Updated November 2026 for the Digital Omnibus on AI (Regulation (EU) 2026/1744). First published December 2025.
Your AI-powered app might already be illegal in the EU. That's not hyperbole.
Since February 2025, specific AI practices have been banned outright in Europe, and two more bans apply from 2 December 2026. Since August 2026, apps that let people talk to an AI or that generate content have had transparency duties. From 2 December 2027, apps in "high-risk" categories will need conformity assessments, quality management systems and CE marking. If you're building apps that use machine learning, recommendation engines, or models from OpenAI, Google or Anthropic, this affects you.
The EU AI Act is the world's first full legal framework for artificial intelligence. Think of it as GDPR for AI, but more technically complex. It applies to any provider placing AI systems on the EU market, regardless of where they're based. Building apps in Manchester for EU users? You're in scope.
We audited an AI-enabled communications platform for the healthcare and care home sectors and found three features that each needed checking against the high-risk rules: a Q&A feature, staff engagement analytics and planned AI content moderation. Its compliance readiness was roughly 15%, and the biggest gap was that it kept no logs of its AI operations at all.
From 2 December 2026, two new bans apply and the grace period for marking AI-generated content ends. Both come from the Digital Omnibus on AI, Regulation (EU) 2026/1744, which has been in force since 27 July 2026.
The Omnibus also moved the high-risk deadlines. Annex III systems (employment, education, credit scoring and the rest) now have until 2 December 2027, and AI in regulated products under Annex I until 2 August 2028. If your app generates images, audio, video or text, or lets people chat with an AI, the rules that matter most to you apply by 2 December 2026.

Article 5 of the Act establishes outright bans. These aren't "comply or pay a fine" situations. These features must be removed from production immediately.
The Act prohibits AI systems that use subliminal techniques or manipulative methods to distort user behaviour in ways that cause harm.
What does this mean practically? Those notification delivery algorithms optimised to target users when they're most vulnerable. Push notification timing that exploits sleep-deprived states to induce purchases. "Dark pattern" UX that AI has optimised to maximise engagement or retention by exploiting psychological weaknesses.
If your recommendation engine's objective function implicitly optimises for vulnerability exploitation, you have a problem.
AI-powered emotion detection is banned in two specific contexts: the workplace and educational institutions.
That "student engagement tracker" using facial analysis? Illegal. HR interviewing tools that analyse candidate "sentiment" or "nervousness" from video feeds? Also illegal.
This doesn't mean emotion recognition is banned everywhere. Driver safety monitoring and medical diagnosis (detecting pain in patients) remain legal. But if you're building a "Zoom for Education" competitor with attention monitoring, you need to rethink that feature.
Using AI to infer race, political opinions, religious beliefs, or sexual orientation from biometric data is prohibited.
Any system that generates a "trustworthiness score" by aggregating disparate data sources (credit, social media, behaviour patterns) leading to unfavourable treatment is banned. If your app architecture resembles China's social credit system, even unintentionally, you're on the wrong side of the law.
Systems assessing crime risk based solely on profiling are prohibited. This one primarily affects public sector applications, but the principle extends to any "risk assessment" based purely on demographic or behavioural profiles.
These two bans matter to any app with image, video or voice generation. For a provider, the ban applies where producing such material is the system's intended purpose, or where the system can foreseeably and reproducibly produce it without significant modification and has no reasonable safeguards to prevent it (Article 5(1a)). In practice, an image generator in your app needs working safeguards against this misuse, and a way to correct misuse once it's reported. For the intimate-image ban, edits that don't increase exposure or change the nature of what's shown aren't caught (Article 5(1b)).

Your app is high-risk if it uses AI for education evaluation, employment task allocation, credit scoring, critical infrastructure management, or biometric identification. Annex III of the Act lists these specific use cases, and regardless of your underlying technology's complexity, falling into these categories triggers conformity assessments, quality management systems and CE marking requirements from 2 December 2027.
You'll also need to register the system in the EU database. The deadline for all of this was 2 August 2026 until the Digital Omnibus moved it to 2 December 2027.
