Last reviewed: 28 September, 2026. Updated for the Digital Omnibus on AI (Regulation (EU) 2026/1744) and extended to cover AutoQA, agent and customer profiles, and EU database registration.

</aside>

About the EU AI Act

The EU AI Act was published in the Official Journal of the European Union on 12 July 2024. The AI Act is a European regulation on artificial intelligence (AI) that assigns applications of AI to risk categories:

Applications and systems that create an unacceptable risk, such as “government-run social scoring of the type used in China”, are banned.
High-risk applications, such as a “CV-scanning tool that ranks job applicants”, are subject to specific legal requirements.
Remaining applications are deemed low-risk, and they are largely left unregulated.

This page describes the impact of the AI Act for the EdgeTier Customer Insights and Automated QA system. The EdgeTier system includes sentiment analysis and quality assurance features to assist in assessing agent performance and compliance in customer service applications.

Therefore, two parts of the AI Act are relevant to EdgeTier, and this page examines each in turn:

Emotion detection: whether our agent empathy and customer emotion features could be deemed an “emotion-recognition system”, which is banned in the workplace under Article 5(1)(f).
AutoQA and agent/customer profiles: whether features that evaluate agent performance, or attach performance insights to natural persons by creating profiles/reports, are “high-risk” under Annex III and Article 6(2).

High-Level Summary

Question	Features in scope	Conclusion
1. Is it a prohibited emotion-recognition system?	Agent empathy detection, customer sentiment and emotion analysis	No. Limited-risk, with transparency obligations only. Our empathy analysis is not based on actual emotion - it’s based only on written language from transcripts. We evaluate whether the agents’ words align with empathetic customer service practices, not whether the agent actually felt empathy towards the customer. No biometric data is used, so the features fall outside the definition in Article 3(39) and the Article 5(1)(f) ban.
2. Is it a high-risk system because it can create profiles of agents and customers?	AutoQA, Coach, agent-level performance insights, customer-level signals, Ask Spotlight	We will register as high-risk as a precaution. Whether our insights “profile” natural persons, which would rule out the Article 6(3) exemption, is a grey area, so we will taking the precautionary route when it is time to register. Technically, it is feasible to ‘Ask Spotlight’ who is underperforming, and use the analysis as basis for coaching decisions that could affect agents’ careers.

EdgeTier will register as a high-risk provider in the EU database and meet the associated obligations by the 2 December 2027 deadline for Annex III systems, unless the deadline is moved again. Since (1) AI outputs are reviewed by humans and can never affect employment conditions without human review, and (2) we recommend that our customers disclose to agents that AI is being used to monitor their conversations, this registration may not be strictly necessary, but we will proceed anyway.

1. Emotion detection

1.1 What the AI Act prohibits

Definition of an “emotion-recognition system”

The AI Act only uses this term where an AI system “identifies or infers emotions or intentions … on the basis of their biometric data”. Recital 18 reiterates that emotion recognition means drawing inferences from physical or physiological signals (e.g. facial micro-expressions, voice tone, heart rate) that qualify as biometric data.
Workplace ban

Article 5(1)(f) prohibits deploying such biometric-based emotion-recognition systems “in the workplace or in educational institutions”, unless the purpose is medical or safety related.

1.2 How our emotion detection works

EdgeTier has two emotion-related features. Both analyse written language only: the text of chat and email messages and of call transcripts. No audio features (tone, pitch, volume, intonation or cadence), images, keystroke patterns or physiological signals are captured or used.