The Synthos team's monthly reading
of the Regulation in practice.

What you will find here

The AI Act is not applied by copying and pasting the Regulation. It is applied through the living law that surrounds it: the AI Office guidelines, data protection authority decisions, national judgments, the positions of competent national authorities, the Digital Omnibus under way.

The Brief holds these pieces together, one per issue, so that European SMEs have something they can act on.

Cadence

Monthly, with acceleration around individual events: publication in the OJEU of the Digital Omnibus, a data protection authority penalty in the AI field, a national judgment of significance.

The archive stays open and free.

Subscription

Leave your address below to receive the next issue. A single email per month. The subscription can be revoked with one click, always.

AI Office guidelines on Article 5:
what changes for European SMEs.

What the guidelines are and are not

The European Commission guidelines published on 4 February 2025 are soft law. They do not have the rank of a Regulation and cannot amend the text of Article 5. Their function is interpretative.

The difference from an implementing act is substantial. An implementing act is binding and has the force of EU secondary law; a guideline is an orientation from the executive that clarifies how the text is to be applied. In Union law, Recital 16 of the Regulation explicitly authorises the Commission to issue operational guidance on prohibited practices.

The practical point is this. When a national supervisory authority opens an investigation into an AI system, it will use these guidelines as an interpretive grid. When a national court (the labour court for HR cases, the administrative court for public-administration decisions, the ordinary court for civil disputes over credit-scoring systems) has to rule on the application of an AI Act penalty, it will cite these guidelines as an auxiliary source.

The Italian Council of State precedent (section VI, judgment 8472/2019) on algorithmic transparency in public-administration procedures shows that an EU national administrative court is already prepared to make interpretive use of non-binding EU documents (Italian case law, illustrative). The AI Office guidelines will enter the same argumentative perimeter.

The Commission has reserved the power of periodic revision. A new version of the guidelines does not require a legislative step. When it is issued, the applicable framework will update automatically.

Cognitive manipulation (Article 5(a)): what is clarified

The literal text of Article 5(1)(a) prohibits 'the placing on the market, the putting into service or the use of an AI system that deploys subliminal techniques beyond a person's consciousness or purposefully manipulative or deceptive techniques, with the objective, or the effect of materially distorting the behaviour of a person', causing them to take a decision they would not otherwise have taken and that causes significant harm.

The guidelines restate the test as three cumulative criteria: the presence of a subliminal, manipulative or deceptive technique; material distortion of behaviour; significant harm, whether actual or reasonably foreseeable. All three must occur together. If one is missing, the prohibition does not apply.

The Commission lists concrete prohibited cases: subliminal visual or auditory messages, distracting attention to prevent a person from noticing relevant content, manipulating the perception of time to induce dependency, mood induction through background sounds and images, and a deceptive chatbot that impersonates a human and causes fraud.

The pattern that protects SMEs is the qualification of 'lawful persuasion'. Standard targeted advertising, marketing personalisation under GDPR Article 6, and emotion analysis to improve customer support with adequate disclosure all remain permitted.

Where the bar lowers for SMEs is in gambling applications, aggressive fintech and dark-pattern UX. A betting platform that uses AI to recognise signs of incipient addiction and to propose intensifying play falls within the prohibition; an ordinary recommendation of financial products, by contrast, is lawful persuasion.

The 'significant harm' criterion is what distinguishes most commercial systems from prohibited practices. The materiality of the harm must be documented in an audit.

Vulnerability (Article 5(b)): who is covered

Article 5(1)(b) prohibits 'an AI system that exploits any of the vulnerabilities of a natural person or a specific group of persons due to their age, disability or a specific social or economic situation', with the effect of distorting behaviour and causing significant harm.

The guidelines give an operational definition of vulnerability in three categories. Age, which covers minors and the elderly. Disability, recognised under EU law. A demonstrated social or economic situation, which includes documented indebtedness, conditions of social marginalisation, and dependence on welfare services.

The Italian Data Protection Authority (Garante) precedent on Replika (limitation of 2 February 2023) is the Italian reference case (illustrative). The chatbot was blocked because it was addressed, among others, to minors and to fragile people without adequate safeguards. The Commission guideline confirms and generalises that interpretation two years later.

