Official EU sources checked on 21 October 2026. This article provides technical and operational guidance to support specialist legal review.
A voice assistant answering appointment questions and an assistant ranking job applicants may use the same underlying model. Their purposes, effects on people and legal requirements need separate assessment.
For a team building or operating agents, readiness starts with understanding that deployment: what the system does, who is responsible, which requirements apply and what evidence shows that the controls work.
A readiness review should produce a documented decision about the service and a plan with named owners for any remaining work. A generic checklist cannot establish that position for every agent.
Work from the current implementation dates
The EU AI Act's application dates have changed since its original adoption. The AI Omnibus amendments entered into force on 27 July 2026. The Commission confirmed extended timelines for the main high-risk rules. [1]
As at 21 October 2026, selected milestones relevant to agent projects are:
| Provision or system category | Date | Practical significance |
|---|---|---|
| Article 50 transparency obligations | 2 August 2026 | Relevant interaction notices, marking and disclosure duties already apply, subject to their scope and exceptions. |
| Limited Article 50(2) transition for certain systems placed on the market before 2 August 2026 | 2 December 2026 | This concerns synthetic-content marking and detection; it is not a general extension for all transparency duties. |
| Main high-risk requirements for Article 6(2) / Annex III systems | 2 December 2027 | Plan the applicable high-risk requirements for the specific listed use. |
| Main high-risk requirements for Article 6(1) / Annex I systems | 2 August 2028 | Assess the relevant product and safety-component conditions. |
The high-risk dates follow the amended Article 113. The Commission's Article 50 FAQ explains the narrower transition for synthetic-content marking. [2] [9]
Create an obligation register with the relevant actor, provision, application date and evidence owner. Check existing-system and other transitional rules where relevant. Later high-risk dates do not postpone transparency or data-protection duties that already apply.
Map the system and its responsibilities
Describe the whole agent service. Include its model, knowledge sources, memory, tools, voice or chat channel, identity controls and connections to business systems. Record the people affected, the actions it can perform and the geographical context for the legal scope assessment.
Use a specific intended-purpose statement. For example: “Answer approved service questions and help verified customers request appointment changes.” Record any excluded uses and the point at which a person takes over.
Then map the legal roles. Under Article 3, a provider develops a system, or has it developed, and places it on the market or puts it into service under its own name or trademark. A deployer uses a system under its authority in a professional context. [3]
The organisation providing the model, the organisation providing the assembled agent and the organisation operating it may differ. Outsourcing implementation therefore still needs a clear allocation of responsibilities and access to the information each party requires.
Revisit that allocation when the system changes. Article 25 identifies circumstances in which rebranding or modifying a system can make another party the provider of a high-risk system. The consequences depend on the particular change and conditions. [4]
Record the assessment alongside supplier documentation and the contract. A platform's general assurance statement does not explain the purpose, configuration or operating controls of the completed service.
Classify the intended use before expanding it
Assess the system against the actual classification rules. Article 6 covers the product-related route and the Annex III route to high-risk classification, including defined exceptions. For Annex III systems, the limited conditions in Article 6(3) need a reasoned assessment. That derogation does not apply to an Annex III system that profiles natural persons. [5]
Annex III expressly includes recruitment and selection activities such as analysing job applications and evaluating candidates. It also identifies particular uses involving education, creditworthiness and essential services. The presence of a chatbot or a human reviewer alone cannot settle how the actual use should be classified. [6]
For the appointment assistant, adding candidate scoring or employee assessment would require a fresh review of purpose and effect. Record how an output is used in the eventual decision, including whether it changes what a person sees or which options reach them.
Screen prohibited practices separately. Article 5 includes a prohibition on AI emotion inference based on biometric data in workplaces and educational institutions, with an exception for medical or safety reasons. A proposed voice-analytics feature needs assessment of what it actually infers and how it is used. [7]
Keep the classification decision, reasoning, responsible reviewer and review date. Use that record when reviewing proposed capabilities and deciding whether the original assessment still covers the service.
Put transparency into the first interaction
For systems intended to interact directly with people, Article 50(1) places a design obligation on providers to inform them that they are interacting with AI, unless that is obvious in the relevant circumstances. Article 50(5) requires the information to be clear, distinguishable and provided by the first interaction or exposure, with applicable accessibility requirements observed. [8]
For a voice service, an illustrative opening could be: “Hello, you’re speaking with an AI assistant for appointment enquiries.” For chat, make the identity of the assistant clear when the interaction begins. Test the notice in each supported channel and language, including the actual entry points customers use.
Map the other transparency duties separately. Article 50 distinguishes the provider's machine-readable marking duties for synthetic outputs from deployer disclosures for deepfakes and certain public-interest text. For that text, the human-review or editorial-control exemption also requires editorial responsibility. The Commission's FAQ explains these distinctions and their scope. [9]
A spoken introduction, a technical content mark and a publication label each address a different requirement. Record which are relevant to the service, what a supplier supplies and what the operating organisation must implement.
