Compliance
Guardrails & audit trails for the EU AI Act
The EU AI Act’s timeline just changed: the digital-omnibus package adopted in June 2026 pushed the high-risk obligations back by more than a year – but not everything moved. Transparency duties still land on 2 August 2026, the GPAI rules are already in force, and the Act’s extraterritorial reach and €35 million / 7% penalty ceiling are untouched. Here is where each obligation actually stands.
The honest part, up front: no tool makes you compliant. Compliance is a property of your organization – its processes, contracts, and documentation – not of any component you install. What a guardrails proxy provides is technical measures and evidence that support the obligations below. Assess your own obligations with counsel.
The timeline after the 2026 omnibus
| Obligation | Original date | Where it stands now |
|---|---|---|
| Prohibited practices (Article 5) | 2 February 2025 | In force – unchanged |
| GPAI model obligations (Articles 51–56) | 2 August 2025 | In force – unchanged; falls mainly on model providers |
| Transparency (Article 50): chatbot disclosure, AI-content marking | 2 August 2026 | Still 2 August 2026 – not deferred (a 4-month watermarking grace applies to systems already on market) |
| High-risk systems, Annex III (employment, credit, essential services…) | 2 August 2026 | Deferred to 2 December 2027 |
| High-risk systems, Annex I (AI embedded in regulated products) | 2 August 2027 | Deferred to 2 August 2028 |
| Extraterritorial reach & penalties (up to €35M or 7% of global revenue) | – | Unchanged |
What the deferral does – and doesn’t – buy you
The omnibus replaced the old cliff with fixed later dates because member states and standards bodies weren’t ready – not because the obligations shrank. Logging and traceability, human oversight, and accuracy and robustness controls all survive intact; they simply bind Annex III deployments from December 2027. And the nearest deadline is now the one teams overlook: Article 50 requires that people are told they are interacting with AI and that generated content is machine-readably marked, from 2 August 2026, for systems in scope regardless of risk class.
For teams running LLM applications and agents, the practical weight of the high-risk chapters lands on a few recurring themes: risk management as a process, logging and traceability, human oversight over consequential operation, and accuracy and robustness controls. These are runtime properties, and the mapping below shows how they relate to runtime guardrails – as education, not as a compliance solution. If the Act applies to you, plan with counsel.
Where in/guard/out helps
Logging & traceability – high-risk systems must automatically record events over their lifetime, and deployers must retain logs (six months at minimum under Articles 19 and 26)
Every request records which checks ran, their verdicts, timings, and costs; agent runs reconstruct into complete traces across model calls, tool calls, and workflow checkpoints. Retention is a plan setting; screened-by-default logging keeps the archive itself low-risk.
Human oversight – Article 14 requires systems be designed so humans can effectively oversee them, understand behavior, and intervene
Approval tiers put a human decision in front of designated actions (a payment, a deletion) via the checkpoint API; the run graph makes what the agent actually did legible to the person overseeing it; PREVENT mode is the intervention path, enforced in line.
Accuracy & robustness – Article 15 requires appropriate accuracy, robustness, and cybersecurity, resilient to errors and manipulation attempts
Grounding and fact-checking measure and enforce output accuracy per request; injection defense addresses the manipulation channel unique to LLMs; format enforcement and fail-closed behavior contain error propagation.
Risk management – Article 9 requires an iterative risk-management process across the lifecycle
Observe-then-enforce is that loop, instrumented: run checks in FIX mode to measure risk on real traffic, review findings, tighten to PREVENT where evidence warrants – with the dashboard as the record of the process.
Transparency to deployers – systems used with or affecting people carry information duties
Response headers and per-request reports state what was checked and what was found – machine-readable evidence your application layer can surface as its transparency obligations require.
Frequently asked questions
Did the August 2026 deadline really move?
For high-risk obligations, yes. The digital-omnibus package – endorsed by Parliament and Council in June 2026 – defers Annex III high-risk duties to 2 December 2027 and Annex I (embedded) to 2 August 2028, as fixed dates. Article 50 was not deferred: chatbot disclosure and AI-content marking still apply from 2 August 2026. Verify the final Official Journal text with counsel.
Does using a guardrails proxy make my AI system EU AI Act compliant?
No – and no product honestly can. Compliance involves classification, documentation, conformity assessment, and organizational processes. in/guard/out contributes technical measures (logging, oversight hooks, robustness controls) and the per-request evidence that your documentation can cite.
Are most LLM applications “high-risk” under the Act?
Most are not – the high-risk category is defined by specific use areas (Annex III) such as employment, credit, and essential services. But the Act’s logging, transparency, and oversight themes are becoming the de-facto bar customers expect, whatever your classification. Classify with counsel.
What does the Act say about general-purpose AI models?
GPAI obligations (transparency, copyright policy, training-data summaries) fall mainly on model providers, with obligations phasing from August 2025. As a deployer building on those models, your duties concentrate on how you use them – which is the layer in/guard/out instruments.
How does the six-month log retention interact with data minimization?
This is exactly why screened-by-default logging matters: transcripts are stored with PII replaced by placeholders, so retaining operational logs for traceability does not mean warehousing personal data. Raw logging remains a deliberate, audited opt-in.
Related
Bring your compliance team to the demo.
The per-request evidence trail tends to answer their questions faster than a slide deck. We’re running a limited demo – sign up and we’ll get you in as soon as we can.