Why sentiment analysis fails
for financial services.

The word problem

Traditional sentiment analysis works by classifying words and phrases as positive, negative or neutral. It was designed for product reviews and social media — contexts where the words someone chooses are a reasonable proxy for how they feel.

In financial services, this assumption breaks down. A customer in genuine financial distress may speak calmly and politely throughout an entire collections call. A customer making a fraudulent insurance claim may use appropriately emotional language. The words are not the signal. The person is the signal.

What sentiment analysis misses

Consider a debt recovery call where the customer says "I understand" in a quiet, steady voice. Sentiment analysis scores this as neutral or mildly positive. But voice prosody analysis reveals: pitch has dropped 40% from baseline, speech rate has halved, micro-pauses have tripled in frequency. The customer is not calm. They are shutting down — a classic vulnerability indicator that keyword analysis will never detect.

FACS-based facial analysis adds another dimension. Involuntary Action Units — brow tension (AU1+4), lip press (AU24), chin raiser (AU17) — signal stress, suppression and cognitive load regardless of what the person is saying. These are physiological responses that cannot be consciously controlled.

The regulatory consequence

When the FCA reviews a complaint escalation, they ask: did the firm identify vulnerability? If the firm's compliance tools classified the interaction as "neutral" because the words were polite, the firm has a documentation gap that no amount of process can fill.

EchoDepth closes this gap by analysing the physiological signal beneath the words — across voice, video, image and text — producing structured VAD scores that document emotional state at every point in the interaction.