Education and Vocational Training
Systems that determine access to education, assign students to institutions, or evaluate learning outcomes are high-risk. That "AI Tutor" app adapting curriculum based on student performance? If its output formally evaluates outcomes or steers the learning process, it's probably high-risk.
Employment and Workers Management
Recruitment tools (CV parsing, candidate ranking) and systems for task allocation or performance monitoring fall here. This is the big one for gig economy apps. If your algorithm assigns rides, deliveries, or tasks to workers, you're operating a high-risk system because it determines "access to self-employment."
Credit and Insurance
AI-powered credit scoring (except for fraud detection) and life/health insurance risk assessment are high-risk. Your "Buy Now, Pay Later" app using AI to determine credit limits? High-risk.
Critical Infrastructure
AI components managing road traffic, water, gas, heating, or electricity supply. If you're building apps that optimise HVAC systems in commercial buildings for energy efficiency, think carefully. If a failure could disrupt heating supply, you're likely high-risk.
Biometrics
Remote biometric identification (even post-event) and biometric categorisation trigger high-risk status.
In that healthcare communications audit, the RAG-based Q&A feature could influence care decisions, so it may count as a medical device. If it needs a notified-body assessment under the MDR, it becomes high-risk under Annex I from 2 August 2028. The engagement analytics fall under Annex III point 4(b) if they're used to monitor and evaluate staff performance. The planned AI content moderation has no Annex III category of its own, so its risk depends on what decisions it feeds.
A system listed in Annex III isn't automatically high-risk if it doesn't pose a "significant risk of harm." The derogation applies when your AI:
There's a catch. If your system performs profiling of natural persons, it's always high-risk. No exceptions.
If you rely on the derogation, you have to document your assessment before launch and still register the system in the EU database (Article 6(4)). The Digital Omnibus shortened what that registration has to contain.
Strategic Implication: You can de-risk a product by narrowing its scope. Instead of "Automated Recruitment Screening Tool" (high-risk), position it as a "Keyword Highlighting Tool for Recruiters" (potentially lower risk). But this requires your "intended purpose" documentation to match what the system actually does.

For high-risk systems, compliance isn't a checkbox exercise. It's architectural.
Training datasets must be "relevant, sufficiently representative, and to the best extent possible, free of errors and complete."
Nobody has error-free datasets at scale. The regulators know this. What they're looking for is a rigorous process of error detection and correction. Automated data quality checks (using tools like Great Expectations) validating schema, distribution, and null values. The logs of these checks become your compliance evidence.
You must examine datasets for biases affecting health, safety, or fundamental rights. The Act allows processing of special category data (race, ethnicity) solely for bias detection and correction. Since the Digital Omnibus this sits in a new Article 4a, which also extends it beyond high-risk systems. It comes with conditions: other data, including synthetic or anonymised data, mustn't be able to do the job, the data must be pseudonymised and access-controlled and never passed to anyone else, and it has to be deleted once the bias is corrected or its retention period ends. This means secure "clean room" environments where demographic data joins training data temporarily to calculate fairness metrics before being discarded.
High-risk systems must automatically generate logs enabling traceability (Article 12). A practical template, built on the minimum Article 12(3) sets for remote biometric identification:
For high-traffic LLM applications, this generates massive storage volumes. And logging inputs might capture PII, conflicting with GDPR's data minimisation principle.
Practical Solutions:
When we audited a client's AI-enabled platform recently, logging was the most significant gap. They had basic Firebase Analytics for telemetry, but zero compliance-grade logging: no AI decision logs, no input data recording, no model version tracking, no human override logging. For their RAG system specifically, there was no record of which documents were retrieved, what confidence scores drove ranking decisions, or which prompt template versions were in use. Rebuilding this from scratch added significant development time to their compliance roadmap.