The Commission's transparency Code of Practice is voluntary and supports particular marking and labelling obligations. Those underlying obligations remain legal requirements when applicable. Use the code's defined scope when deciding what evidence it provides. [10]
Build evidence around the applicable requirements
For a high-risk system, determine the formal provider and deployer obligations and their relevant application dates. Provider obligations in Article 16 include the relevant system requirements, quality management, documentation, conformity assessment, declaration and marking arrangements. These need assessment for the particular system and role. [12]
Article 26 sets out distinct deployer duties, including use according to instructions, competent human oversight, monitoring and specified record-keeping. Plan those duties where they apply; avoid transferring the entire provider checklist to every business using an agent. [13]
For delivery teams, the following is a practical readiness pack. It supports review and implementation; the formal documentation and assessment requirements still need to be established for the deployment.
| Readiness area | Evidence to assemble | Decision it supports |
|---|---|---|
| Purpose and system boundary | Intended use, affected users, architecture and permitted actions | What is being assessed? |
| Roles and classification | Dated reasoning, supplier responsibilities and applicable provisions | Who needs to do what, and when? |
| Transparency | Voice/chat notices, output-marking arrangements and channel tests | Can people recognise the AI interaction and applicable synthetic content? |
| Human oversight | Handover rules, approval boundaries, operating instructions and pause procedure | Can an authorised person intervene effectively? |
| Quality and security | Representative evaluations, access tests, failure cases and recorded limitations | Does the configured service behave within its intended scope? |
| Data handling | Data flows, permissions, retention choices and privacy assessment records | Is the information handled appropriately? |
| Operation and change | Version history, incident routes, release decisions and review triggers | Can the organisation keep its assessment current? |
Use evidence from the configured environment. In an Azure or hybrid deployment, include the actual identity scopes, knowledge access, trace configuration, storage and business-system integrations. Test a failed verification, a denied tool action and a handover that someone actually receives.
For human oversight, identify the operator's authority and available controls. With the appointment assistant, rehearse a case where verification fails or a change has an uncertain outcome. Check that the person taking over can see the relevant information, pause the action and resolve or route the request.
Support the people who operate the agent
The amended Article 4 requires providers and deployers to support the development of AI literacy among staff and others operating or using systems on their behalf, taking account of their knowledge and the context of use. It also states that they are not required to guarantee a particular literacy level for each individual. [11]
Translate that into practical learning for the work. A service operator should understand the agent's permitted scope, known limitations, verification steps and escalation route. A person approving a business action needs to know what they are approving and how to check its result.
Keep records of the measures taken: relevant guidance, worked examples, practice sessions and updates when the service changes. Review whether people can use the controls in a realistic situation. Choose the approach for the roles and risks involved, and document the reasoning.
Keep data protection and impact assessments in scope
Map the information passing through the full voice or chat service: inputs, transcription, model calls, retrieval, tool results, memory, logs and storage. Identify the organisations processing it and document the relevant purposes, lawful basis, notices, access and retention arrangements.
The Commission's data-protection guidance explains information duties, security and data protection by design. It also identifies the GDPR trigger for a data protection impact assessment: processing likely to create a high risk to individuals' rights and freedoms. [14]
Assess that processing on its own facts. An AI-interaction notice does not determine whether recording or other personal-data processing is lawful, and an AI Act classification does not replace the data-protection assessment.
There is also a separate AI Act fundamental-rights impact assessment under Article 27 for specified deployers and high-risk uses. Its scope is not every organisation using an agent. Determine which assessment duties apply, how they relate and when they need to be completed. [15]
Reassess readiness as the service changes
Assign an owner for keeping the assessment and supporting evidence current. Screen changes to purpose, models, tools, knowledge, permissions and the people affected. Establish which changes need specialist reassessment and which can follow the existing release process.
For the appointment assistant, a new opening-hours source may need content validation. A move into evaluating people calls for a different review of classification and responsibilities. Record the reason for the decision and the evidence considered.
Keep a route for complaints, errors and incidents, with authority to pause affected functions. The operating team should know what information to preserve, whom to contact and who assesses any reporting duties. Update the tests and instructions when a failure reveals a gap.
A useful readiness review ends with named actions: an unclear role is resolved, a missing notice is implemented, a denied-access test is passed or a planned capability is held for further assessment. The service owner can then explain the intended use, the applicable requirements and the evidence supporting the next deployment decision.
Preparing an AI voice or chat service for your organisation? Explore Governance, Security & Reliability or discuss your use case with RZK. RZK provides technical and operational readiness support for internal governance and specialist legal review.
Sources
Official EU material consulted on 21 October 2026. The AI Act Service Desk reproduces the consolidated text as at 27 July 2026; the legally authentic instruments are published in the Official Journal. The workflow examples and readiness-pack table are practical recommendations.
- European Commission — AI Omnibus enters into force
- AI Act Service Desk — Article 113: Entry into force and application
- AI Act Service Desk — Article 3: Definitions
- AI Act Service Desk — Article 25: Responsibilities along the AI value chain
- AI Act Service Desk — Article 6: Classification rules for high-risk AI systems
- AI Act Service Desk — Annex III: High-risk uses
- AI Act Service Desk — Article 5: Prohibited AI practices
- AI Act Service Desk — Article 50: Transparency obligations
- European Commission — Transparency obligations under Article 50: FAQ
- European Commission — Code of Practice on Transparency of AI-generated Content
- AI Act Service Desk — Article 4: AI literacy
- AI Act Service Desk — Article 16: Obligations of high-risk system providers
- AI Act Service Desk — Article 26: Obligations of high-risk system deployers
- European Commission — Data-protection obligations
- AI Act Service Desk — Article 27: Fundamental-rights impact assessment