Automating Customer Support for a Nigerian Fintech

The challenge

The support team was fast, but the queue was faster. Macros helped, but they could not reason across account history, product changes, or policy nuance. Leadership needed automation that felt like an extension of the team—not a brittle chatbot bolted onto the side.

Our approach

• Discovery & strategy — We mapped ticket categories, escalation paths, and the knowledge sources support engineers actually trusted.
• Architecture & design — We designed an agent workflow with retrieval over approved documentation, strict guardrails for money movement, and human handoff when confidence dropped.
• Engineering & integration — We implemented the agent service, logging, and review tooling, then integrated with the existing helpdesk so agents never switched context.
• Deployment & optimization — We launched behind feature flags, tuned retrieval with real ticket feedback, and iterated weekly on deflection quality.

The solution

A retrieval-augmented support agent with policy-aware escalation, structured audit logs, and a review queue for continuous learning—without exposing sensitive actions to unauthenticated automation.

The outcome

Deflection climbed steadily as coverage expanded. Escalated tickets arrived with richer context, so engineers resolved them faster. The system stayed under change management as policies evolved.

Key learnings

• Grounding beats flair — retrieval quality mattered more than model size.
• Handoff is a product surface — the transition to humans had to feel seamless.
• Measure what operators feel — time-to-first-useful-reply beat vanity automation rates.