How One Fintech Cut Customer Response Time by 94% with AI Agents

In January 2025, a Nairobi-based mobile lending platform was drowning. Their support team of 12 agents was processing 3,400 tickets per week — mostly account queries, repayment confirmations, and loan status checks. Average response time had climbed to 4.2 hours. CSAT sat at 2.9 out of 5. Churn was accelerating.

Six months later, 94% of those queries were being handled automatically within 14 seconds. The 12-person team still existed — but they now handled only the escalations that genuinely required human judgment. CSAT climbed to 4.7. This article breaks down exactly how that transformation happened and what made it work.

The Anatomy of the Problem

Three query types made up 81% of all tickets: loan balance inquiries (34%), repayment confirmation requests (29%), and account suspension queries (18%). Every one of these had a defined resolution path — but they required accessing live account data, which the old rule-based chatbot couldn't do. The agents weren't slow because they were bad at their jobs. They were slow because each ticket required logging into three separate internal systems, manually pulling data, and composing a response from scratch. The bottleneck was system integration, not human effort.

Why They Chose AI Agents Over Hiring More Staff

Adding six more agents would cost approximately $84,000 per year in salaries — without addressing the underlying integration problem. The bottleneck would simply resurface at higher volumes. An AI agent connected to live banking APIs could handle the same queries without per-ticket labor cost and would scale horizontally without additional headcount.

Architecture: Three Layers of Knowledge

The Chatzuri deployment used three distinct knowledge layers. The first was a static knowledge base: FAQs, policy documents, product terms, and repayment schedules — uploaded as PDFs and indexed with chunked retrieval. The second was dynamic data access via API tool calls into the core banking system to pull real-time balance, repayment status, and account flags. The third was conversation context: the agent maintained session state so customers didn't have to repeat account details mid-conversation.

The static knowledge base alone would have failed. Most queries required live data. The combination of retrieval-augmented generation for policy questions and real-time API tool calls for account data was what made the agent genuinely useful — not just capable of responding.

What Went Wrong in Week One

Two problems surfaced immediately. The agent was occasionally retrieving outdated policy content when vector similarity matched an old chunk rather than the updated version. And for account suspension queries — which often carry emotional weight — the tone was too transactional. Customers felt dismissed.

The retrieval issue was fixed by adding metadata filters that prioritised recently-updated chunks and by stamping each policy document with a 'last updated' date that the retriever weighed. The tone problem required prompt revision: the agent was given explicit instructions to acknowledge the customer's frustration before moving to resolution. Small change, measurable CSAT lift.

The Numbers After 90 Days

What Generalises From This Case

The pattern here isn't unique to fintech or to East Africa. Any support operation where a defined set of query types makes up 70%+ of volume, and where resolution requires accessing live system data, is a strong candidate for AI agent deployment. The bottleneck is almost always integration design, not the AI model itself.

Start by categorising your last 500 tickets. If more than 60% share a resolution pattern, you have a deployable AI opportunity. The rest — fraud disputes, regulatory queries, complex escalations — will always benefit from human judgment. The goal is not to automate everything. It's to automate everything that doesn't require human judgment, so the humans can focus on the work that does.