstacle for them.

## Triage: common SaaS and app AI features against the rules

Use the table below as a starting point for your own inventory, not as a legal classification. Roles depend on how you sell and who controls the system.

| Feature | Your likely role | Transparency duty | Deadline | Practical note |
|---|---|---|---|---|
| Customer-support chatbot in your app | Provider | Disclose AI interaction | Already applies | Label in the header and the first message; offer a handoff to a human |
| Voice agent you sell to businesses | Provider (your customer is the deployer) | Disclose AI interaction | Already applies | An audible disclosure at the start of every call |
| AI writing assistant producing long drafts | Provider | Machine-readable marking of text over 200 tokens (Code) | 2 Dec 2026 if launched before 2 Aug 2026 | Grammar-only tools may count as standard editing; document it |
| Image or video generation | Provider | Multi-layer marking plus detection | 2 Dec 2026 if launched before 2 Aug 2026 | Check your CDN and transcoding don't strip metadata |
| Avatar or voice-clone videos made by your customers | Customer is deployer; you are provider | Deepfake label (deployer); marking (you) | Already applies / 2 Dec 2026 | Give customers built-in EU icon and label options |
| Auto-published AI news, briefings or blog posts | Whoever publishes is the deployer | Label public-interest text unless editorially reviewed | Already applies | Name an editor and document the review |
| Long AI meeting or document summaries | Provider | Likely marking of text over 200 tokens | 2 Dec 2026 if launched before 2 Aug 2026 | Often missed because the output is "internal" |
| Emotion or sentiment detection on calls or video | Deployer (and you as provider) | Inform affected people | Already applies | Check Article 5 prohibitions first |
| CV screening or candidate ranking | Provider | Transparency, plus high-risk obligations | High-risk from 2 Dec 2027 | Recruitment is a listed Annex III area |
| Credit scoring in a fintech app | Provider | High-risk obligations | 2 Dec 2027 | Credit scoring is a listed Annex III area |
| Product recommendations, search ranking | Provider | Usually none under Article 50 | n/a | Still covered by consumer and data protection law |

*Annex III examples: [Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/). Text threshold: [Freshfields on the final Code](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx).*

If you are still deciding which AI features to build at all, start with [our practical guide to adding LLM features](/insights/blog/how-to-add-ai-to-your-product). Transparency is much cheaper to design in than to bolt on.

## High-risk systems: what the delay does and does not change

Moving Annex III obligations to 2 December 2027 gives teams in recruitment, education, credit scoring and access to essential services more than a year of extra time. Gibson Dunn calls it a [deferral rather than a dismantling](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/): the risk-based structure and core obligations remain. White & Case puts it plainly: the extension gives more time but [does not materially reduce the work required](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act).

What the work involves, in engineering terms:

- **Risk management** across the whole lifecycle.
- **Data governance** covering provenance, representativeness and bias testing. The agreed Omnibus lets all AI systems and GPAI models process special-category data for bias detection, under a [strict necessity standard](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/).
- **Logging** good enough to reconstruct what the system did and why.
- **Human oversight** built into the workflow, with real authority to override.
- **Accuracy, robustness and cybersecurity testing**, with results you can show.
- **Technical documentation and a quality management system**, simplified for SMEs and small mid-caps but not removed.

Two changes may take some products out of scope:

- **The narrower "safety component" definition.** AI built into a regulated product that only helps users or optimises performance is [not automatically high-risk](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act), unless its failure would create health or safety risks.
- **Less duplication in regulated sectors.** For medical devices and other sectors with their own rules, the Commission can now limit duplicate requirements through delegated acts.

> **The honest trade-off:** if your roadmap includes an Annex III feature such as automated candidate ranking, ask whether you need it before 2027 at all. A decision-support tool where a human makes the call, built with logging and oversight from the start, is often the better product and the easier one to certify. If you can't staff the governance work, don't ship the automated version.

Good evaluation practice carries straight over into high-risk evidence. Our guide to [evaluating LLM features in production](/insights/blog/evaluating-llm-features-in-production) covers the test sets and metrics that become your accuracy documentation.

