AI Nudging Engine

How B-Accesible detects user difficulty in real time and offers contextual, non-manipulative guidance.

The AI nudging engine monitors behavioral signals on each screen and offers contextual help when a user appears to be struggling. Nudges are informational and non-intrusive — they are never used to push financial products or influence decisions.

How it works

1. Collect behavioral signals

   As the user interacts with a screen, the frontend tracks the following signals:

   Signal	Description
   screen	Screen identifier, e.g. send-money
   num_validation_errors	Number of form validation errors triggered
   time_on_screen_seconds	Total time spent on the current screen
   num_back_navigations	Number of times the user navigated backwards
   steps_total	Total steps in a multi-step flow
   current_step	Current step index in the flow

2. Send signals to the nudging API

   Signals are sent to the nudging inference endpoint:

   POST /ai/nudge
   

   {
     "screen": "send-money",
     "num_validation_errors": 3,
     "time_on_screen_seconds": 75,
     "num_back_navigations": 1,
     "steps_total": 4,
     "current_step": 2
   }
   
3. Receive nudge recommendation

   The model responds with a difficulty score and the recommended nudge type:

   {
     "screen": "send-money",
     "model_version": "nudging-v1.0",
     "result": {
       "needs_help": true,
       "difficulty_score": 0.82,
       "recommended_nudge_type": "assist",
       "reason": "high_error_rate_or_time"
     }
   }
   

   • needs_help — Whether a nudge should be shown.
   • difficulty_score — Continuous score from 0 to 1 indicating estimated difficulty level.
   • recommended_nudge_type — One of assist, info, or warning.
   • reason — Human-readable code explaining the trigger.
4. Display the nudge

   If needs_help is true, the frontend renders the appropriate nudge component. The user can dismiss the nudge at any time.

---

Nudge types

The nudge type is determined by the difficulty_score returned by the model.

Nudge type	Score threshold	Behavior
assist	> 0.75	Direct step-by-step help is offered to guide the user through the current action
info	> 0.4	An informational hint is shown, providing context about the current screen
warning	≤ 0.4	A soft security reminder is displayed, e.g. to verify recipient details

The assist nudge is the most active intervention. It presents the user with a guided walkthrough of the current screen, broken down into clear steps.

---

Fallback logic

When the AI model is unavailable or not loaded, the frontend applies a rule-based fallback to ensure users still receive help.

A nudge is triggered if either of the following conditions is met:

• num_validation_errors >= 2
• time_on_screen_seconds > 60

The fallback always triggers an assist-level nudge to err on the side of providing more help.

---

Model performance

The nudging model (nudging-v1.0) was trained on synthetic user interaction data.

Metric	Value
Accuracy	0.83
F1 score (macro avg)	0.80

---

Ethical principles

The nudging engine is designed around the following commitments:

User in control

Every nudge includes a dismiss option. The user is never forced to engage with a nudge.

Non-commercial

Nudging is strictly informational. It is never used to promote financial products or influence purchase decisions.

Transparent

Every nudge includes a brief, plain-language explanation of why it was triggered.

Respectful tone

Nudge copy avoids technical jargon and is written in a calm, supportive tone appropriate for all age groups.