Challenge
At a licensed credit institution, every reply sent to a customer has to be auditable and compliant with regulations. The team had to juggle three languages and two portals (modena.ee for loans and modena.capital for investors) at the same time, under constant pressure to keep replies fast and quality consistent with a small team.
Solution
- 11 AI workflows, one platform. AI is woven into 11 different customer support workflows rather than bolted on as a separate chatbot; every AI output is reviewed by a human before it reaches the customer.
- Cost-efficient three-model strategy. Triage runs on Haiku, drafting replies on Sonnet and complex reasoning on Opus – the administrator can see the choice and override it per scenario.
- RAG over the Zendesk archive. Each new ticket is given the three most similar resolved tickets as context, so the AI learns from how people actually resolved the same question.
- Portal-aware knowledge base. Every knowledge base article is tagged by portal – an investor ticket never sees a lending FAQ, and vice versa.
- Self-improving knowledge base. Every Sunday, Claude reviews the resolved tickets and drafts the three most frequent FAQs; after review, the administrator clicks “publish”.
- Graceful degradation. If the Claude API is down, the platform keeps working – AI is an enhancement, not a critical dependency.
Results
- Response time dropped from 5–10 minutes to 1–2 minutes per reply.
- Escalated cases reach the right person the same day, not from the back of the queue.
- Reply quality is more consistent across agents – the same knowledge base, the same past resolutions.
- New staff use the platform from day one – no separate “prompt engineering” training needed.
- Direct software savings of approximately €7,140 per year compared with Zendesk Suite and Advanced AI.
Key takeawayAI-native customer support doesn’t need an AI-startup budget – it needs careful workflow integration, a cost-efficient model strategy and a human in every decision.