Case studies
Turkish conversational AI and document processing for a Turkish bank
For a bank whose assistant answered a narrow set of questions in stilted Turkish, a morphology-aware, grounded system improved self-service.
The challenge
The bank's assistant answered a narrow set in stilted formal Turkish and failed on anything phrased naturally. Underneath was a Turkish NLP problem: the intent model was built on tokenisation that could not handle agglutinated forms. The bank was clear that a wrong fee or rate answer was unacceptable.
What Corpshore did
We built two connected capabilities. The assistant is retrieval-grounded: every answer is grounded to the bank's content and constrained from answering outside it. We rebuilt the query understanding with morphology-aware processing and rewrote the source content into clear Turkish before indexing. The document capability extracts fields with confidence scoring.
Results
- Self-service containment rose substantially against the previous assistant
- Turkish query understanding accuracy improved markedly, largest on morphologically varied and colloquial queries
- Straight-through document processing reached 64 percent of volume
- Zero substantiated incidents of the assistant giving incorrect product or fee information in the first operational period
Why it worked
The constraint was the design. Grounding every answer and refusing to answer outside the corpus produced lower coverage than an unconstrained model would claim, and an assistant a regulated bank could actually put in front of customers.
Let us build your Türkiye team
Tell us your function, your scale and your language needs. We will come back to you within six hours.