A rapidly growing U.S.-based digital bank with a global customer base but no physical branches. All customer interactions are conducted through digital channels, including mobile apps, websites, email, and phone. With over $1 billion in assets under management, delivering seamless, responsive support at scale was critical to maintaining customer satisfaction and trust.
1:1 Human Interaction Is Expensive
As the bank expanded globally, it faced a major challenge: its customer service operations could not scale quickly enough using human agents alone. With rising volumes of inquiries—averaging thousands of chats daily—it became clear that:
- Hiring human agents was cost-prohibitive: U.S. salaries for customer support roles can range from $40,000–$50,000 annually or $28–$38/hour in call centers.
- Customers required 24/7 multilingual support (English, Spanish, Mandarin, etc.), which was difficult to staff.
- There was a growing “no-reply” rate to service requests.
The bank needed to scale its support team across digital platforms—without relying solely on human agents.
AI-Powered Support Can Manage Most Queries
Factored implemented a three-pillar AI-powered system that combined advanced infrastructure, software engineering, and applied AI/ML to automate customer interactions efficiently and securely.
Pillar 1: Infrastructure
- Kubernetes-based architecture with two environments (staging and production) enabled fast iteration and high system reliability.
- Enabled full automation pipelines with A/B testing for new models and deployment strategies.
Pillar 2: Frontend & Backend Engineering
- Real-time communication enabled via real-time APIs.
- Built using Python and high performance APIs, ensuring scalability and rapid deployment of updates.
- Integrated with existing digital banking APIs to access customer profiles, loan status, payment history, etc., in real-time.
Pillar 3: AI & Automation
- Deployed a multimodal LLM architecture using APIs from 3 industry leading models.
- One of the models were for image-to-text parsing to optimize accuracy based on context:
- Document validation.
- Payment collections.
- Escalation routing to agents.
- Document validation.
AI Chatbot Capabilities
- Intent detection: Identifies the primary intent (e.g., loan disbursement, payment delays, etc.)
- Function calling: Handles critical branching decisions—e.g., automating minor issues or escalating users with critical needs to 1:1 human support.
- Language understanding: Processes multilingual input.
- Structured validation: Uses a second LLM call to verify model outputs and prevent hallucinations.
- Agentic architecture: Each "sub-agent" handles a specific intent and performs actions such as escalating to a human, closing the chat, or redirecting to a specialized agent.
Evaluation & Testing:
- We created mock conversations and replayed past transcripts in staging to evaluate new models.
- Compared responses using:
- Baseline (long prompt with manual logic).
- Agentic approach (LLM chooses from small prompts/sub-agents).
- A/B testing conducted to optimize response accuracy and automation rate.
+60% Total Reduction 1:1 Human Interactions
- Achieved 70% automation on text-based chats, and 30% automation on chats with image—critical in reducing overhead and wait times.
- Maintained compliance and security standards by avoiding sensitive data exposure to LLMs (all personal information pulled through secure APIs).
- Decreased customer service costs significantly by reducing the need for new hires and minimizing errors in multilingual settings.
Through a robust AI-driven approach, the U.S.-based digital bank successfully scaled its global customer support operations—achieving high automation rates, reducing operational costs, and delivering faster, round-the-clock service to users worldwide. Factored's three-pillar solution ensured rapid iteration, accuracy, and adaptability in a highly regulated financial environment.
