The EU AI Act's High-Risk Deadline Just Moved to 2027. Here's What Salesforce Teams Should Do With the Extra Year.

If you built a 2026 compliance plan around the EU AI Act's August deadline, that plan changed in July.

On 24 July 2026 the European Union published the Digital Omnibus on AI — Regulation (EU) 2026/1744 — in the Official Journal. It entered into force on 27 July 2026, and it is the first formal amendment to the AI Act since the Act was adopted in June 2024. Its most consequential effect: the high-risk obligations that were scheduled to bite on 2 August 2026 have been deferred by sixteen months.

A lot of compliance content still says August 2026. A lot of internal roadmaps still say August 2026. Here is what actually applies now, and — more importantly — what did not change at all.

What exactly changed on 27 July 2026?

The Omnibus moved the application date for high-risk AI system obligations. It split that deferral in two, depending on how the system is classified:

Stand-alone high-risk systems — those classified under Article 6(2) and listed in Annex III — now apply from 2 December 2027. Annex III covers biometrics, critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services (including creditworthiness assessment and life and health insurance pricing), law enforcement, migration, and the administration of justice.

Embedded high-risk systems — those falling under Article 6(1) and Annex I Section A, meaning AI that is a safety component of a product already regulated under EU harmonised product-safety legislation — now apply from 2 August 2028.

That is the change. It is a timing amendment, not a rewrite of the obligations themselves. Articles 9 through 17 still describe what providers of high-risk systems must do: risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness. Article 26 still describes deployer duties. The substance survived intact. Only the clock moved.

What did not change?

This is the part that gets lost in the headlines, and it is the part that matters most for anyone running AI inside a CRM.

The prohibited-practice bans are in force and have been since 2 February 2025. Social scoring, emotion inference in workplaces and educational settings, untargeted scraping of facial images, and manipulative or exploitative AI systems are banned now. The Omnibus did not touch them. These carry the AI Act's heaviest penalty tier — up to €35 million or 7% of global annual turnover, whichever is higher.

General-purpose AI transparency obligations are in force and have been since 2 August 2025. If you build on or distribute a GPAI model, those duties apply today.

GDPR was not amended by any of this. This is the single most important sentence in this article. The AI Act is a product-safety regulation layered on top of European data protection law — it does not replace it, suspend it, or defer it. Every obligation you have under GDPR for personal data that flows into an AI system is an obligation you have this morning:

• You need a lawful basis under Article 6 for processing personal data to train, fine-tune, or ground an AI system. "We already had the data in our CRM" is not a lawful basis for a new processing purpose.
• Purpose limitation under Article 5(1)(b) still binds you. Data collected to service a customer relationship was not collected to train a model.
• Data minimisation under Article 5(1)(c) still binds you. Feeding an entire production object into a prompt because it was easier than scoping the query is the textbook failure here.
• The right to erasure under Article 17 still binds you — and it reaches AI-generated records, embeddings, and derived data, not just the original row you deleted.

Regulators have been consistent that the AI Act's phase-in does not create a grace period for data protection. The deferral bought time on one regulation. It bought nothing on the other.

Which Annex III use cases actually run on Salesforce?

It is tempting to read "high-risk AI" as someone else's problem — autonomous vehicles, medical devices, border control. For enterprise Salesforce orgs, the Annex III list is closer to home than most teams assume.

Creditworthiness assessment. AI that evaluates the credit standing of a natural person is explicitly Annex III. If Einstein scoring, a custom model, or an Agentforce agent influences lending decisions in Financial Services Cloud, that is a candidate for high-risk classification.

Life and health insurance risk assessment and pricing. Also explicitly Annex III. Insurance orgs running AI-assisted underwriting or pricing on Salesforce should be scoping this now.

Employment and worker management. AI used to recruit, screen applications, evaluate candidates, or make decisions affecting the terms of a working relationship. Any org running AI-assisted candidate screening or performance analysis on Salesforce data is in scope territory.

Education and vocational training. AI that determines admission, evaluates learning outcomes, or monitors students. Higher-education orgs running admissions workflows on Salesforce should read Annex III carefully.

Access to essential public and private services. Broader than it sounds, and worth a careful read against your service-eligibility automations.

Classification is a genuinely fact-specific legal exercise and this article is not a substitute for counsel. But the practical point stands: if AI in your Salesforce org materially influences a decision about a person's money, job, education, insurance, or access to a service, you should assume Annex III is a live question rather than an academic one.

Why training-data governance is still the work

Here is the trap in the deferral. The obligations that moved to December 2027 are precisely the ones with the longest lead time. Article 10 — the data governance requirement — expects training, validation, and testing data sets to be relevant, sufficiently representative, and to the best extent possible free of errors and complete. Article 12 expects automatic logging over the system's lifetime. Article 11 expects technical documentation that describes, among other things, the provenance of your data.

None of those are things you can produce in the last quarter before a deadline. They are properties of how your data was handled from the beginning. An organisation that spends the next sixteen months feeding unmasked production PII into AI systems will arrive at December 2027 with a data estate it cannot document, cannot explain, and cannot retroactively clean.

The teams that use this deferral well will spend it on four things:

Know where personal data actually is. You cannot govern what you have not found. Custom objects and custom fields accumulate PII in places no data map predicted, and AI grounding tends to reach exactly those places.

Mask non-production environments now. Sandboxes, test orgs, and development environments are where AI experimentation actually happens, and they are the least governed part of most Salesforce estates. Masking them removes the largest category of accidental training-data exposure at a stroke — and it is the change with the shortest implementation path.

Make erasure reach AI artefacts. If a data subject exercises Article 17 today, can you remove their data from prompts, logs, embeddings, and generated records — not just the source object? For most orgs the honest answer right now is no. That gap is a live GDPR exposure, not a 2027 one.

Capture consent at the granularity you will need. Consent to be contacted is not consent to have your data train a model. Orgs that recorded only the former will find the latter expensive to reconstruct.

What to do in the next sixteen months instead of nothing

Treat 2 December 2027 as a documentation deadline, not a build deadline, and work backwards:

Now through Q4 2026 — inventory AI use across your Salesforce estate, including the shadow uses. Classify each against Annex III. Mask non-production data. Get an accurate personal-data map.

2027 H1 — build the Article 11 technical documentation and Article 12 logging while the systems are being developed, not after. Close the erasure gap so deletion propagates to AI-derived data.

2027 H2 — run conformity assessment, human-oversight design under Article 14, and deployer-duty readiness under Article 26 against a data estate you already trust.

The deferral is a gift to organisations that use it and a trap for organisations that read it as permission to stop. The AI Act's high-risk clock now runs to December 2027. The GDPR clock never stopped.

The short version

The EU AI Act's high-risk obligations moved from 2 August 2026 to 2 December 2027 for Annex III stand-alone systems, and to 2 August 2028 for Annex I embedded systems, under Regulation (EU) 2026/1744. Prohibited practices and GPAI transparency obligations were not deferred and remain enforceable. GDPR was not amended and continues to govern every piece of personal data that reaches an AI system in your Salesforce org today.

If your AI touches credit, hiring, insurance, education, or access to services, you have sixteen extra months — and a data-governance backlog that will take most of them.

---

For the regulation-by-regulation view, see EU AI Act compliance for Salesforce; for the platform-side gaps, see what the Einstein Trust Layer does not cover.

Sources

• Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, 24 July 2026 — in force 27 July 2026
• Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 5, 6, 9–17, 26, 99, Annexes I and III
• Regulation (EU) 2016/679 (GDPR), Articles 5, 6, 17