The useful starting point is your firm’s actual AI workflow: where it operates, what the system does and who is responsible for it.
The EU AI Act can apply to non-EU firms, but having an EU client does not automatically make every AI tool high-risk. The timetable has also changed. The Digital Omnibus on AI entered into force on 27 July 2026, moving the principal high-risk application dates to December 2027 and August 2028. Some other duties already apply. European Commission update
Checked 2 October 2026. This guide explains the EU framework for planning purposes. It is educational information, not legal advice for a particular firm, matter or jurisdiction.
Start with scope
Article 2 covers several connections to the EU, including providers bringing systems or general-purpose AI models to the EU market, deployers established or located in the EU, and third-country providers or deployers where their system’s output is used in the EU.
For an international firm, map the deploying entity, office, workflow and destination of the output. A Paris office and a US team whose AI output is used in the EU need different factual assessments. London is outside the EU; a London-based operation may nevertheless have a relevant EU connection.
The nationality of the data subject or the governing law of a contract does not, on its own, answer the AI Act scope question. Data-protection obligations require their own analysis.
Legal research is not automatically high-risk
The justice category in Annex III, point 8(a) concerns systems intended to assist judicial authorities, or operate on their behalf, in researching facts and law or applying law to particular facts. It also covers similar use in alternative dispute resolution. Recital 61 explains the relevance of legally effective ADR outcomes and excludes purely ancillary administration from that justice rationale.
A private firm’s research, contract review or first-draft tool does not enter that category merely because its work influences a client’s case. That conclusion is an application of the category’s wording, not a blanket exemption for law firms.
Check other uses too. Recruitment or employee-evaluation tools can engage separate Annex III categories. Article 6 contains a conditional exception for some Annex III systems without significant harm risks, including certain procedural or preparatory tasks. It is not a general human-in-the-loop exemption; profiling within Annex III remains high-risk.
The dates to put in your register
Use the revised Article 113 and Commission timeline:
- 2 February 2025: initial prohibitions and general provisions began applying
- 2 August 2025: GPAI-model provider obligations began applying, subject to transitions
- 2 August 2026: Article 50 transparency duties generally began applying
- 2 December 2026: the Article 50(2) transition ends for providers of qualifying synthetic-content systems already marketed before 2 August 2026
- 2 August 2027: the compliance deadline for GPAI models placed on the market before 2 August 2025
- 2 December 2027: Chapter III sections 1–3, except Article 6(5), apply to Annex III high-risk systems
- 2 August 2028: those provisions apply to the Annex I high-risk product route
Existing high-risk systems have additional transitional treatment under Article 111, including rules involving significant design changes. Record the relevant market/service dates. A single firm-wide deadline will miss these distinctions.
Separate your duties from the vendor’s
High-risk providers are responsible for system compliance, quality management, documentation and the applicable conformity assessment.
High-risk deployers have different responsibilities: following instructions, assigning competent human oversight, monitoring operation and keeping available logs for the required period, among other duties. Buying software does not automatically transfer its developer’s entire compliance programme to the firm.
Roles can change. Rebranding a high-risk system, substantially modifying it, or changing a system’s intended purpose so it becomes high-risk can bring provider responsibilities under Article 25. Review that before turning a purchased tool into a client-facing product.
For ordinary procurement, ask which entity supplies the model and which supplies the application. Article 53 sets GPAI-provider duties such as documentation, copyright policies and training-content summaries, with defined exceptions. A vendor’s compliance statement does not settle your own obligations.
Transparency depends on the use
Article 50 contains specific requirements:
- Providers of directly interactive systems must inform people they are dealing with AI, unless this is obvious in context
- Synthetic-content providers have machine-readable marking duties; exceptions include standard editing assistance or uses that do not substantially alter the input or its meaning
- Deployers have disclosure duties for deepfakes and certain AI-generated text published to inform the public on public-interest matters
The public-interest text rule has an exception where human review or editorial control occurs and a person or organisation holds editorial responsibility. Deepfakes within evidently creative or fictional works have a tailored disclosure rule, rather than a blanket exemption. These provisions do not create one universal label for every AI-assisted client document. Check applicable court directions, professional rules and client commitments separately.
Ask for evidence you can use
Avoid an undefined demand for an “EU AI Act certificate.” Evidence should match the system, role and applicable rule. For some high-risk systems, the relevant material includes an EU declaration of conformity; Article 43 specifies different conformity-assessment routes.
A practical vendor file should identify the system and version, intended and prohibited uses, relevant dates, oversight instructions, logging capabilities and the documentation available to support your assessment. Escalate missing evidence rather than treating a badge as the answer.
What to do now
Build an inventory with one entry per workflow. Assign an owner, document the scope and classification rationale, and record the applicable duties and dates.
Support staff AI literacy. The amended Article 4 still requires contextual measures, while clarifying that organisations need not guarantee a particular literacy level for every individual.
Use your firm’s AI governance policy to turn approved uses, data handling, human review and incident escalation into daily practice. Obtain qualified legal review where the classification or jurisdictional analysis is uncertain.
a documented assessment of the actual workflow is more useful than a blanket warning about every legal AI tool.
Penalties depend on the breach and undertaking
Article 99 sets ceilings of €35 million/7% for prohibited practices, €15 million/3% for specified other breaches and €7.5 million/1% for specified incorrect, incomplete or misleading information. Turnover uses the preceding financial year’s worldwide annual total. The higher ceiling normally applies; SMEs receive lower-of treatment for paragraphs 3–5, while qualifying small mid-cap enterprises receive it only for paragraphs 4–5. Actual fines require a proportionate, case-specific assessment.
Read Article 99 for the precise categories and assessment factors.
AI-Assisted Research. This update uses AI-assisted research and editing. It is published under the byline of Manu Ayala. For deeper takes and the perspective behind the research, follow me on LinkedIn or email me directly.