EU AI Act, 2 August 2026: What Actually Changes for European SaaS Buyers

Three concrete obligation sets switch on. Each is enforced by the European AI Office and, in parallel, by the national AI supervisory authorities that member states have designated under Article 70.

**1. Technical documentation and transparency for all GPAI providers (Article 53).** Any provider placing a general-purpose model on the EU market must maintain up-to-date technical documentation covering training data provenance, energy consumption, evaluation methodology, and known limitations. A summary sufficient for downstream deployers to comply with their own obligations must be made publicly available. "Placed on the market" is interpreted broadly by the AI Office guidance published 18 April 2026 — API access from EU IP addresses is placement, even when the provider is US-headquartered.

**2. Copyright and training data disclosure (Article 53(1)(c)–(d)).** Providers must implement a policy to comply with EU copyright law, including the Article 4(3) text-and-data-mining opt-out reservation, and publish a sufficiently detailed public summary of training content. The AI Office template released in May 2026 requires disclosure at the level of dataset families and major web crawls, not individual documents.

**3. Systemic-risk obligations for frontier models (Article 55).** Models trained with more than 10^25 FLOPs — currently understood to include GPT-5.6, Gemini 3 Ultra, Claude Opus 4.5, and Llama 4-405B — carry additional duties: model evaluations conducted according to state-of-the-art protocols, systemic risk assessment and mitigation, serious incident reporting to the AI Office within 15 days, and adequate cybersecurity protection for model weights.

Fines for GPAI-specific breaches are capped at €15 million or 3% of global annual turnover. Other AI Act breaches carry higher ceilings (up to 7% for prohibited practices), but those fine categories relate to obligations with later effective dates.

The Council's Digital Omnibus package, politically agreed on 29 June 2026, revised several downstream deadlines. It is important to understand precisely what moved and what did not.

**Deferred to 2 December 2026:** the obligation for providers of high-risk AI systems listed in Annex III (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice) to have completed conformity assessment and CE marking. Originally scheduled for 2 August 2026 alongside GPAI enforcement, the postponement gives providers additional time to align with harmonised standards that CEN-CENELEC JTC 21 has not yet finalised.

**Deferred to 2 August 2027:** obligations for AI systems that are safety components of products already covered by the Union harmonisation legislation listed in Annex I (machinery, medical devices, in-vitro diagnostics, radio equipment). This aligns AI Act conformity with the underlying sectoral regime.

**Not deferred:** all GPAI obligations under Article 53 and 55, the prohibited practices regime under Article 5 (in force since February 2025), the AI literacy obligation under Article 4 (in force since February 2025), and the governance provisions establishing the AI Office and Board.

The practical effect is that on 2 August 2026 there is a discontinuity in obligation depth: model providers face live enforcement while system-level deployers of high-risk AI have three to twelve additional months. For SaaS buyers this means the immediate compliance conversation is almost entirely upstream — with the model providers embedded in your vendors' stacks.

Most European organisations do not train foundation models. They procure SaaS products that embed them. The AI Act treats this chain explicitly through the "downstream provider" and "deployer" categories in Articles 25 and 26, and the exposure map is not intuitive.

**Layer 1 — Foundation model providers.** OpenAI, Anthropic, Google DeepMind, Meta, Mistral, Cohere, xAI. Direct Article 53 and (for the largest) Article 55 obligations. Enforcement risk sits with them, not you.

**Layer 2 — Model-integrating SaaS vendors.** Notion AI, Salesforce Einstein, HubSpot Breeze, Atlassian Rovo, GitHub Copilot, Microsoft 365 Copilot, Zendesk AI. These vendors are "downstream providers" under Article 25 when they modify or substantially adapt an upstream GPAI model, or "deployers" under Article 26 when they merely wrap it. The classification determines whether they inherit provider obligations or only user-facing ones. Most vendors have chosen deployer classification, which pushes documentation duties back onto the underlying model provider.

**Layer 3 — Your organisation as deployer.** You inherit Article 26 obligations whenever you use an AI system in a professional capacity. These include ensuring human oversight, monitoring for anomalies, cooperating with authorities, and — critically — informing workers and their representatives before deploying AI systems in the workplace (Article 26(7)). For high-risk deployments there are additional fundamental rights impact assessment duties, but those are on the deferred timeline.

**Layer 4 — End users.** Your customers or employees receive the Article 50 transparency obligations: they must be informed when interacting with an AI system, when content is AI-generated, and when biometric or emotion recognition is used.

