Data Department builds generative artificial intelligence (AI) systems that work on your company's own documents and data: RAG (retrieval-augmented generation) assistants, NL2SQL (natural-language questions answered from your databases), document AI for certificates, delivery notes and supplier offers, and agents built on LLMs (large language models) with scoped permissions and human approval. Every system ships with an evaluation set, so quality is measured, not assumed.
What problems does generative AI solve?
Generative AI is useful where people spend time finding, reading or retyping information. The problems we work on:
- Finding the right answer in your own documents. Procedures, technical specifications, contracts and quality manuals, with the source shown next to each answer.
- Asking the database a question in plain language. "What were sales of product X in the Aegean region last quarter?" without waiting for a report.
- Reading documents that arrive every day. Extracting fields from material certificates, delivery notes and supplier offers, checking them against orders, and putting supplier offers side by side.
- Multi-step tasks across systems. An agent checks stock, drafts a purchase order and waits for a person to approve it.
- Routing and drafting. Classifying incoming requests and drafting replies for a person to review.
What a project looks like
Discovery (2 weeks). We pick one use case with a clear owner. We collect real questions or documents from the people who will use the system, with the correct answers, and turn them into an evaluation set. We check document quality, access permissions and where the data lives. You get a go/no-go decision.
Pilot (6–10 weeks). We build the system and measure it on the evaluation set: is the answer correct, is the cited source the right one, and does the system say "I don't know" when the answer is not in the documents? Then a small group of real users works with it, and their feedback becomes new test cases.
Production and handover. Access control follows your existing document and database permissions, so nobody sees through the assistant what they could not see before. We add logging, quality monitoring and usage cost tracking, and train your team to maintain the evaluation set. Models can be hosted APIs (application programming interfaces) or open-weight models running on your own infrastructure.
Which generative AI pattern fits which job?
| Pattern | Good for | Watch out for |
|---|---|---|
| RAG assistant | Answers from your documents, with citations | Outdated document versions, permissions, poor scans |
| NL2SQL | Ad-hoc questions to databases | Ambiguous business definitions; we use read-only access and check every query |
| Document AI extraction | Structured fields from certificates, delivery notes, offers | Handwriting and low-quality scans; low-confidence fields go to a person |
| Agent with tools | Multi-step tasks across systems | Scoped permissions, approval before any write action, full audit log |
| Fine-tuning a model | A fixed format or style at large volume | Rarely the right first step; try retrieval and better prompts first |
How we keep agents safe
An agent that can act in your systems needs the same controls as a new employee. Each agent gets only the permissions its task needs. Any action that changes data, sends a message or commits money waits for human approval. Every step is logged, so you can see what the agent read, what it decided and why.
Proof and experience
We do not yet publish a numeric generative AI case study, and we will not invent one. What we can show is how we measure: an evaluation set built from your questions, scored before and after every change. For a deeper look at one common use case, read generative AI for documents in manufacturing. If you already have an assistant and are unsure how good it is, a RAG audit from our AI strategy services is a good starting point.
When is generative AI NOT the right tool?
- The answer must be exact and calculated. Pricing rules, tax calculations and stock totals belong in code or SQL (structured query language), not in a language model.
- The question set is small and fixed. A good search page or a list of frequently asked questions may do the job.
- The source documents are wrong or outdated. The assistant will repeat those errors with confidence. Clean the source first.
- Decisions carry legal liability. Use the model to draft and a person to decide.
- "We want a chatbot" is the whole brief. Without a defined user and question, there is nothing to measure.
Frequently asked questions
How do you stop the assistant from making things up?
It answers only from retrieved sources, shows its citations and is designed to say it does not know. We measure this on the evaluation set before release and monitor it after.
Will our data be used to train someone else's model?
We choose deployment options where your data is not used for model training. If needed, open-weight models run entirely on your infrastructure.
What about KVKK and GDPR?
If documents contain personal data, sending them to a model hosted abroad is a cross-border transfer under KVKK (Türkiye's Personal Data Protection Law), and GDPR (General Data Protection Regulation) has its own rules. We handle this with masking, regional hosting or on-premises models.
Can it connect to our file servers, SharePoint and ERP?
Yes. We connect to file shares, document management systems, email inboxes and ERP (enterprise resource planning) databases, with read-only access unless a write action is explicitly approved.
What drives the cost of a generative AI project?
The volume and quality of documents, the number of source systems, the number of users and queries, the choice of model and hosting, and how complex your permission rules are.