The AI Act for marketplaces and local services

AI ACT · PLATFORMS AND RECOMMENDATIONS

In a marketplace, one engine may rank offers, detect fraud, answer questions and block an account. These functions do not share one purpose and should not be assessed as a single “AI algorithm”. Classification starts with the concrete effect on people.

Map decisions, not merely features

For each component record inputs, output, audience and subsequent action. A category recommender may merely order products; a fraud score may block payment; moderation may remove a provider; matching may decide who receives an opportunity. Consequence, not marketing name, drives assessment. ^{[EU-REG] [EU-RISK]}

Separate ranking, recommendation, semantic search, pricing, verification, moderation, fraud, support and profile generation in the register. Link each to an owner, vendor, data, metrics, override and appeal. This lets you change one decision’s control without declaring the entire marketplace “compliant” or “non-compliant”. ^{[EU-DESK] [EU-LIT]}

Ordinary recommendation is not automatically high-risk

The AI Act classifies high risk through Article 6 conditions and Annexes I and III in light of intended purpose. Recommending providers by distance, availability and preference does not become high-risk merely because it influences a commercial choice. Nor is “minimal risk” an official certificate. ^{[EU-REG] [EU-RISK]}

The analysis changes where a system enters recruitment, access to essential services, creditworthiness or another listed area and produces the relevant effect. A services marketplace should not be presumed high-risk, but check whether an apparently ordinary function is later used for an Annex-listed decision. ^{[EU-REG] [EU-RISK]}

Ranking needs explanation and control

Define accepted factors, relevant weights or rules, excluded data and success. Test new providers, small localities, languages, price ranges and groups that may be indirectly disadvantaged. Click metrics may favour sensational content or incumbents without improving real matching. ^{[EU-LIT] [EU-REG]}

Keep a non-personalised path, a reasonable explanation of factors and a way to correct profile data. Distinguish sponsored content from organic recommendation. Do not use AI to hide payment behind “relevance” users cannot understand. ^{[EU-REG] [EU-LIT]}

Moderation and fraud need human appeal

A model may prioritise cases, but account suspension or payment blocking should not be hidden inside an opaque score. Define thresholds, evidence, duration, approver and appeal. Review staff need signals and limits, not just a “fraud” verdict. ^{[EU-LIT] [EU-REG]}

Measure false positives, remediation time and differences between categories, not only fraud stopped. An aggressive system can look effective because it rejects many legitimate users. Keep rollback and limit automation until error and appeal are understood. ^{[EU-LIT] [EU-RISK]}

Avoid manipulation and disproportionate profiling

Article 5 prohibits certain practices, including forms of manipulation or exploitation of vulnerabilities causing significant harm and certain social scoring. Commission guidance provides interpretation and examples but does not replace the law. Do not optimise interfaces to pressure vulnerable people or hide alternatives. ^{[EU-REG] [EU-PROH]}

Do not build a general “trust” score from conversations, reviews, location and behaviour without limited purpose and validation. Use signals for the specific decision with controlled retention and access. Anti-fraud should not quietly become a general profile of a person’s worth. ^{[EU-PROH] [EDPB-28] [EU-REG]}

Prudent scenarios for AdverLink and CanUHelp APP

For AdverLink, a recommended model would explain publisher-campaign matching factors, separate sponsorship, enable profile correction and escalate blocks. For CanUHelp APP, local-service recommendation would keep clear criteria, a non-personalised option, chatbot disclosure and support handoff. ^{[EU-A50] [EU-LIT] [EU-REG]}

These scenarios do not mean the products are automatically high-risk or compliant. Final classification requires real functions, intended purpose, data and consequences. Where a platform enters an Annex III area or materially affects rights, reassess before release, not after complaints. ^{[EU-RISK] [EU-REG] [EU-DESK]}

Official sources and verification date

1. Regulation (EU) 2024/1689 — Artificial Intelligence Act
2. European Commission — AI Act Service Desk and Compliance Checker
3. European Commission — AI Literacy Questions & Answers
4. European Commission — transparency obligations under Article 50
5. European Commission — high-risk AI system classification
6. European Commission — guidelines on prohibited AI practices
7. European Data Protection Board — Opinion 28/2024 on AI models