01. Context
The Challenge
These concerns increased turnover, hindered onboarding, and raised customer support expenditures. The client's goal was clear: lower ticket volume and response time without adding manpower, while improving user satisfaction and onboarding efficiency. A mid-sized FinTech company that offered B2B payments and reconciliation technologies was suffering with rapid customer growth and an increasing support burden. Their application was used by accountants, financial teams, and small-to-medium businesses to confirm payments in a timely manner, handle disputes, and troubleshoot integration issues. As the user base tripled in 12 months, assistance volume outpaced employment capacity. Typical pain points were: - High ticket traffic for routine, repeating queries such transaction status, reconciliation discrepancies, invoice matching logic, and API interface issues. - Long first-response times (12-24 hours) and average time to resolution measured in days for non-urgent issues. - Inefficient use of senior engineers' time on low-value support activities, which raises operating costs. - Inadequate self-service coverage: knowledge base pages were outdated, difficult to search, and not optimised for how users stated their questions. - Friction during onboarding: new clients encountered frequent setup issues that necessitated human intervention.
02. Solution
How We Solved It
We created and implemented a custom AI chatbot that seamlessly integrated with the client's product, knowledge assets, and support workflows. The project was carried out in four phases: discovery, build, pilot, and scale. 1. Discovery and Alignment - User and ticket analysis: We examined six months' worth of support tickets to determine top categories, typical phrasing, peak hours, and ticket funnels. The top ten categories accounted for 68% of all tickets, with the majority being predictable and scriptable. - Stakeholder workshops: We held workshops with product, engineering, support, and compliance teams to sketch out acceptable automated actions, escalation rules, data protection limits, and success metrics (ticket deflection, response time, CSAT). - Knowledge base audit: We analysed existing help articles, API documentation, and error logs to identify and prioritise content gaps. 2. Conversational Design and Architecture - Intent taxonomy: We created a comprehensive intent taxonomy for common queries (e.g., "transaction status," "reconciliation mismatch," "API webhook setup," "reset integration keys," "billing query"). - Hybrid approach: We used retrieval-augmented generation (RAG) to find certain KB articles and a rule-based flows engine for sensitive activities (such as changing API keys or starting refunds). RAG enabled the model to mention certain documentation snippets and connect to source articles. - Context and authentication: The chatbot authenticated users using single sign-on (SSO) and obtained context (account ID, recent transactions, integration status) via secure API calls, allowing replies to contain personalised information without exposing sensitive data. - Escalation rules: If the confidence score was < 0.6, the bot provided a clear escalation path to a human agent, including the chat transcript, tags, and recommended resolution procedures. 3. Building and integration - Knowledge engineering: We streamlined and rearranged KB information into short, user-friendly snippets along with examples and step-by-step directions. Each sample was indexed for synonyms, related issues, and troubleshooting actions. - Platform selection: We used a fine-tuned LLM for generating replies to non-sensitive queries, combined with a vector store for quick retrieval of KB passages. A microservice handled secure API requests to retrieve user-specific information. - Actionable flows: For procedural chores (e.g., "resend webhook test," "regenerate API key"), we used rule-driven flows that needed explicit user confirmation and recorded actions in audit trails. - UX and channel rollout: We released the bot through the web app's chat widget and intercom-style help channels. To lessen cognitive strain, the UI prioritised recommended replies, quick links to knowledge base articles, and "Show steps" toggles. 4. Pilot and iteration. - Controlled pilot: We began with 15% of incoming traffic (new users and basic-tier clients) and tracked intent recognition, deflection rate, escalation quality, and CSAT. - Rapid improvement: We employed active learning for two weeks to examine low-confidence or misrouted interactions, retrain the intent classifiers, and add new cases to the KB index. - Compliance review and security testing: Because this was a FinTech setting, we conducted privacy impact assessments and penetration testing of the bot's integrations and logs.
03. Impact
Results
Within 90 days of full-scale deployment, the FinTech customer saw substantial improvements in support, onboarding, and cost metrics. Key outcomes: - Ticket deflection: Overall support tickets declined by 67% compared to the previous quarter. The chatbot answered a large number of routine questions, including those about transaction status, reconciliation help, and basic API connectivity. - Faster response and resolution: For chat engagements, the average first response time was reduced from 12-24 hours to less than 2 minutes. Mean time to resolution for automated or bot-assisted tickets decreased by 78%. - Increased agent efficiency: Human agents reported a 55% reduction in repetitive tasks, freeing them to focus on complicated escalations and product improvements. Despite a threefold increase in user numbers, the customer was able to avoid recruiting more younger agents. - Increased onboarding velocity: Automated onboarding flows and in-chat guided configuration cut onboarding time for new customers by roughly 40%, cutting time-to-first-payment and speeding up value realisation. - Customer satisfaction: CSAT for support interactions using the bot increased from 82% to 91%. Qualitative feedback indicated a preference for quicker responses, step-by-step instructions, and context-aware solutions. - Cost impact: Support cost savings (headcount and handling time) are estimated to be 38% each year. Engineers and product teams saved time, which resulted in faster product iterations and problem solutions, indirectly reducing future support demand. Why it worked. - Data-driven prioritisation: Focusing on the most common ticket types resulted in a quick impact without over-automating dangerous areas. - Hybrid architecture: Combining retrieval of exact documentation with controlled action sequences ensured that replies were accurate and auditable, which is crucial in FinTech. - Context-aware responses: Using account-specific data reduced back-and-forth queries while making answers actionable. - Continuous improvement: Active learning loops and regular KB updates kept the bot updated as product and error patterns changed. - Clear escalation rules: Providing a clear access to human assistance maintained trust in complex or sensitive circumstances.
Ready to Build Together?
Let's discuss how we can deliver similar impact for your business.