Standard security (firewalls, encryption) isn't enough. The Act requires protection against AI-specific vulnerabilities:
Data Poisoning: Attacks manipulating training data to introduce backdoors
Adversarial Attacks: Inputs designed to trick the model (adding noise to misclassify images)
Model Extraction: API abuse to reverse-engineer model weights
Prompt Injection: Manipulating LLM behaviour through malicious inputs
Article 15 expects high-risk systems to be as resilient as possible to errors and faults, and says that can include backup or fail-safe plans. In practice that means a fallback: if AI confidence drops below a threshold, the system hands over to heuristic rules or a person instead of guessing.
Most modern apps are wrappers around third-party APIs. OpenAI, Google Vertex AI, Azure AI, Anthropic's Claude. The EU AI Act makes this relationship complicated.
Deployer: You use a generic "Sentiment Analysis API" to sort customer feedback internally. Your obligations are limited (transparency and literacy).
Provider: You take a foundation model, fine-tune it on medical data, and release "Dr. AI" as a medical diagnosis assistant.
In the second scenario, you're the Provider of a High-Risk AI System. OpenAI or Anthropic isn't the provider of your specific medical system. You are. You're fully responsible for conformity assessment, data governance, and risk management of the entire system, including the base model you don't control.
We've seen this scenario play out with a client's white-label platform. They provided a generic AI communications tool that healthcare organisations could rebrand. Under Article 25, those deployers risked becoming providers simply by putting their own name on the system. The liability implications weren't in anyone's contracts. We now advise explicitly addressing this in terms of service.
If you become a Provider based on a third-party model, you face a "compliance gap." You need technical details about the base model (training data, validation results) to complete your conformity assessment. The API provider might consider this a trade secret.
Article 53 of the Act obliges GPAI providers to supply technical documentation enabling downstream providers to comply. Article 25(4) also requires the provider of a high-risk system and any third party supplying an AI system, model, tool or component for it to agree in writing the information, technical access and help the provider needs to comply. Free and open-source tools are exempt unless they're general-purpose AI models. The Digital Omnibus made explicit that this covers AI models, and breaching it can now be fined under Article 99(4). Your contracts with API providers should reference both Article 53 and Article 25(4). If a provider won't supply the compliance data you need, you may not be able to complete a conformity assessment, and without one you can't place a high-risk system on the market.

Compliance with the EU AI Act isn't something you implement once. It requires ongoing habits.
Define "Intended Purpose" narrowly. The legal definition of your product's purpose determines its risk class. Avoid marketing language like "AI that assesses potential" if the tool simply matches keywords. Precise language prevents regulatory headaches.
Maintain a Model Inventory. Know every place AI is used in your product, its risk classification, and its data sources. Treat this like an asset register.
Check for Substantial Modification. Before approving major model updates or prompt engineering changes that alter system behaviour, verify whether it triggers re-assessment of compliance.
Vendor Due Diligence. When buying AI tools, ask: "Is this CE marked? Where's your Declaration of Conformity?"
Treat Prompts as Code. Prompts define system behaviour. Version control them (Git), review them, test them. They're not throwaway strings.
"I am a Robot" by Default. Hardcode transparency disclosures. Article 50(1) requires users to be told they're interacting with an AI unless it's obvious, and Article 50(2) requires generated images, audio, video and text to be marked as AI-generated in a machine-readable way. Both have applied since 2 August 2026 (for marking, from 2 December 2026 if your app was already on the market). Recital 133 lists watermarks, metadata, cryptographic provenance and fingerprints as ways to mark content, and lets the marking sit in the model, so check whether the model you call already marks its output and that the marking is still there in what your app delivers.
Log the "Why." Don't just log outputs. Log configuration parameters (temperature, top_p) and prompt template versions. You need to explain why the system produced a specific output.
Red Team Your Own Features. Don't just test if it works. Test how it fails. Try to make your chatbot produce harmful outputs. Try to bias your classifier. Document these tests. They're your "risk management" evidence.
Use Standard Libraries. Use open-source compliance tools: C2PA libraries for content provenance metadata, a watermarking library if that metadata may be stripped, and fairness toolkits (e.g. Fairlearn) for bias detection. Don't reinvent these wheels.

