Emotion AI for investment advisory: reading confidence and hesitation

Twistag built an emotion-aware advisory interface for NVISO, a Swiss FinTech serving wealth managers and private banks, producing a 35% improvement in client comprehension of complex investment products across a pilot at two Swiss banks. The interface reads gaze and micro-expression signals from NVISO's affective computing engine, detects confusion during product disclosures, and surfaces a cue to the human advisor — never a decision, and never to the client. Because Swiss financial regulation (FINMA) and data protection law (FADP) treat every emotion signal as personal data, Twistag designed consent-first: clients can see when sensing is active and pause it.

Personal wealth management is a conversation, not a transaction. A client's real comfort with a recommendation lives in hesitation — a pause before clicking, a gaze lingering on a risk disclosure, a micro-expression before verbal commitment. Human advisors read these signals intuitively; their software ignores them entirely. Complex products such as structured notes, multi-leg derivatives and ESG-weighted allocations need not just comprehension but confidence, and onscreen a client's comprehension stays invisible until they make a decision, by which point it is too late to adapt. Swiss regulation made this harder rather than easier: under FINMA and FADP, every emotion signal is personal data, so trust had to be architectural.

Twistag built an emotion-sensing UI layer that reads client behaviour in real time and adapts the advisory interface without replacing the advisor. It draws on NVISO's affective computing engine — micro-expression and gaze-tracking sensors detecting cognitive load, confidence shifts and hesitation — and Twistag mapped those signals to advisory workflows: confusion during a product disclosure, optimal moments for intervention, recommendations that may not have landed.

The critical architectural choice was consent-first. Rather than collect emotion data and ask permission later, Twistag inverted the flow: clients see a visual indicator when sensing is active, understand which signals are read (gaze and expression, not audio), and control when collection pauses. Signals surface only to the advisor, as a subtle cue rather than an alarm, and the advisor chooses the response. The machine never makes the recommendation.

React on the front end, Node.js orchestrating the signal pipeline, AWS hosting real-time processing. The framework is modular by design and capability-agnostic, so the emotion-responsive layer transfers to other regulated advisory workflows without architectural rework. The 35% improvement is a measurement of interface design working, not of model accuracy. Most firms selling AI would treat this engagement as an integration; the reason a number exists at all is that it was treated as a product problem inside a regulatory constraint. Under FINMA and FADP every emotion signal is personal data, so trust had to be architectural rather than a checkbox.

Stack: React, Node.js, AWS. Regulatory context: FINMA, FADP, consent-first data capture. Services: Product engineering, product design, AI integration.

User comprehension of complex investment structures improved by 35% in pilot testing across two Swiss banks. Product-to-client matching accuracy increased as advisors could see and respond to uncertainty in real time. The advisory relationship strengthened: clients were heard at the moment of hesitation rather than after a decision. The emotion-responsive layer now transfers to other regulated advisory workflows — insurance, pensions and credit decisions — without architectural rework. Advisors adopted it because the system made them better at their job rather than watched while they did it, and because the client could see at any moment that sensing was active and could stop it.

Advisors who trust the system deploy it; those who feel monitored hide it. Framing decides which.