Modena Estonia OÜ is a credit and payment institution operating under a license (4.1-1/100) issued by the Estonian Financial Supervision Authority. The company provides flexible payment and financing solutions to both consumers and merchants. Modena’s product portfolio includes a “Pay in 3” zero-interest installment solution, instalment plans of up to 48 months, consumer loans, and Open Banking-based bank payments. The business is split across two portals — modena.ee (consumer-facing lending and payment services) and modena.capital (investor portal).
As a licensed financial institution, Modena operates in an environment where every customer communication, response, and escalation must be auditable, regulation-compliant, and consistently high-quality. The customer support team handles inbound emails in three languages (Estonian, English, Russian) and serves two distinct audiences with very different needs and regulatory frameworks — borrowers and investors.
Solution
- SRINI built Modena Helpdesk — an AI-native customer support platform where artificial intelligence is not a separate “chatbot on the side” but is deeply integrated into 11 distinct business workflows. The core design principle was clear: AI gives agents superpowers, but humans always remain the final decision-makers — every AI-generated output (a reply, a category, a FAQ draft, a trend report) goes through human review before it reaches a customer or the database.
- The platform is built on a deliberately lean and controllable tech stack — a PHP 8 monolith with a MySQL database, vanilla JavaScript on the front end, and cron jobs for background tasks — and it uses Anthropic’s Claude Messages API across three model tiers (Haiku, Sonnet, Opus). Each AI scenario uses the model that is optimal for cost-efficiency: triage and classification (high volume, simple task) run on the cheaper Haiku; reply drafting and contextual reasoning use Sonnet; Opus is reserved for the most complex cases.
- Integration with Modena’s existing infrastructure — Microsoft 365 mail via Microsoft Graph, Azure AD SSO, and Zendesk archive sync — was delivered in a way that didn’t disrupt operations for a single day.
- AI
- Backend
- PHP
- Self-Service
- UX
- gain (5 agents): ~€27,420
- Annual direct financial
- €20 280 / year
- Labor cost savings
- Zendesk Suite + Advanced AI: ~€7 140 / year
- Direct software savings vs.
- ~27.5h per week 12 h per month ≈ 0.7 FTE
- Time saved by AI features:
Standout features
- Per-ticket AI chat box. Every ticket has a side-panel chat where agents can talk to Claude in the context of that specific ticket. At the start of each conversation, the AI is given a dynamic context block — ticket metadata, the last 20 comments, three similar previously resolved Zendesk tickets, and portal-specific knowledge base articles. A classic RAG pattern, without a vector store.
- Automatic email triage. Every inbound email is classified in the background: priority, category, language, a 120-character summary, and an escalation flag. Angry customers automatically rise to the top of the queue.
- AI-generated reply drafts. The agent clicks “Generate reply” and gets a complete email draft in the customer’s language, in the right tone, grounded in Modena’s knowledge base. The draft is never sent automatically — the agent reviews and sends it.
- 🧙 Personal prompt wizard. Each agent goes through a 5–7 question conversation, after which the AI builds them a personalized response prompt — asking about their role, areas of responsibility, tone, favorite phrases, and words to avoid. Users don’t need to learn “prompt engineering.”
- Similar tickets RAG. From the Zendesk archive of resolved tickets, the system surfaces the three most similar cases for every new query and feeds them as context. The AI gets to see how a human handled the same kind of question before.
- Portal-specific knowledge base. Every KB article is tagged with a portal (modena.ee OR modena.capital) — an investor ticket never sees a lending FAQ and vice versa. This solves the classic AI failure mode of “saying correct things, but for the wrong context.”
- Weekly trend analysis. Every Monday morning, Claude emails the admin team a summary of patterns from the past seven days — recurring issues, escalated topics, possible technical glitches. Operational intelligence without an analyst on payroll.
- Self-extending knowledge base. On Sunday evenings, Claude reviews comments from resolved tickets and drafts the three most frequently recurring FAQs. The admin gets a notification and clicks “publish” after review. The KB grows automatically based on what agents are actually answering.
- Cost-aware model selection. Admins can see and override the model used in each AI scenario through the UI. Defaults are set per use case — this is a genuinely dollar-cost-aware AI architecture.
- Audit transparency. Every AI action is logged: who, when, what context, what result. Mandatory in financial services, useful everywhere else.
- Graceful degradation. If the Claude API is down, the platform keeps running without it — categories stay empty, triage returns defaults, agents reply manually. AI is an enhancement, not a critical path
Results
Based on a conservative scenario (a 5-person support team handling ~80 emails and ~50 tickets per day):
– Quality improvements that are measurable but not converted to euros:
– Response time dropped from 5–10 minutes to 1–2 minutes (review-and-send time)
– Escalations now reach the right person on the same day, not from the end of the queue
– Reply quality is more consistent across agents — everyone draws from the same KB and prior resolutions
– New hires can use the platform on day one without any “prompt engineering” training
Key takeaway
Modena Helpdesk shows that an AI-native support operation doesn’t require an “AI startup” budget — what it requires is careful integration into existing business workflows, a cost-conscious model strategy, and the discipline to keep humans in the loop. The result is a platform where AI is an invisible layer working everywhere on agents’ behalf, but never taking the final decision out of human hands.
The same approach is transferable to any organization that handles a meaningful volume of customer support, sales, or internal request flows — especially in regulated sectors (fintech, healthcare, public sector) where audit transparency and human oversight are mandatory.
Want a similar solution for your business? Let’s talk about how SRINI can build an AI-native platform tailored to your workflows.
[Get in touch →] rasmus@srini.ee
SRINI OÜ · Software development and IT modernization · Tallinn, Estonia