The Act applies in phases, and the Digital Omnibus moved some of them. These are the dates as they stand now:
When we first wrote this guide in December 2025, the Digital Omnibus was a Commission proposal and the high-risk dates might have slipped. It became law in July 2026 and they did: Annex III moved from 2 August 2026 to 2 December 2027, and Annex I from 2 August 2027 to 2 August 2028. It also eased the rules for smaller companies: SMEs can now meet parts of the quality management system requirement in a simplified way (Article 63), and small mid-cap companies join SMEs in having most fines capped at the lower of the two limits (Article 99).
Conformity work still takes months, and the bans and transparency duties that apply to most apps didn't move, so the extra time is best spent on testing.
Here's the thing. The EU AI Act isn't just regulatory burden. It's becoming a quality standard.
Just as GDPR became the de facto global privacy benchmark, "EU AI Act compliant" is emerging as a trust signal. Customers in the US, Asia, and elsewhere increasingly expect the same standards. Building for EU compliance isn't a regional cost. It's a global quality investment.
Companies that implement rigorous risk management, bias testing, and transparency now will be audit-ready while competitors scramble. The "Compliance-as-a-Service" middleware market is emerging (automated bias checking, secure logging vaults, hallucination firewalls). Early adopters will have smoother paths to market.
The regulatory environment rewards organisations that treat "Trustworthy AI" as a feature, not a constraint.
If you're building apps with AI capabilities for EU users, start here:
The high-risk deadline is now 2 December 2027. A year sounds comfortable until you factor in development cycles, testing, and conformity assessments.
The healthcare platform we audited was at roughly 15% readiness. Logging infrastructure, documentation and conformity assessment preparation take longer than teams expect.
The EU AI Act represents a significant shift in how AI-powered apps must be designed, documented, and operated. If you're unsure how your current or planned mobile app fits within these regulations, or you need help architecting compliant systems from the ground up, we're here to help.
At Foresight Mobile, we build AI-integrated mobile applications with compliance considered from day one. Whether you're concerned about risk classification, logging requirements, marking AI-generated content, or third-party API dependencies, get in touch to discuss your specific situation.
We offer a structured discovery process through our App Gameplan service. In four weeks, you'll know exactly what to build, what compliance considerations apply, and how to move forward with confidence.
Yes, if the app uses AI and is available in the EU, or its AI output is used there. Article 2 covers providers that place AI systems on the EU market wherever they're based, and providers and deployers outside the EU whose AI output is used in the Union. A UK company's AI-powered app in the EU App Store or Google Play is in scope.
Mainly the high-risk rules. Annex III systems now have until 2 December 2027 and Annex I systems until 2 August 2028. The bans, the general-purpose AI rules and the Article 50 transparency duties kept their dates, apart from a four-month transition to 2 December 2026 for marking by generative systems already on the market. The Omnibus also added two new bans that apply from 2 December 2026.
If you're the provider of a system that generates images, audio, video or text, yes. Article 50(2) requires the output to be marked in a machine-readable way so it can be detected as AI-generated. It doesn't apply to standard editing tools or to output that doesn't substantially change the input. Separately, a deployer that publishes deepfake images, audio or video has to disclose that they're artificial (Article 50(4)).
Using a banned practice can cost up to €35 million or 7% of worldwide annual turnover, whichever is higher (Article 99(3)). Breaching the transparency duties or most high-risk obligations can cost up to €15 million or 3% (Article 99(4)). For SMEs and start-ups the cap is whichever figure is lower, and since the Omnibus small mid-cap companies get the same treatment for everything except the bans.
Only once its design changes significantly. Under Article 111(2), a high-risk system already on the market before the high-risk rules apply is covered from the point it's significantly changed. The Omnibus clarified (recital 39) that further units of the same type and model share that grace period, as long as the design stays the same. Systems intended for public authorities have to comply by 2 August 2030 regardless. Whether a model or feature update counts as a significant change is worth deciding before you ship it.
Sources: