Back to the comparison
DownloadOpenAI “deep research”

EU AI Act: Real Obligations for a Commercial Software Company Selling into the EU

Research date: 31 August 2026. The operative statute is Regulation (EU) 2024/1689 (Artificial Intelligence Act), not “Regulation (EU) 2021/XXXX.” The “2021” reference comes from the Commission’s original legislative proposal. The AI Act was adopted on 13 June 2024, published on 12 July 2024, and entered into force on 1 August 2024. It has since been materially amended by Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted on 8 July 2026, published on 24 July 2026, and in force from 27 July 2026. This report therefore uses the EUR-Lex consolidated text dated 27 July 2026, read together with the authentic amending act. [1][2][3]

Executive summary

For a normal commercial software company, the AI Act is not a single certification requirement applying to every AI feature. The legal outcome depends first on what the company is doing in relation to each feature: provider, deployer, importer, distributor, product manufacturer, downstream integrator, or provider of a general-purpose AI model. It then depends on the system's risk classification and functionality. A SaaS company that builds an AI feature and sells it under its own name will ordinarily be the provider of that AI system; the same company is separately a deployer when it uses AI internally, for example for hiring, employee management, customer support, fraud decisions, or sales. [4][5]

The most important practical conclusions as of 31 August 2026 are:

Takeaway What it means for a software vendor now
Some obligations are already live for essentially all relevant providers/deployers. AI-literacy measures under Article 4 have applied since 2 February 2025; most Article 5 prohibited practices have applied since the same date; and Article 50 transparency rules became applicable on 2 August 2026. [6][7][8]
The major high-risk compliance package is not yet generally due. Following the July 2026 Omnibus, Chapter III Sections 1–3 apply to Annex III high-risk systems from 2 December 2027, and to Article 6(1)/Annex I product-related systems from 2 August 2028. The Commission expressly attributed the postponement in part to delayed standards. [7][9][10]
Generative-AI transparency is a live obligation now. Providers of systems generating synthetic audio, images, video or text generally need machine-readable/detectable marking under Article 50(2); providers of conversational systems have a disclosure obligation under Article 50(1); deployers publishing deepfakes or certain public-interest AI text have their own Article 50(4) disclosure duty. Legacy synthetic-content systems placed on the market before 2 August 2026 get a narrow Article 50(2) transition until 2 December 2026. [8][11][12][13]
Using an LLM does not automatically make the software vendor a GPAI-model provider. Article 53 applies to a provider of the general-purpose AI model. A company integrating a third-party model into its application will normally instead be the provider of the resulting AI system/downstream provider, unless it develops the model itself or makes a modification significant enough to assume GPAI-provider responsibilities. The Commission has issued specific GPAI-provider guidelines on this issue. [4][14]
GPAI-model rules are already applicable. Articles 51–55 have applied since 2 August 2025 to newly placed GPAI models. Providers of GPAI models placed on the market before that date have until 2 August 2027. [15][16][13][7]
SME status is mostly relief, not exemption. SMEs remain subject to substantive Article 4, 5, 50, GPAI, and high-risk duties when those provisions apply. They receive proportionality/support measures, possible simplified quality-management treatment, reduced conformity-assessment fees, and lower fine ceilings. [17][18][19]
Much of the law is settled; the compliance machinery is not. The duties themselves are now legislation, but harmonised high-risk standards remain under development; final Article 6 classification guidelines are still pending; Commission guidance/templates are still due in several areas; and multiple delegated-act powers can change classifications, thresholds, documentation and conformity details. [20][10][21][22]

The most common mistake would be to defer all AI Act work until the high-risk dates. That is wrong. Article 4, most of Article 5, Article 50, and—if the company itself supplies GPAI models—Chapter V are already operative. [7][13]

The other common mistake would be the opposite: assuming every AI-enabled SaaS feature must already meet Articles 8–15, obtain CE marking, and run a notified-body assessment. Those measures concern high-risk systems, and the principal Chapter III high-risk duties now have the later 2027/2028 dates. Most ordinary productivity, drafting, summarisation, recommendation, support and generative-software features will not become high-risk merely because they use AI; intended purpose and the specific Article 6/Annex III categories control. [23][20][7]

The financial exposure is nevertheless material. Article 99 provides a headline maximum of €35 million or 7% of worldwide annual turnover for Article 5 violations and €15 million or 3% for specified operator obligations, including Articles 16, 22–26 and 50; supplying incorrect, incomplete or misleading information to authorities can reach €7.5 million or 1%. For SMEs the applicable cap is the lower rather than the higher of the fixed amount and percentage; the 2026 Omnibus also added specified relief for small mid-cap companies. [19][24]

Scope, roles, and risk classification

The AI Act has deliberately broad territorial reach. Article 2 covers providers placing AI systems or GPAI models on the EU market or putting AI systems into service in the EU regardless of where the provider is established; deployers located in the EU; and providers or deployers outside the EU where the output produced by the system is used in the EU. Importers, distributors, certain product manufacturers and affected persons in the EU also appear expressly within the Regulation's scope. [5]

How to determine your role

For each distinct AI feature or model, role classification should be done separately.

