Emotion Detection Software: 44 Action Units vs Generic AI

Enterprise interest in emotion detection software has grown significantly in the last three years, driven by expanding use cases in financial services, sales, HR, defence and healthcare. The market includes a wide range of platforms with very different technical foundations and very similar marketing language. The gap between “AI-powered emotion detection” on a vendor website and what the platform can actually do in a real-world deployment is often substantial.

This guide is intended for procurement teams, technology evaluators and business leaders considering emotion detection platforms. It covers the scientific foundations that separate valid from invalid claims, the accuracy questions to ask, the governance requirements you will need to satisfy, and the deployment considerations that determine whether a platform actually works in your environment.

The FACS question — ask it first

The most important distinction in emotion detection software is whether the platform is built on Facial Action Coding System (FACS) methodology or on generic image classification. FACS maps 44 specific facial muscle movements — Action Units — to emotional states, based on decades of peer-reviewed research by Paul Ekman and Wallace Friesen at UCSF. It has validated reliability coefficients, published cultural calibration data and a scientific record that spans 40 years.

Generic image classification systems assign emotion labels to faces using machine learning models trained on labelled datasets — “happy face,” “angry face,” “sad face.” These systems lack the anatomical grounding that makes FACS defensible. They are typically trained on dataset populations that skew toward Western, English-speaking demographics. They produce outputs that are less accurate, less consistent and less culturally calibrated.

The question to ask every vendor: is your emotion classification based on Action Unit detection from FACS-coded training data, or on a classification model trained on labelled emotional expression datasets? The answer will tell you more about the platform than any marketing material.

Evaluating accuracy claims

Every emotion detection platform makes accuracy claims. Almost all of them measure accuracy on controlled laboratory datasets — clean, well-lit, cooperative subjects in academic settings. The gap between lab accuracy and real-world deployment accuracy is substantial and vendor-controlled. Ask specifically for accuracy figures from real-world deployment in environments similar to yours.

Governance requirements you cannot skip

Facial expression analysis constitutes biometric data processing under UK GDPR Article 9 — special category data. Any enterprise deployment requires: explicit informed consent from all data subjects; a documented purpose of processing statement; a completed Data Protection Impact Assessment; a signed Data Processing Agreement with the vendor; and a data retention schedule. These are not optional additions — they are legal requirements. Any vendor that does not provide full governance documentation as standard should not be on your shortlist.

What accuracy claims actually mean

Most emotion detection vendors publish accuracy figures. "94% accuracy" appears in sales materials without context. Accuracy against what? Measured on which dataset? Validated in which cultural context? The number is not meaningless — but it is rarely interpretable without knowing what it is measuring.

The most important distinction is between recognition accuracy (can the system identify that a facial movement occurred?) and classification accuracy (can the system correctly label the emotional state that movement represents?). Recognition accuracy can be very high — 90%+ — while classification accuracy varies significantly depending on the cultural calibration of the model, the quality of the training data, and the context of deployment.

EchoDepth uses the FACS standard — measuring 44 specific Action Units rather than attempting direct emotion classification. This approach is more conservative and more reliable: it reports what the facial muscles did, not what emotion that action represents, and leaves interpretation to context-aware human review. For enterprise deployment in regulated environments, this is the correct architectural choice.

The enterprise vendor landscape, compared

Accuracy has become close to a commodity across the serious platforms. The differences that decide a procurement are modality coverage, where inference runs, what the measurement is grounded in, and whether the compliance position survives legal review. The table below compares the platforms most often shortlisted for enterprise deployment in 2026.

Compiled September 2026 from vendor documentation and public product material. Capabilities change; verify current specifications directly with each vendor before procurement.

Two market facts are worth knowing before shortlisting. Microsoft deprecated emotion inference from the Azure Face API, so a general-purpose cloud emotion endpoint is no longer a safe long-term dependency. And the research consensus has moved against reporting Ekman's six basic emotions as universal — cross-cultural validity does not hold — which is why Action-Unit reporting and arousal/valence framing have replaced direct emotion labels in defensible deployments.

Where each platform is lawful under the EU AI Act

This question now precedes the technical evaluation for any EU deployment, and it applies to the use case rather than the vendor. EU AI Act Article 5(1)(f) prohibits emotion recognition systems in workplaces and educational institutions, except where deployed for medical or safety purposes. It does not impose a blanket ban: using these systems to assess customer emotion — a contact centre caller's rising frustration, for example — is not prohibited.

The practical consequence is that a platform can be entirely suitable for contact centre, market research or investor communication work and unlawful for recruitment screening in the same organisation. Scope the use case first, then the vendor. Any vendor unable to state its position on Article 5(1)(f) in writing has not done the work.

Separately, under UK and EU GDPR, facial analysis that uniquely identifies an individual is special category biometric data under Article 9. Analysis that measures observable delivery signals without identification sits differently, and the distinction is worth establishing with counsel before deployment rather than after.

The six questions to ask before a procurement decision

When evaluating emotion detection software for enterprise deployment, six questions should form the basis of the technical due diligence: (1) What standard does the system use — FACS-grounded AUs, or proprietary emotion categories? (2) How many cultural cohorts is the model calibrated against? (3) What consent architecture is required and how is it implemented? (4) Is biometric data retained beyond the processing window? (5) What governance documentation is provided for regulated sectors? (6) Can the system provide a sample analysis of content from your specific use case before commercial commitment?

A vendor who cannot answer questions 1–5 clearly should not progress to procurement. A vendor who declines question 6 — a free sample analysis on your own content — is not confident that their platform performs in your context. EchoDepth provides a free sample analysis as standard: submit content, receive a full scored report within 5 working days, before any commercial conversation begins.

Frequently Asked Questions