## Engineering effort: what compliance looks like in hours and weeks

These are rough planning ranges for a small, senior team working on an existing product with one or two generative features. They are not benchmarks. Your stack, the number of surfaces and what your model vendor already provides will move them in either direction.

| Workstream | Typical effort | Who | Notes |
|---|---|---|---|
| AI feature inventory and role mapping | 2 to 5 days | Product lead, tech lead, counsel | The foundation for everything else; often turns up forgotten features |
| Disclosure UI for chat, voice and messaging | 1 to 3 days per surface | Frontend, design | Copy reviewed by counsel; translated for each EU language you support |
| Signed metadata for images and video | 1 to 3 weeks | Backend, DevOps | Signing service, key management, CDN and transcoding fixes |
| Watermarking via your model vendor | Days, if the vendor supports it | Backend | Confirm in writing what the vendor marks and how |
| Watermarking for self-hosted models | Several weeks or more | ML engineers | Text watermarking is still a research problem; consider a vendor that provides it |
| Detection endpoint or SDK | 1 to 2 weeks | Backend | Hosted in the EU or run locally; rate limiting; free access for authorities |
| Labelling tools for your customers | 1 to 2 weeks | Frontend, design | EU icons, placement options, audible disclaimers |
| Evidence pack and AI literacy records | 2 to 4 days, then ongoing | Tech lead, operations | Decisions, configurations, test results, training logs |

The biggest variable is self-hosting. If you run open-weight models yourself, all of the marking work is yours. For many teams the sensible answer is a hosted model whose provider documents its marking, revisiting self-hosting once the tooling matures.

## Step by step: getting ready before 2 December 2026

1. **Build the inventory.** List every feature that calls a model, including internal tools, background jobs, email generators and anything behind a feature flag. Record the model, the vendor, the output type and when EU users first had access.
2. **Assign a role to each feature.** Decide whether you are provider, deployer or both, and write down why. Take particular care with white-labelled or heavily modified third-party systems.
3. **Confirm your launch dates.** The 2 December grace period only covers systems placed on the market before 2 August 2026. Morgan Lewis recommends that providers relying on it [document the basis for doing so](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august). Pull release notes and deployment history now.
4. **Close the disclosure gaps first.** Interaction disclosure has applied since 2 August with no grace period. It is cheap to fix, so ship it this sprint.
5. **Get your vendor's marking position in writing.** Ask your model providers which outputs they mark, and with which techniques; whether metadata is signed; how detection works; and what their interoperability plan is for February 2027.
6. **Trace every output path.** Follow each generated file from the model response to what the user receives. Find where metadata is stripped or text is rewritten, and fix it.
7. **Decide on the Code.** Signing gives you a recognised benchmark. If you don't sign, document the equivalent measures you use instead.
8. **Ship tools for your customers.** Add EU icons, label placement and audible disclaimers so customers can meet their own obligations inside your product.
9. **Stand up detection.** Offer an endpoint or SDK, host it in the EU or make it runnable locally, and set a fair-use policy if you qualify for the fee carve-out.
10. **Write the evidence pack.** Keep the inventory, role decisions, configurations, test results, vendor confirmations and literacy measures in one place, with an owner and a review date.
11. **Add it to your release checklist.** New systems must comply from day one, so every new AI feature should pass a transparency check before launch.
12. **Plan the 2027 work.** If anything you ship could fall under Annex III, start the risk, data and logging work in the first half of 2027, not the second.

A focused team can often finish steps 1 to 5 within a week. Steps 6 to 9 depend on how many media types you generate. If you want an outside view of where you stand, an [independent software audit](/services/software-audit) covering AI features and data flows is a sensible first step.

![Hand tools and a wax seal stamp laid out in an orderly row](https://r78zyqbigtq5h1ka.public.blob.vercel-storage.com/posts/eu-ai-act-transparency-rules-saas-apps-chatbot-disclosure-content-marking-image-3-QrWl4RfiMQBdzluthz6ruBifEdgyoj.webp)
*A methodical checklist gets most teams ready within a few weeks.*

## Mistakes to avoid and when not to build

- **Assuming the model vendor covers everything.** The lab provides the model; you probably provide the system. Their marking helps, but your interface, pipeline and documentation are yours.
- **Relying on a visible badge alone.** Article 50(2) is about machine-readable marks. A visible "AI" sticker helps users but does not replace signed metadata or a watermark.
- **Building your own text watermarking.** Unless watermarking is your product, it is rarely worth the research effort for a small team.
- **Ignoring the nudifier ban if you offer open-ended image or video generation.** "We didn't intend it" is not enough when the output is foreseeable and safeguards are weak. Test your filters with adversarial prompts and keep the results.

Sometimes the right call is not to build. A marginal generative feature can drag in signing infrastructure, detection hosting and labelling tools. If it would, cut it or use a non-generative alternative.

## Who enforces it and what is at stake

[According to the Commission](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations), three bodies enforce Article 50:

- National market surveillance authorities.
- The AI Office, for systems it supervises.
- The European Data Protection Supervisor, where EU institutions are involved.

The Omnibus strengthened the AI Office ([Gibson Dunn](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/)). It alone now supervises AI systems where the same company builds both the general-purpose model and the system. The same goes for AI in very large online platforms and search engines. It can investigate, inspect on site and fine.