Provider. Article 3(3) defines a provider as the person or entity that develops an AI system or GPAI model, or has one developed, and places it on the market or puts the AI system into service under its own name or trademark, whether paid or free. A commercial SaaS company developing a feature and selling that feature under its brand will therefore normally be a provider. [4]

Deployer. Article 3(4) covers an entity using an AI system under its authority, except purely personal/non-professional use. Your customers will frequently be deployers of your product; your own company is a deployer when it uses AI internally. [4]

Importer and distributor. An EU entity placing a third-country supplier's branded system on the market may be an importer; another supply-chain entity making an AI system available without being provider/importer may be a distributor. These distinctions matter substantially for high-risk systems because Articles 23 and 24 impose independent verification, corrective-action and cooperation duties. [4][25]

Authorized representative. This is an EU-established person with a written mandate from a provider. Importantly, there is no general rule forcing every non-EU vendor of ordinary non-high-risk SaaS to appoint an AI Act representative. The explicit representative duties relevant here arise principally for third-country providers of high-risk systems under Article 22 and of GPAI models under Article 54, subject to the statutory exceptions. [26][27]

Product manufacturer. For specified high-risk AI safety components in regulated products, Article 25 can make the product manufacturer the AI-system provider, particularly where the system is marketed or put into service under the manufacturer's name. [28]

Downstream/GPAI distinction. Integrating an externally supplied GPAI model into an application does not by itself turn the application vendor into the original model's GPAI provider. Article 53 actually requires the GPAI provider to supply downstream AI-system providers with documentation sufficient to understand the model's capabilities and limitations and comply with the Act. The Commission's GPAI guidelines further address when modifications are sufficiently significant to transfer GPAI-provider obligations. [29][14]

Risk triage

A commercial vendor should classify each product through the following sequence.

First, determine whether the software actually falls within the Article 3(1) definition of an AI system: a machine-based system operating with varying levels of autonomy that may exhibit adaptiveness and infers from inputs how to generate outputs such as predictions, content, recommendations or decisions affecting physical or virtual environments. The Commission has issued non-binding AI-system-definition guidelines, but the Regulation itself is controlling. [4][30]

Second, test Article 5. A prohibited use cannot be “made compliant” through certification. Most Article 5 bans are already effective. [31][7]

Third, test Article 6. An AI system is high-risk under Article 6(1) when it is a qualifying safety component/product governed by listed harmonisation legislation and the relevant product is required to undergo third-party conformity assessment; the 2026 amendment narrowed what counts as a safety component by excluding systems used solely for non-safety assistance, optimisation, efficiency, automation, convenience or quality control, while preserving coverage where failure or malfunction would endanger health or safety. [23]

Separately, Article 6(2) makes the use cases in Annex III high-risk. They cover defined applications in biometrics, critical infrastructure, education and vocational training, employment and worker management, access to essential public/private services, law enforcement, migration/asylum/border control, and administration of justice and democratic processes. [23][32]

Article 6(3) contains an important but narrow Annex III carve-out where the system does not present a significant health, safety or fundamental-rights risk and satisfies one of the specified conditions—for example, carrying out a narrow procedural or preparatory task. Systems performing profiling of natural persons remain high-risk. A provider relying on the carve-out must document the assessment and, once the provision is operative for the system, register itself and the system under Article 49(2). [20][8]

That carve-out is currently one of the Act's largest unresolved interpretive areas. Article 6(5) required Commission practical guidelines by 2 February 2026, but the Commission instead issued draft guidance for consultation in May 2026; the consultation closed on 23 July 2026, and the Commission says final guidelines will be adopted by the end of 2026. [20][9][33]

Open-source status is also not a blanket exemption. The AI Act contains an exemption for qualifying free/open-source AI systems, but it does not eliminate the Article 5 and Article 50 rules or the high-risk regime. GPAI models have a separate, narrower open-source treatment: Article 53(2) can remove the technical-documentation/downstream-information duties in Article 53(1)(a) and (b) for qualifying open-source models, but not the copyright-policy or public training-summary obligations, and the exception does not extend to systemic-risk GPAI models. [34][15]

Obligation map

The following is the core compliance map for an unspecified commercial software company. P = AI-system provider; D = deployer; G = GPAI-model provider; G-SR = provider of GPAI model with systemic risk; AR = authorized representative; I = importer; Dist = distributor; PM = product manufacturer; TP = third-party supplier into a high-risk system.

Obligation Article(s) Roles What the company must actually do Applicable date / deadline Legal status
AI literacy Art. 4; Art. 113(a) [6][7] P, D Take measures supporting development of sufficient AI literacy among staff and other persons operating/using AI on its behalf, taking account of their knowledge, training and use context. The July 2026 amendment expressly avoids requiring a guarantee of a particular literacy level. Since 2 Feb 2025; amended wording effective with 2026/1744 from 27 Jul 2026. Settled/current. Method is deliberately flexible; no mandatory certification course.
Avoid prohibited AI practices Art. 5(1)(a),(b),(c)–(h); Art. 113(a) [31][7] Any actor placing, putting into service or using the relevant system; commonly P/D/I/Dist Do not supply/use systems engaging in prohibited manipulation/exploitation, unlawful social scoring, specified individual criminal-risk prediction, untargeted facial-image scraping, workplace/education emotion inference except permitted cases, prohibited biometric categorisation, or prohibited real-time remote biometric identification. Since 2 Feb 2025. Settled/current. Commission prohibited-practices guidance exists but is non-binding.
New sexual/intimate-content and child-abuse prohibitions Art. 5(1)(ba), (bb), Art. 5(1a)–(1b); Art. 113(a) [31][7] Relevant suppliers/users Do not engage in the newly prohibited conduct involving qualifying non-consensual realistic sexual/intimate AI depictions of identifiable persons or systems designed/adapted for prohibited child-sexual-abuse material, subject to the article's precise conditions. 2 Dec 2026. Settled future law, but older Article 5 guidance predates these 2026 additions.
Tell people they are interacting with AI Art. 50(1) [8][11] P Design directly interactive AI so people are informed they are interacting with AI unless this is obvious to a reasonably well-informed, observant and circumspect person in context. Since 2 Aug 2026. Settled/current; guidance available.
Mark synthetic AI output Art. 50(2); Art. 111(4) [11][13] P Systems generating synthetic audio, image, video or text must mark outputs in a machine-readable format and make them detectable as artificially generated/manipulated, using effective/interoperable/robust/reliable technical solutions as technically feasible. New systems: 2 Aug 2026. Systems placed before that date: Article 50(2) compliance by 2 Dec 2026. Settled/current; technical implementation evolving. Final Code of Practice is an accepted voluntary compliance route.
Emotion-recognition / biometric-categorisation notice Art. 50(3) [12] D Inform exposed natural persons when using the covered emotion-recognition or biometric-categorisation systems; personal-data rules remain separately applicable. Since 2 Aug 2026. Settled/current.
Deepfake and public-interest AI-content disclosure Art. 50(4) [12] D Disclose that qualifying deepfake image/audio/video has been artificially generated/manipulated; disclose specified AI-generated/manipulated text published to inform the public on matters of public interest, subject to statutory exceptions including certain editorial/human-review situations. Since 2 Aug 2026. Settled/current; guidance/code available.
Determine systemic-risk GPAI status and notify Commission Arts. 51–52 [35] G A GPAI model is presumed high-impact when cumulative training computation is greater than 10²⁵ FLOPs; when the threshold is or will be met, notify the Commission without delay and in any event within two weeks, unless the statutory procedure justifies otherwise. GPAI regime since 2 Aug 2025; legacy model transition below. Binding, but scope dynamic: Commission can change thresholds/benchmarks by delegated act.
GPAI technical documentation Art. 53(1)(a), Annex XI [29][36] G Create and keep up-to-date model documentation covering training/testing/evaluation and Annex XI information; provide it to AI Office/national authorities on request. New GPAI models since 2 Aug 2025; pre-2 Aug 2025 models by 2 Aug 2027. Settled; methodologies/annex can evolve by delegated act.
Information for downstream AI-system providers Art. 53(1)(b), Annex XII [29][15] G Give downstream providers sufficient current information about model capabilities and limitations so they can meet their own AI Act obligations. Same GPAI dates. Settled; guidance and GPAI Code available.
GPAI copyright policy Art. 53(1)(c) [29][37] G, including qualifying open-source providers Maintain a policy to comply with EU copyright/related-rights law, specifically including recognition of Article 4(3) DSM Directive text/data-mining rights reservations using state-of-the-art technologies. Same GPAI dates. Settled; voluntary GPAI Code gives an accepted implementation route.
Publish GPAI training-content summary Art. 53(1)(d) [29][38][39] G, including qualifying open-source providers Publish a sufficiently detailed summary of content used to train the model using the AI Office template. Same GPAI dates. Settled/current; Commission template has been issued.
EU representative for third-country GPAI provider Art. 54 [27] Non-EU G, AR Before placing a GPAI model on the EU market, appoint a written-mandate EU representative; representative verifies documentation/compliance, keeps documentation for 10 years and cooperates with the AI Office. Qualifying open-source non-systemic models have an exemption. Same GPAI dates; appointment precedes placing on market once obligation applies. Settled/current.
Systemic-risk GPAI safety obligations Art. 55 [16] G-SR Conduct state-of-the-art model evaluations and adversarial testing; assess/mitigate Union-level systemic risks; track/document/report serious incidents without undue delay; maintain adequate model/infrastructure cybersecurity. Since 2 Aug 2025 for covered new models; legacy GPAI by 2 Aug 2027. Binding; implementation strongly state-of-the-art dependent. GPAI Code currently offers accepted compliance methodology pending harmonised standards.
Annex III high-risk classification and carve-out documentation Art. 6(2)–(4), Art. 49(2) [23][20][8] P Determine whether Annex III use case is high-risk. If relying on Art. 6(3)'s “not significant risk” exception, document the assessment before market/service and register the system when the regime applies. Profiling systems cannot use the carve-out. 2 Dec 2027 for Annex III systems. Future binding; important guidance pending. Final Article 6 guidelines promised by end 2026; Articles 6–7 contain delegated-act powers.
Product-related high-risk classification Art. 6(1), (1a)–(1c), Annex I [23][7] P, PM Determine whether AI is a regulated safety component/product subject to third-party conformity assessment; apply the post-Omnibus safety-component limitations. 2 Aug 2028. Future binding; sectoral scope remains dynamic. Art. 2(13) delegated relief is due by Aug 2027 for certain equivalent sectoral requirements.
High-risk risk-management system Arts. 8–9 [40] P Establish a continuous, iterative lifecycle risk-management system: identify/analyse known and reasonably foreseeable risks, evaluate risks under intended use/misuse, test controls and manage residual risk. Annex III: 2 Dec 2027; Art. 6(1) product systems: 2 Aug 2028. Future binding; standards pending.
High-risk data and data governance Art. 10; where relevant Art. 4a [41][6] P Training/validation/testing datasets must satisfy specified governance and quality controls, including relevance, representativeness and examination for bias where applicable. Any use of special-category data for bias work must fit the Act's narrow safeguards and applicable data-protection law. Same high-risk dates; Art. 4a itself is current after the 2026 amendment. Core law settled; GDPR interaction remains fact-specific.
High-risk technical documentation Art. 11, Annex IV [41][42] P Prepare technical documentation before market/service sufficient for authorities/conformity assessors to evaluate compliance; keep it current. Same high-risk dates. Future binding. Annex IV can be updated by delegated act; simplified smaller-company documentation machinery remains an implementation area to monitor.
Automatic logging Art. 12 [43] P Design high-risk system with automatic logging capabilities adequate to trace operation, relevant risks, post-market monitoring and substantial modification. Same high-risk dates. Future binding; standards pending.
High-risk transparency and instructions Art. 13 [44] P Make the system sufficiently transparent for deployers to interpret/use outputs appropriately and supply instructions covering capabilities, limitations, accuracy, foreseeable circumstances, oversight, logging and other required information. Same high-risk dates. Future binding.
Human oversight by design Art. 14 [45] P Design high-risk systems so natural persons can effectively oversee them, understand limitations, detect anomalies, disregard/override/reverse outputs where appropriate and intervene or stop safely. Same high-risk dates. Future binding; method depends on system/context.
Accuracy, robustness and cybersecurity Art. 15 [46] P Achieve appropriate accuracy, robustness and cybersecurity throughout lifecycle, disclose accuracy metrics and protect against AI-specific attacks such as poisoning, adversarial examples and other vulnerabilities where appropriate. Same high-risk dates. Future binding; heavily standards/state-of-the-art dependent.
High-risk provider quality-management system Arts. 16(c), 17; SME Art. 63 [46][47][18] P Operate a documented QMS covering regulatory strategy, design/development/testing, data controls, risk management, post-market monitoring, incident reporting, documentation, communications, responsibilities and resources. Same high-risk dates. Future binding. SME simplification exists in law but Commission SME-QMS guidance is still an implementation dependency.
Retain high-risk documentation and logs Arts. 18–19 [48][49] P Keep specified technical/QMS/conformity documentation for 10 years after market/service; retain automatically generated logs under provider control for at least six months, unless other applicable law specifies otherwise. Same high-risk dates, then retention periods run from relevant event. Settled future requirement.
Corrective action and authority cooperation Arts. 20–21 [49][26] P If a system is non-compliant or risky, take corrective action including bringing into compliance, withdrawal/disablement/recall as appropriate and provide required notifications; furnish authorities with information/documentation/log access on reasoned request. Same high-risk dates. Settled future requirement.
EU representative for third-country high-risk providers Art. 22 [26] Non-EU P, AR Appoint an EU authorized representative before making the relevant high-risk system available; representative performs mandate, retention and cooperation tasks. Relevant high-risk date: 2 Dec 2027 or 2 Aug 2028. Settled future requirement.
Importer verification Art. 23 [50][51] I Before placing a high-risk system on market, verify conformity assessment, technical documentation, CE marking, provider/representative details; do not market a suspected non-conforming/risky system; preserve compliance and cooperate with authorities. Relevant high-risk date. Settled future requirement.
Distributor verification Art. 24 [25] Dist Check CE marking, declaration/instructions and key provider/importer obligations before distribution; withhold non-conforming systems; maintain suitable storage/transport conditions; correct/withdraw/recall and notify where required; cooperate with authorities. Relevant high-risk date. Settled future requirement.
Role conversion after rebranding/modification/purpose change Art. 25(1)–(3) [25][28] D/I/Dist/TP/PM A distributor, importer, deployer or other party can become the legal provider if it puts its brand on a high-risk system, substantially modifies it, or changes intended purpose so it becomes high-risk; product manufacturers can likewise become providers in specified cases. Relevant high-risk date. Settled future law; “substantial modification” remains fact-sensitive.
Mandatory high-risk supply-chain agreement Art. 25(4) [28] P, TP High-risk provider and third party supplying an AI system/model/tool/service/component/process used in that high-risk system must, by written agreement, specify the necessary information, capabilities, technical access and assistance allowing the provider to comply. Certain open-source components are exempt, but the exclusion does not generally remove GPAI-model issues. Relevant high-risk date. Important new/settled requirement; model contractual terms remain voluntary/pending.
High-risk deployer controls Art. 26(1)–(7), (12) [28][52] D Use according to instructions; assign competent/trained/authorised human oversight; ensure controlled input data is relevant/sufficiently representative; monitor the system, respond to risk/incident indications, keep controlled logs for at least six months, and cooperate with authorities. Annex III normally 2 Dec 2027; product systems 2 Aug 2028, depending on class. Future binding.
Employee and affected-person notices Art. 26, incl. workplace and Art. 26(11) notices [52][53] D Before workplace deployment of covered high-risk AI, give the statutory information to worker representatives/affected workers; for Annex III systems making or assisting decisions about individuals, inform people that they are subject to use of the high-risk system. Relevant high-risk date. Future binding. Employment/privacy law may add more.
Fundamental-rights impact assessment Art. 27 [53] Certain D Before first use of qualifying Article 6(2) systems, public-law bodies, private entities providing public services and deployers of specified Annex III 5(b)/(c) systems must assess processes, affected groups, risks, human oversight and mitigation/governance arrangements and update where necessary. 2 Dec 2027 for the relevant Annex III systems. Future binding; templates/guidance may evolve.
Conformity assessment Art. 43; Annexes VI–VII [54] P Complete the applicable conformity-assessment route before placing/putting into service a high-risk system. The route differs by category; for Annex III point 1 systems, absence/non-use of harmonised standards or common specifications can require notified-body involvement. Art. 43 formally lies outside delayed Sections 1–3 and is generally applicable from 2 Aug 2026, but its operative market-entry gate attaches when the relevant system becomes subject to the high-risk regime: Dec 2027/Aug 2028. Binding architecture; standards are a major pending dependency.
EU declaration of conformity and CE marking Arts. 47–48 [55][8] P Draw up/sign and maintain an EU declaration for each high-risk system, keep it for 10 years, assume responsibility for conformity, and affix the required CE marking. Same practical high-risk market-entry timing. Binding architecture; Annex V can be updated by delegated act.
EU-database registration Art. 49 [8] P/AR; public D Before market/service, register Annex III high-risk systems (subject to statutory exceptions) and register systems for which provider relies on Art. 6(3); specified public deployers also register their use. Practical trigger 2 Dec 2027 for Annex III classification, subject to exact category. Binding future requirement. Database implementation remains operationally important.
Real-world testing outside sandbox Arts. 60–61 [56] P/prospective P Where using the statutory real-world-testing route, comply with testing plan, registration/governance/safety requirements and obtain documented, informed consent from participating subjects where Article 61 applies. Chapter VI generally applicable since 2 Aug 2026. Current but conditional: only relevant when using real-world testing provisions.
Post-market monitoring Art. 72 [22] P Establish/document a proportionate system that actively and systematically gathers and analyses lifetime performance data and evaluates continued compliance; maintain a post-market monitoring plan in the technical documentation. Art. 72 formally generally applies from 2 Aug 2026, but substantive operation for Article 6 high-risk systems follows their high-risk compliance trigger. Commission plan guidance/template due 2 Sep 2027. Binding architecture; guidance expressly pending.
Serious-incident reporting Art. 73 [57] P; D's awareness can start timing Report qualifying serious incidents to relevant market-surveillance authority. General maximum: 15 days after awareness, once causal link/reasonable likelihood established; widespread infringement/specified serious incidents: no later than 2 days; death: no later than 10 days. Investigate and take corrective action. Same practical high-risk timing. Settled reporting rules.
Explanation of significant individual decisions Art. 86 [58] D Where an affected person is subject to a qualifying decision based on an Annex III high-risk system that legally or similarly significantly adversely affects them, provide clear and meaningful explanations of the AI system's role and main elements of the decision, unless an applicable exception/other EU right controls. Article 86 formally applies from 2 Aug 2026; because it predicates the right on an Annex III “high-risk” system, its ordinary practical significance aligns with the 2 Dec 2027 Annex III regime. Binding, with a timing/predicate nuance.

Two timing nuances deserve emphasis. First, the 2026 Omnibus specifically postpones Chapter III Sections 1–3—classification, technical requirements and operator obligations—rather than literally rewriting the application date of every later article mentioning “high-risk AI system.” Thus Articles 43, 49, 72, 73 and 86 formally fall within the generally applicable part of the Act from 2 August 2026, but their predicates depend on a system being subject to the relevant high-risk regime. The sound compliance interpretation is therefore to build them into the Dec-2027/Aug-2028 high-risk launch gate rather than conclude that non-yet-applicable Article 6 systems unexpectedly needed CE marking in August 2026. This reading is also consistent with the Commission's public description of the 2026 amendment as postponing the high-risk rules to December 2027/August 2028. [7][10]

Second, Article 111 gives substantial grandfathering to existing high-risk systems: except for Article 5, operators of systems placed on the market or put into service before the applicable Chapter III date generally become subject to the Regulation where the system undergoes significant design changes after that date. High-risk systems intended for use by public authorities have an ultimate compliance date of 2 August 2030. [59]

What is settled and what is still moving

The AI Act is now mature enough that it is misleading to call the whole regime “pending.” The better distinction is between binding duties whose content is already fixed, binding duties whose technical means remain under development, and areas where the Commission retains explicit power to change scope or detail.

The July 2026 Digital Omnibus is law, not a pending Commission proposal. The November 2025 proposal that had suggested simplification and postponement was enacted, in modified final form, as Regulation (EU) 2026/1744. The final regulation changed the high-risk timetable, clarified/narrowed parts of the product safety-component analysis, modified Article 4's literacy language, introduced the new Article 5 sexual/intimate/child-abuse prohibitions, expanded proportionality measures including for small mid-caps, changed enforcement arrangements and made other implementation changes. Any compliance memo still stating that the high-risk rules became generally applicable on 2 August 2026 is therefore obsolete. [60][2][3][7]

High-risk classification guidance is late and not final. Article 6(5) itself specified a Commission-guidelines deadline of 2 February 2026. The Commission issued draft guidelines only in May 2026; consultation closed on 23 July, and the Commission now promises final adoption by the end of 2026. This matters especially to enterprise-software vendors selling “assistive” HR, education, credit, insurance or public-sector tools whose output may or may not materially influence the protected decision. [20][9]

Harmonised standards remain a major implementation dependency. Under Article 40 and the New Legislative Framework structure, following a harmonised standard referenced in the Official Journal is voluntary but can create a presumption of conformity for requirements it covers. The Commission's current standardisation materials say standards are still being developed, and its AI Act FAQ explicitly identifies delayed standard availability as a reason for moving the high-risk application dates. Vendors therefore should not treat CEN/CENELEC drafts as binding legislation, although aligning architecture and evidence with likely risk-management, QMS, data-governance, robustness and cybersecurity standards can reduce future rework. [61][21][10]

For certain biometric high-risk systems, this is more than a paperwork issue. Article 43 permits internal control in specified circumstances when harmonised standards/common specifications are used; where they do not exist or are not applied as Article 43 specifies, a notified-body route can be required. [54]

Article 50 is considerably more mature. The Commission has issued implementation guidance and a final Code of Practice on Transparency of AI-generated Content. On 31 July 2026 it stated that the Commission and AI Board had confirmed the Code as an adequate voluntary tool for demonstrating Article 50 compliance. The statutory obligations remain binding whether or not a company signs the Code; the advantage of adherence is an accepted evidentiary/compliance pathway. [62][63]

The Article 50 Code matters particularly for Article 50(2)'s technical requirement. A product statement saying “our service uses AI” is not a substitute for the provider-side obligation to make synthetic output machine-readable and detectable as artificially generated/manipulated where Article 50(2) applies. Separately, the customer's/deployer's Article 50(4) disclosure duty for deepfakes is not automatically discharged merely because the provider placed metadata in a file. [11][12][62]

GPAI implementation is also relatively mature. The Commission has final GPAI-provider guidelines, a voluntary GPAI Code of Practice whose Transparency and Copyright chapters map to Article 53 and whose Safety and Security chapter maps to Article 55 systemic-risk duties, and an official training-content-summary template. A GPAI provider that does not adhere to an approved code or relevant harmonised standard must demonstrate alternative adequate means of compliance for Commission assessment. [14][37][39][15][16]

The GPAI regime is not completely frozen, however. Article 51(3) permits delegated changes to systemic-risk thresholds/benchmarks; Article 53(5)–(6) permits delegated methodologies and amendments to Annexes XI/XII; and the eventual arrival of harmonised GPAI standards can alter the most convenient route for demonstrating conformity. [35][15][36]

Product-sector overlap is deliberately still moving. New Article 2(13) empowers the Commission to limit duplicative high-risk AI requirements for Article 6(1) product systems where listed Union harmonisation law already supplies an equivalent or higher level of health, safety or fundamental-rights protection. Crucially, the amended statute says the Commission shall adopt those delegated acts by 2 August 2027. Product vendors—medical/device, radio, machinery and other regulated-product businesses in particular—therefore cannot finalize all AI Act/sector-law overlap assumptions today. [4]

The Omnibus itself also changed the product-law perimeter, including moving the Machinery Regulation into the sectoral approach reflected in Annex I Section B. A generic software-only vendor may never encounter this issue, but a company embedding its software in regulated products should perform a product-specific Article 2/6/Annex I analysis rather than reuse a generic SaaS checklist. [64]

Post-market-monitoring guidance is expressly outstanding. Amended Article 72 now requires the Commission, taking utmost account of the Board's opinion, to adopt guidance including a post-market-plan template by 2 September 2027. The obligation to operate post-market monitoring is already in the statute; the precise Commission template is not. [22]

SME simplification is partly dependent on further guidance. Article 63 allows qualifying SMEs—including start-ups, provided they do not have partner or linked enterprises within the meaning of the referenced SME framework—to comply with certain QMS elements in simplified form, but requires the Commission to develop guidelines on which elements can be simplified. Article 63 expressly says this does not relieve those companies of the substantive Articles 9–15, 72 and 73 obligations. [18]

Likewise, Article 62 requires conformity-assessment fees to take account of SME provider needs and be reduced proportionately to size, market size and related indicators. The presence of these measures should not be mistaken for an exemption from conformity assessment. [17]

In the EUR-Lex implementation materials reviewed for this report, the two prominent dedicated Commission implementing regulations already adopted are Commission Implementing Regulation (EU) 2025/454, governing the scientific panel of independent AI experts, and Commission Implementing Regulation (EU) 2026/1755, setting detailed arrangements for certain Commission proceedings under the AI Act. The latter is principally enforcement/procedure machinery rather than a specification of how an ordinary SaaS product must be engineered. [65][66][67]

A practical consequence is that no company should tell an auditor or customer that “EU implementing acts now completely define high-risk compliance.” They do not. Technical standards, final classification guidance, post-market guidance and certain delegated acts remain outstanding. [9][10][22]

Corrigenda require a language-version caveat. EUR-Lex records AI Act corrigenda in 2026; for example, the 27 March 2026 corrigendum expressly states that it does not concern the English version. The Official Journal browser also lists an AI Act corrigendum entry dated 18 August 2026. The search material available for that later entry does not establish a substantive English-language correction, so it would be unsafe either to claim that it changes the English obligations or to pretend it does not exist. For English-language analysis this report therefore uses the consolidated 27 July 2026 text plus the authentic 2026/1744 amendment, while flagging the 18 August entry for language-specific verification where a local-language text controls operational work. [68][69][1]

