Start with the directive's definition

The directive does not cover every workplace that uses an algorithm. A digital labour platform, as defined in the text, has four features. It provides a service at least partly at a distance through electronic means. The service responds to a recipient's request, organizes paid work as a necessary and essential component, and uses automated monitoring or automated decision-making systems.

The rules apply when the platform work is performed in the EU, regardless of where the platform is established. Personal-data protections in the algorithmic-management chapter also reach people performing platform work without an employment contract. Other provisions, including parts of the health-and-safety and worker-consultation regime, are specific to platform workers. Article 11's human-review right has an exception for people who qualify as business users under Regulation (EU) 2019/1150.

Legal teams need to resolve that scope in each market. Engineers can begin by listing systems before debating their labels. Put software that allocates jobs, sets prices, scores performance, controls access to work, or affects an account on the map.

Six data practices have to stop

Article 7 prohibits digital labour platforms from using automated monitoring or decision systems to process several kinds of personal data:

• emotional or psychological state;
• private conversations, including exchanges with other workers or their representatives;
• data collected while a person is not offering or performing platform work;
• data used to predict the exercise of fundamental rights, including association and collective bargaining;
• data used to infer protected or highly sensitive traits, including racial or ethnic origin, migration status, political opinions, religion, disability, health, union membership, sex life, or sexual orientation; and
• biometric data used to establish identity by comparing it with stored biometric data in a database.

These restrictions apply from the start of recruitment or selection. They also cover automated systems that take or support decisions affecting a person performing platform work in any way, not only products marketed as AI.

A blocklist in a model prompt does not solve the problem. The restriction has to reach event collection, analytics, feature stores, embeddings, training data, vendor payloads, logs, and support tools. A prohibited inference can reappear downstream even when the original field never reaches a dashboard.

A generic notice cannot describe a decision system

Article 9 requires disclosure of automated monitoring and decision systems. For monitoring, that includes the data and actions observed, the purpose and method, and any recipients or transfers. For decision systems, it includes the categories of decisions, the data and main parameters used, their relative importance, how a person's data or behavior affects decisions, and the grounds for adverse account, payment, or contractual decisions.

The information must be written, accessible, intelligible, and in plain language. A person performing platform work must receive concise information about systems that directly affect them by the first working day, before relevant changes, and whenever they request it. Recruitment systems require notice before the procedure begins.

Maintain a versioned decision inventory with the system version, policy version, input categories, main parameters, downstream effects, and owner. Connect it to deployment controls so a material model or policy change cannot reach production while the notice still describes the previous system.

Human review requires authority and evidence

The directive requires an impact evaluation of individual decisions at least every two years, with worker representatives involved. Oversight staff must have the competence, training, and authority to override automated decisions. They are protected against dismissal, discipline, or adverse treatment for doing that work.

For account restriction, suspension, termination, or an equivalent detriment, the human is the decision-maker. People performing platform work can request an oral or written explanation for decisions taken or supported by an automated system. Certain adverse decisions require written reasons no later than the date they take effect. If a decision infringes a person's rights, the platform must rectify it without delay and within two weeks of the decision's adoption.

The product needs a reviewable state machine:

1. The automated system flags or recommends an action.
2. The case enters a review queue with evidence, model and policy versions, and a readable reason trace.
3. A qualified reviewer approves, rejects, or changes the recommendation.
4. The system records the human decision and reasons, then sends the required notice.
5. A separate path supports explanation, reconsideration, correction, and audit.

The observability used to explain agent failures helps reconstruct the path, but worker decisions need stricter access and retention controls. If models are routed through a shared endpoint such as api.ish.chat, log the route that handled the case and keep personal data inside the approved processing boundary. Use synthetic cases for comparative testing instead of live worker records.

Test one adverse action before December

Choose a real pathway, such as a fraud flag that can restrict job access. Follow it from raw event to final notification:

• Did any prohibited data or inference enter the pipeline?
• Can the team identify every automated system that influenced the outcome?
• Does the explanation match the deployed model and policy versions?
• Can a qualified reviewer inspect the evidence and override the recommendation before the action takes effect?
• Can the platform produce the decision record, explanation, reconsideration, and correction history?
• Are worker representatives involved in the impact assessment where required?

This test also separates the Platform Work Directive from the EU AI Act. The AI Act classifies and regulates certain AI systems. The platform-work rules cover automated monitoring and decision systems more broadly, including software that only supports a decision. Our Article 50 disclosure guide addresses another transparency obligation.

The European Commission says the directive entered into force in December 2024. It describes the algorithmic-management rules as helping people understand task allocation and pricing and contest decisions. The Council's adoption notice also points to qualified oversight and the right to contest automated decisions.

National rules will not be identical after transposition. Before deployment, a platform should be able to reproduce the evidence, system versions, human authorization, notice, and review history for each adverse account action. If those records do not exist, neither a worker nor an auditor can verify who made the decision.

Primary sources

• Directive (EU) 2024/2831 on improving working conditions in platform work
• European Commission: The future of work
• European Commission staff report on transposition of Directive (EU) 2024/2831
• Council of the EU: adoption of the platform-work rules
• ILO: Algorithmic management of work and its implications