The exposure buyers most consistently miss is Layer 2. When a Layer 2 vendor's underlying model provider is out of compliance on 2 August, the vendor's service may become unlawful to use in the EU even if the vendor itself has done everything correctly. Contractual pass-through of AI Act warranties is now standard in enterprise MSAs; it should be standard in mid-market ones too.

The GPAI Code of Practice, finalised by the AI Office on 10 July 2025 and endorsed by the AI Board in September 2025, functions as a de facto safe harbour. Providers who sign and adhere to it benefit from a presumption of conformity with Article 53 obligations; those who do not must demonstrate equivalent compliance through their own means, without the presumption.

As of June 2026, the confirmed signatories include Mistral, Anthropic, Google, OpenAI, Microsoft, IBM, Cohere, and Aleph Alpha. Meta declined to sign, citing concerns about the copyright chapter. xAI signed only the safety and security chapter. Several Chinese providers (including DeepSeek and Alibaba's Qwen team) are not signatories and have not indicated intent to sign.

For buyers this creates a bifurcated risk landscape. When your SaaS vendor's underlying model is provided by a signatory, the compliance evidence chain is standardised, auditable, and mapped to the Code's chapters on transparency, copyright, and — for systemic-risk models — safety and security. When the underlying model is from a non-signatory, the vendor must produce equivalent evidence bilaterally, and the AI Office has publicly indicated (guidance of 12 June 2026) that it will apply "enhanced scrutiny" to non-signatory GPAI models during the first enforcement phase.

The practical implication is that signatory status now belongs in vendor questionnaires alongside SOC 2, ISO 27001, and data residency. It is a leading indicator of enforcement risk on 2 August and beyond.

Generic AI Act readiness statements from vendors are close to worthless. The following six questions are calibrated to produce evidence rather than reassurance, and map directly to the obligations that become enforceable on 2 August.

**1. Which GPAI models are embedded in the service, and is each provider a signatory to the GPAI Code of Practice?** Ask for the list at the model version level (GPT-5.6-turbo, not "OpenAI"). Signatory status per Article 56 is the fastest proxy for upstream compliance.

**2. For each embedded model, provide the Article 53(1)(a) technical documentation summary and the Article 53(1)(d) training data summary.** Both must be publicly available for signatory providers. If the vendor cannot supply URLs, the underlying provider is likely not compliant.

**3. Under Article 25, do you classify yourself as a downstream provider or a deployer of the embedded model?** The distinction determines whether the vendor inherits provider obligations. Deployer classification means you should verify the upstream provider directly.

**4. What is your process for notifying customers when an embedded model is replaced or materially updated?** Article 25(4) requires downstream providers to inform deployers of substantial modifications. This should be a contractual notification obligation, not a courtesy.

**5. Do any embedded models meet the Article 51 systemic-risk threshold, and if so, provide evidence of the Article 55 evaluations conducted?** Systemic-risk models require model evaluations, adversarial testing, and incident reporting. The vendor should be able to point to the upstream provider's Model Evaluation Report.

**6. What is your position on Article 26(7) worker information obligations when your service is used in employment contexts?** This is a deployer obligation but vendors targeting HR, productivity, or communication use cases should provide standard notification templates.

If a vendor cannot answer four of the six by mid-July 2026, it is a signal to escalate the account for legal review before 2 August.

The window between publication of this article and the enforcement date is short. A pragmatic procurement response has four components.

**Inventory.** Enumerate every SaaS product in the estate that has an AI feature, whether or not it is enabled. The definition to apply is Article 3(1): an AI system is one that infers, from input, how to generate outputs that influence physical or virtual environments. This is broader than "LLM-powered". It captures ranking, recommendation, anomaly detection, forecasting, and classification.

**Segmentation.** For each system, identify the underlying GPAI model and its Code of Practice signatory status. This is a two-question vendor request and should complete within two weeks for a mid-sized estate.

**Contract review.** Review AI-related warranties in current MSAs. The 2026 enterprise standard clause requires the vendor to (a) warrant Article 53 compliance of embedded models, (b) notify of material model changes, and (c) indemnify for regulatory fines arising from upstream non-compliance. Amendments before 2 August are difficult but not impossible for the top ten vendors by spend.

**Fallback plans.** For any critical system where the underlying model is provided by a non-signatory, prepare a substitution plan. This does not mean immediate migration — enforcement will not be instantaneous on 2 August — but it means knowing which sovereign or signatory-based alternative you would move to if the AI Office issues a suspension order.

The organisations that treat 2 August as a moving target — not a single date — will manage the transition better than those that treat it as a compliance event to be survived.