The Commission is also required by Article 112 to review the Annex III high-risk list and Article 5 prohibited-practices list regularly. A Commission Article 112 assessment, COM(2026) 234 final of 20 May 2026, is part of that continuing review process. Such assessments are not themselves amendments: current obligations remain those in the enacted/consolidated Regulation unless a delegated act or legislation actually changes them. [70][13]

EU, data-protection, and member-state interpretations

The AI Act does not displace the GDPR. Where an AI product trains on, ingests, infers from, stores, outputs or otherwise processes personal data, the company has to run the AI Act and data-protection analysis in parallel. France's CNIL states this explicitly and updated its AI-development recommendations in January 2026: where personal data is used in AI development, both regimes apply. [71][32]

That distinction is especially important for generative AI. Article 50 may tell a provider to mark synthetic content, but it does not provide the lawful basis for processing personal data contained in prompts or model training. Likewise, satisfying Article 10's high-risk data-governance requirements does not by itself satisfy GDPR purpose limitation, lawful basis, transparency, data minimisation, security or data-subject-rights obligations. [71][72]

The EDPB's Opinion 28/2024 on AI models is therefore a material adjacent source. Among other things, the EDPB has said that legitimate interests may, in appropriate circumstances, provide a legal basis for development or deployment of an AI model/system; that is a case-specific GDPR conclusion, not a blanket AI exception. The EDPB/EDPS reiterated this position while commenting on the Digital Omnibus in February 2026. [73][74]

The EDPB and EDPS also scrutinized the Omnibus proposal's special-category-data derogation, supporting the goal of a narrow mechanism for incidental/residual sensitive-data processing while recommending clearer scope and lifecycle safeguards. Because Regulation 2026/1744 has since been enacted, that joint opinion is useful interpretive history but the final amended Article 4a is the governing statute. [73][75][3]

Article 4a should consequently not be read as a general authorization to build sensitive-trait models. It is a specialized AI Act mechanism with safeguards; the remaining GDPR framework continues to apply. [6][71]

There is already meaningful national supervisory guidance, but companies should distinguish it from the directly applicable Regulation.

Germany. The Bundesnetzagentur has issued AI-literacy material interpreting Article 4 broadly across providers and deployers and recommending a context-specific approach rather than a mandatory standardized training course; its guidance also emphasizes documentation of measures and encompasses personnel acting on the organization’s behalf. [76]

There is an important temporal caveat: national Article 4 materials written before 27 July 2026 may use the original “ensure a sufficient level of AI literacy” formulation. The Omnibus has since softened/clarified the operative text toward taking measures to support development of AI literacy and explicitly says a particular literacy level need not be guaranteed. The amended EU text controls over stale wording in earlier guidance. [6][77][3]

Netherlands. The Dutch Autoriteit Persoonsgegevens' AI Act information was updated on 14 August 2026. It treats the 2 August transparency regime as live and has specifically advised organizations to consider signing the transparency Code of Practice; the AP also describes itself as coordinating Dutch supervision of algorithms and AI presenting risks to fundamental values and rights. [78][79][80][81]

France. CNIL's current approach emphasizes cumulative AI Act/GDPR compliance rather than treating AI Act conformity as a data-protection safe harbor. Its 2026 recommendations address AI-development issues such as legal basis, information to data subjects and data-source accountability, making them highly relevant to SaaS vendors using customer/user data or web-sourced data in model development. [71][82][72]

Spain. Spain's AESIA has developed practical AI Act compliance materials, including guidance dealing with Article 50 transparency and high-risk-system questions. These are valuable implementation aids but do not substitute for the EU-level legal text or eventual Commission/harmonised-standard framework. [83][84][85]

Member-state guidance can matter greatly to enforcement practice, but substantive AI Act obligations remain directly applicable EU law. National differences remain relevant for enforcement procedure and penalties, sector-specific rules, data protection, employment law, and areas where the AI Act expressly allows national restrictions—for example, Article 26 permits Member States to introduce more restrictive rules concerning certain post-remote-biometric-identification uses. [7][53][19]

This means that a software vendor selling throughout the EU should have one EU-level product compliance classification, but may still need country overlays for specific deployments such as employment, healthcare, law enforcement, public-sector use and biometrics. [53][71]

SME compliance checklist and timeline

For a typical SME SaaS vendor that develops AI-enabled software but does not train its own frontier/GPAI model, the practical 2026 compliance program can be relatively focused.