For SMEs the operational risk cases are three. Apps aimed at minors (gaming, social, edutainment): it is mandatory to demonstrate in design records that the AI does not exploit cognitive immaturity. High-margin financial products aimed at weaker categories: the Mediaworld precedent (Garante, injunction order of 4 July 2024, EUR 800,000 penalty) for credit scoring without an adequate impact assessment marks the standard (illustrative). Healthcare systems for the elderly with personalisation of treatments or value-added commercial offers: the combination of health vulnerability and economic exploitation is the riskiest scenario.

The operational audit that SMEs must carry out is documentary. It must be demonstrated, in writing and before placing on the market, that the system is not designed to exploit the covered vulnerabilities. Documentation prepared in advance is the most relevant mitigating factor in supervision.

Emotion recognition at work (Article 5(f)): the tightened reading

Article 5(1)(f) prohibits 'AI systems to infer emotions of a natural person in the areas of workplace and education institutions, except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons'.

The guidelines confirm the restrictive interpretation of the prohibition. Inferring emotions means deducing emotional states (joy, sadness, anger, anxiety, stress) from facial, vocal or physiological biometric data. The workplace covers all working contexts, including internal corporate training.

The Commission gives three typical prohibited cases. Software that assesses tone of voice in call centres to measure operators' engagement. Analysis of micro-expressions in HR videoconferencing during selection interviews. Systems that infer the sentiment of employees from internal company chats for performance-review purposes.

The 'medical or safety reasons' exception is narrow. The guidelines limit it to cases in which stress detection is explicitly connected to a safety risk for the worker or for third parties: an operator of dangerous machinery, a worker in chemical plants, a driver of transport vehicles. A general need for organisational well-being is not enough.

Coordination with national legislation raises the bar further. In Italy, the Workers' Statute, Article 4 (Law 300/1970) requires a collective agreement or authorisation from the labour inspectorate for remote-monitoring systems; comparable rules apply in other Member States, so check your national measures. GDPR Article 9 classifies biometric data as a special category. The Foodinho/Glovo precedent (Garante, injunction order of 5 July 2021, EUR 2.6 million penalty) consolidated the standard of algorithmic transparency in the workplace (Italian DPA, illustrative).

An SME that runs call centres, HR tech or training platforms must suspend any emotion-detection functionality before 2 December 2026 (the date of application of the related Article 50 obligations) and refer the system to an audit documented by the partner lawyer.

What to do now, in four steps

The AI Office guidelines have been in force since 4 February 2025. The Article 5 prohibitions are directly applicable from 2 February 2025. SMEs that want to reduce penalty risk over the next three months have four operational steps.

First. Audit the systems at risk under Article 5(a), (b) and (f). The audit is taxonomic: list every AI system in use or in development, check whether it falls within one of the four covered situations (cognitive manipulation, exploitation of vulnerability, emotion recognition in work or education, facial scraping), and record the outcome. The AI Act Check pre-screening covers this phase.

Second. Document the design choices that exclude the triggers. For each system identified, produce an internal document (1-3 pages) explaining how the system's design prevents the triggers of the prohibition. Documentation prepared before any challenge is the most relevant mitigating factor recognised by the European Commission and by data protection authorities. It must be dated, signed by the technical lead and archived.

Third. Coordinate with the partner lawyer on the grey-zone cases. Systems that fall into a borderline area (sentiment analysis for customer support, personalised recommendations for demographically sensitive categories, monitoring of concentration in corporate training) require a precise legal assessment. Synthos Logic refers these to its partner law firms. Preventive advice costs less than even a minor penalty.

Fourth. Prepare the file for the supervisory authority. The competent national authority for AI Act supervision differs per Member State (in Italy, the National Cybersecurity Agency (ACN) under Law No. 132 of 23 September 2025 — check your national designation). SMEs operating in exposed sectors (HR tech, fintech aimed at sensitive categories, edutainment for minors, digital health, gambling) must pre-assemble the file: description of the system, the Article 5 audit, design documentation, any FRIA under Article 27, and the partner lawyer's contact. The file is prepared once and updated monthly.

Digital Omnibus, the 2027-2028 deadlines:
what really changes for the European SME, and what does not.

