AI Assistant Handles 20% of Queries and Saves £300K/Month for Fintech Platform

Context

A fast-growing fintech platform for entrepreneurs provided banking, accounting, and business management tools for business clients. The organization needed to scale client service capacity without proportional headcount growth while preserving the high Net Promoter Score (NPS) it was known for. The objective was to build an AI assistant as a daily business companion for entrepreneurs: one that could handle routine support queries, execute authorised banking operations, and deliver personalised financial expertise. Given the regulated financial context and the high-trust relationships with clients, the team treated the assistant as a voluntary alternative to human agents rather than a mandated replacement, and set out to validate technical viability, accuracy, and client willingness to adopt before committing to full-scale development.

Challenges

Key challenges were regulatory and behavioural. LLMs deliver the best answers when they have full conversational context, but that context in banking contains protected personal and financial data that cannot be passed to external models under strict financial services regulation. The platform’s internal knowledge base was fragmented and had never been designed for AI retrieval, making accurate, up-to-date responses impossible without a structural rebuild. Operationally, clients in this sector default strongly to human support and could not be forced away from agents without risking NPS and long-term customer relationships. Finally, the team needed to validate whether an AI assistant could deliver sufficient accuracy and earn client trust before investing in full development and scaling.

Implementation

The Fractional Head of AI led a hands-on validation and delivery programme. A working prototype was built in one week that answered questions across a single banking topic area with 87% accuracy. Two engineers integrated the prototype into the live product interface to produce a proof of concept that confirmed both technical viability and client willingness to engage with AI-assisted support. To address data protection, the team designed and built a proprietary data masking layer applied before any data reached the LLM; this covered 20+ PII categories, including data clients enter themselves mid-conversation, preserving model context while ensuring no personal or financial data was transmitted or stored. Recognising the fragmented knowledge base, the team created a proprietary machine-readable knowledge system composed of structured fragments optimised for LLM retrieval, and instituted a governance methodology that ensured every product or process change was reflected in the knowledge base within 24 hours to eliminate information drift. Quality and compliance were enforced via a three-stage automated validation pipeline assessing tone, answer relevance, and correct context usage. The assistant was positioned as an opt-in channel and surfaced through contextual UX entry points tied to specific client moments (for example, transaction issues or account setup steps), which encouraged voluntary adoption without forced routing.

Results

Within 12 months of launch the assistant achieved 60K monthly active users and 25% Day-28 retention, driven entirely by voluntary adoption. Twenty percent of all support queries were handled voluntarily by the AI in a sector where clients strongly default to human support. The assistant achieved an NPS of 8.7, outperforming the company average of 7.5, demonstrating that clients trusted the AI as an alternative channel. Operationally, the platform avoided £300K per month in hiring costs as the client base scaled, with no proportional increase in support headcount required. The solution also received external recognition, winning the International AI Championship in Abu Dhabi in 2025. The combined approach—rapid prototyping, strict data masking, a machine-readable knowledge layer with tight governance, and voluntary, context-driven adoption—validated that an AI assistant can deliver sufficient accuracy and trust in a regulated financial context while materially reducing support load and protecting client relationships.