For a non-EU SME, add a specific representative analysis: Article 54 applies if the company is the third-country provider of a GPAI model, and Article 22 applies to third-country providers of covered high-risk systems once that regime becomes operative. [27][26]

The core timetable after the 2026 Omnibus is:

timeline
    title EU AI Act compliance milestones for a software vendor
    2024-08-01 : AI Act enters into force
    2025-02-02 : Article 4 AI-literacy duties apply
               : Most Article 5 prohibitions apply
    2025-08-02 : GPAI-model regime applies to new models
               : Governance and most penalty chapter provisions apply
    2026-07-27 : Regulation 2026/1744 Digital Omnibus enters into force
    2026-08-02 : General application date
               : Article 50 transparency obligations apply
    2026-12-02 : New Article 5 sexual/intimate and child-abuse prohibitions apply
               : Pre-Aug-2026 synthetic generators must meet Article 50(2)
    2027-08-02 : Pre-Aug-2025 GPAI models must comply
               : Article 2(13) sectoral-overlap delegated acts due
    2027-09-02 : Commission Article 72 post-market-plan guidance/template due
    2027-12-02 : Annex III high-risk Sections 1-3 apply
    2028-08-02 : Article 6(1) / Annex I product high-risk Sections 1-3 apply
    2030-08-02 : Public-authority legacy high-risk systems compliance backstop
    2030-12-31 : Specified legacy large-scale EU IT systems compliance date

The dates above come from amended Articles 111 and 113, Article 72, and the new Article 2(13) deadline. [7][13][22][4]

The compliance priority today is consequently not “get high-risk CE certification immediately.” It is: get scope and roles right; stop prohibited uses; have defensible AI-literacy measures; make Article 50 work technically; comply with GPAI rules if the company really is a model provider; classify potential high-risk products early; and build the evidence/control system during the now-expanded transition period. [7][14][10]

Primary sources and current implementation materials

The authoritative starting point is the consolidated Artificial Intelligence Act, Regulation (EU) 2024/1689, version of 27 July 2026, together with the authentic Regulation (EU) 2026/1744 — Digital Omnibus on AI. [1][3]

For current Commission implementation status, the most useful official sources are the Commission's Navigating the AI Act FAQ, the AI Act standardisation page, and the high-risk classification draft-guidelines consultation. The last source confirms that the consultation closed on 23 July 2026 and that final guidelines are expected by the end of 2026. [10][21][9]

For general-purpose models, the principal official sources are the Commission's GPAI-provider guidelines, the General-Purpose AI Code of Practice, and the AI Office's training-content-summary materials. Together they now provide a materially more developed compliance framework than existed when Chapter V first became applicable in August 2025. [14][37][39]

For Article 50, the Commission's Code of Practice on Transparency of AI-generated Content is now final and has been recognized by the Commission and AI Board as an adequate voluntary compliance tool. [62]

The two dedicated implementing regulations particularly worth tracking are Commission Implementing Regulation (EU) 2025/454, concerning the scientific panel, and Commission Implementing Regulation (EU) 2026/1755, concerning Commission proceedings and enforcement procedure. [65][66]

For the GDPR overlay, the EDPB's Opinion 28/2024 and subsequent EDPB/EDPS Digital Omnibus opinions are important interpretive sources; France's CNIL maintains especially detailed 2026 AI-development/GDPR recommendations. [74][73][71]

At Member-State level, useful official implementation materials currently include the Netherlands Autoriteit Persoonsgegevens' updated AI Act and Article 50 materials, Germany's Bundesnetzagentur AI-literacy guidance, France's CNIL materials, and Spain's AESIA practical AI Act guides. These sources are valuable indicators of supervisory expectations, but where their wording conflicts with the post-27-July-2026 consolidated Regulation, the amended EU legislation takes precedence. [78][79][76][32][83]

Overall legal assessment as of 31 August 2026: the immediate obligations for a typical AI-enabled software vendor are substantially settled—Articles 4, 5 and 50, plus Chapter V where the company is itself a GPAI-model provider. The future high-risk duties are also enacted law, and their new Dec-2027/Aug-2028 dates are now fixed by Regulation 2026/1744; what remains materially uncertain is less whether those duties will exist than precisely which borderline systems fall into them and which technical evidence will create the safest presumption of conformity, because final Article 6 guidance, harmonised standards, sector-overlap delegated acts and several implementation templates are still outstanding. [7][9][21][4][22]

Sources

  1. No link given.
  2. No link given.
  3. No link given.
  4. No link given.
  5. No link given.
  6. No link given.
  7. No link given.
  8. No link given.
  9. https://di
  10. No link given.
  11. No link given.
  12. No link given.
  13. No link given.
  14. No link given.
  15. No link given.
  16. No link given.
  17. No link given.
  18. No link given.
  19. No link given.
  20. No link given.
  21. No link given.
  22. No link given.
  23. No link given.
  24. No link given.
  25. No link given.
  26. No link given.
  27. No link given.
  28. No link given.
  29. No link given.
  30. No link given.
  31. No link given.
  32. No link given.
  33. No link given.
  34. No link given.
  35. No link given.
  36. No link given.
  37. https://di
  38. No link given.
  39. https://digital-strategy
  40. No link given.
  41. No link given.
  42. No link given.
  43. No link given.
  44. No link given.
  45. No link given.
  46. No link given.
  47. No link given.
  48. No link given.
  49. No link given.
  50. No link given.
  51. No link given.
  52. No link given.
  53. No link given.
  54. No link given.
  55. No link given.
  56. No link given.
  57. No link given.
  58. No link given.
  59. No link given.
  60. No link given.
  61. No link given.
  62. No link given.
  63. No link given.
  64. No link given.
  65. No link given.
  66. No link given.
  67. No link given.
  68. No link given.
  69. No link given.
  70. No link given.
  71. No link given.
  72. No link given.
  73. No link given.
  74. No link given.
  75. No link given.
  76. No link given.
  77. No link given.
  78. No link given.
  79. No link given.
  80. No link given.
  81. No link given.
  82. No link given.
  83. No link given.
  84. No link given.
  85. No link given.