What happened, exactly

On 7 May 2026 the Council of the European Union reached a political agreement on the package of amendments to Regulation (EU) 2024/1689 known as the 'Digital Omnibus'. A political agreement is not a publication in the OJEU: it is the moment when Ministers express the Council's position. From here come the trilogues with the European Parliament, formal adoption by both co-legislators, and publication in the Official Journal of the European Union. Only at that point do the dates become binding. The current state is therefore: a politically agreed text, a provisional calendar of application, and a window for possible (but unlikely) technical revision in trilogue.

The three operational changes

First change: postponement of the substantive high-risk deadlines. Annex III of the Regulation (eight categories: biometrics, critical infrastructure, education, employment/HR, essential services, law enforcement, migration, justice) was due to take effect on 2 August 2026. It moves to 2 December 2027. Annex I (AI systems as safety components of regulated products — MDR, the Machinery Directive, lifts, vehicles) was due to take effect on 2 August 2027. It moves to 2 August 2028. Second change: a new prohibition, in force from 2 December 2026, the prohibition of 'nudifier apps', AI systems designed to generate non-consensual sexually explicit images of real people. It is added to the Article 5 list without amending the existing one. Third change: the watermarking obligation under Article 50(2) and the deep-fake labelling obligation under Article 50(4) apply from 2 December 2026 also to systems already placed on the market before then, so there is no 'grandfathering' of pre-existing generative systems.

What changes for a European SME

Three readings, one per type of SME. (a) An SME providing HR systems, credit scoring, diagnostic healthcare systems, or educational systems (Annex III): nine more months to close out the quality management system under Article 17, technical documentation under Article 11, conformity assessment under Article 43, CE marking under Article 48, EU database registration under Article 49(1). Time is not gained by standing still; it is gained to do well what should have been done in July. (b) A manufacturing, medtech or automotive SME that integrates AI as a safety component of regulated products (Annex I): one more year to coordinate the dual compliance of the AI Act + product legislation (MDR Regulation 2017/745, Regulation (EU) 2023/1230 which will replace the Machinery Directive from 14 January 2027). The calendar is realigned: machinery 14 January 2027, AI Act Annex I 2 August 2028. A single well-structured conformity assessment covers both. (c) An SME that already places generative AI systems on the market (chatbots, text or image generators): the transparency obligation under Article 50(1), watermarking under paragraph 2, and deep-fake labelling under paragraph 4 take effect on 2 December 2026. There is no grandfathering for pre-existing systems. Seven clear months to update the product.

What does NOT change

The Digital Omnibus does not amend Article 5 (prohibited practices): everything already prohibited since 2 February 2025 stays prohibited. The body of enforcement built by the Italian Data Protection Authority (Garante) on Clearview AI (decision No. 50 of 10 February 2022, EUR 20 million penalty), Replika (limitation of 2 February 2023), ChatGPT (limitation of 30 March 2023 and injunction order of 2 November 2024, EUR 15 million), and Foodinho/Glovo (injunction order of 5 July 2021, EUR 2.6 million) remains binding within its jurisdiction and is an illustrative reference for the rest of the EU. Article 4 (AI literacy) remains applicable from the outset to providers and deployers. The calendar for GPAI (Chapter V) stays at 2 August 2025. The national authorities designated by each Member State remain in full mandate (in Italy, under Law No. 132 of 23 September 2025: AgID for notification, ACN for supervision, the Garante for personal data — your national designations will differ).

The Synthos reading

The postponement is political, not technical: no one has changed their mind about what the Regulation requires; it was recognised that the European system was not ready to enforce deadlines this close. For the European SME the concrete consequence is not 'relax': it is 'use the extra months well'. The main risk we observe in practice is the illusion of time. The FRIA under Article 27 for a credit-scoring system requires three to four weeks of serious work; the quality management system under Article 17 is not written in a weekend; the technical documentation under Article 11 is built in parallel with development, not afterwards. Nine months is a horizon of advisory work, not a holiday. For those who have chosen the Article 6(3) derogation, we recall that the provider's documentation (Article 6(4)) must be produced before placing on the market, not after, and now applies to the new calendar. For those who develop GPAI, we recall that Article 53 is already active for new models from 2 August 2025; the Digital Omnibus does not touch it.