Penalties are set in [Article 99](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99). They scale with worldwide turnover, and the top tier is reserved for prohibited practices. The Omnibus keeps lower penalties for SMEs and small mid-caps. For most software companies, though, the bigger commercial risk is not a fine. It is an enterprise customer's procurement team asking for your marking arrangements and your evidence pack, and you having neither.

## How Authect approaches it

We treat AI transparency as part of the product, not paperwork added at the end. When we build or extend [generative AI features](/services/generative-ai), the inventory, role mapping, disclosure UI and marking pipeline go into the same backlog as the feature itself. A senior engineer leads the work. Delivery is AI-accelerated, and an engineer reviews every line, including the code that signs, watermarks and labels outputs.

You talk directly to the founders who will build your project, never a salesperson or account manager. Teams are small. You own the code, infrastructure and designs from day one, so your compliance evidence lives in your repositories, not ours. We work as a fixed-price project, in monthly sprints, or through ongoing support and evolution, depending on scope. Support and evolution suits the 2027 high-risk work well.

[See how we work](/about/how-we-work), get a second opinion through our [AI consulting](/services/ai-consulting) service, or [start a conversation online](/contact) with the people who would do the work.

## FAQs

### Does the EU AI Act apply to my company if we are based outside the EU?

Yes, if you place an AI system on the EU market or put it into service there. The provider definition depends on [placing a system on the EEA market](https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august) under your name, not on where you are incorporated. A UAE or UK company selling an AI feature to EU customers is in scope for that feature.

### We use OpenAI, Anthropic or Google models through an API. Isn't compliance their job?

Only partly. They provide the general-purpose model and carry the GPAI obligations, which have applied since August 2025. If you build an AI system on that model and sell it under your brand, you are likely its provider for Article 50 purposes. Use your vendor's marking where it exists, but confirm it in writing and make sure your pipeline keeps it intact.

### What exactly is due on 2 December 2026?

Machine-readable marking of synthetic outputs for generative systems placed on the EU market [before 2 August 2026](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act). Interaction disclosure and deployer duties are not delayed; they have applied since 2 August. Gibson Dunn's summary of the agreed text says the transition period for the new nudifier and CSAM ban also ends that day.

### Do short chatbot replies need watermarks?

Under the final Code of Practice, free-form text under 200 tokens [does not need to be watermarked](https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-33-the-final-code-of-practice-on-transparency-of-ai-generate-102n4yx). Longer outputs such as drafts, reports and summaries do. The chatbot itself must still tell users they are talking to AI unless that is obvious.

### Do we have to sign the Code of Practice?

No, it is voluntary. If you don't sign, the Commission expects you to show you meet the marking and labelling obligations [through alternative, equally adequate means](https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations). For small teams, building to the Code's specifications is usually the simplest way to show that.

### Our product does CV screening. When do we need to comply?

Recruitment is a listed Annex III high-risk area, with obligations from [2 December 2027](https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act). Transparency duties may already apply to parts of the product. The high-risk work is substantial, so start it early in 2027.

### Is a visible "Generated by AI" badge enough?

Not for providers. Article 50(2) requires machine-readable marking that allows detection. Under the Code, that usually means at least two layers for images, audio and video, such as signed metadata plus a watermark. Visible labels matter for deployers publishing deepfakes and public-interest text, and they help users, but they don't replace machine-readable marks.

## About Authect

- Authect is a founder-led software development studio: clients talk to the founders who build their project, never a salesperson.
- Every project is led by a senior engineer. Delivery is AI-accelerated, and an engineer reviews every line before it ships.
- Clients own their code, infrastructure and designs from day one, with no lock-in.
- Legal entity: AUTHECT - FZCO, Trade License No. 70505 (IFZA). Registered office: Building A1, Dubai Digital Park, Dubai Silicon Oasis, Dubai, United Arab Emirates.
- The team works in Dubai (Dubai Marina, Business Bay, Dubai Silicon Oasis) and in Spain (Madrid and Marbella), and works with partners in Lithuania.
- Clients across Europe, the Middle East and beyond; areas served include Dubai, Madrid, Marbella, Málaga, United Arab Emirates, Spain, Lithuania, Sweden, Norway, Denmark.
- Founders: Pranas Mickevicius (Chief Executive Officer), Oscar Arson Saiz (Chief Technology Officer).
- Cooperation models: fixed-price projects, monthly sprints, and support and evolution.
- Contact: info@authect.com, +971 50 246 5223, or https://authect.com/contact. A founder replies within one working day.
- Profiles: LinkedIn https://www.linkedin.com/company/authect/, Instagram https://www.instagram.com/useauthect/, X https://x.com/useauthect, TikTok https://www.tiktok.com/@useauthect, Facebook https://www.facebook.com/profile.php?id=61581141524137.
