Kiwi.com's legal team was sitting on more than 20,000 documents they couldn't really search. We built them an assistant that reads the entire library and answers questions in plain language — with every answer tracing back to the exact document it came from.
Where things stood
Kiwi.com's legal team had built up a library of more than 20,000 documents — contracts, NDAs, data processing agreements, policies, and filings. Everything was organized and searchable by metadata: type, counterparty, date. For everyday filing, that worked fine. The trouble started the moment someone needed an answer that lived inside the documents:
⏳ Questions that used to take all day
Answering even one meant opening documents one by one and reading them by hand. Across a library this size, that friction showed up in four ways:
Finding specific clauses or provisions across thousands of documents required hours of manual search — often with incomplete results.
Legal analysis that should take minutes consumed entire days. Cross-referencing obligations across vendor agreements was practically impossible at scale.
Critical knowledge was locked inside documents nobody remembered existed. Onboarding new team members meant weeks of archaeology.
Repetitive review tasks consumed senior legal resources. External counsel was engaged for questions the team could have answered internally — if they could find the right document.
What we built
The idea was simple: take the whole document library, have the AI read every document, and let the legal team ask questions in plain language — with every answer pointing back to the exact document and clause it came from. Making that genuinely useful for a legal team — where being precise matters more than sounding clever — came down to four things.
Every file — PDFs, Word documents, even scans — goes through a pipeline that reads it and breaks it into meaningful sections. Not arbitrary slices: it respects clause structures, paragraphs, and tables, so a clause stays a clause. That's what lets the AI understand what a section is actually about, not just which words it contains.
Ordinary search finds documents that contain certain words. This finds documents that are about a topic — even when they use different words for it. Ask about "limitation of liability" and it also surfaces clauses titled "cap on damages" or "maximum aggregate liability", because it knows they mean the same thing.
In legal work, an answer you can't verify is worthless. The system was built to cite its sources as a hard rule, not a nice-to-have. Every statement links back to the specific document and clause it came from — and if the answer isn't in the documents, it says so rather than guessing.
The team works with all of this through a simple chat. A lawyer types a question the way they'd ask a colleague and gets a clear, sourced answer back in seconds. Conversations remember context, so follow-up questions work without repeating yourself.
What changed
Cross-document analysis — comparing terms across vendor contracts — took a full day of manual review.
New team members spent weeks reading through the library to learn the company's contractual landscape.
Questions like "do any of our contracts contain clause X?" were answered with "probably, but we'd need to check" — and checking rarely happened.
Metadata search found documents by type and date, but couldn't search within the actual content of agreements.
The same analysis takes a single query and returns a sourced answer in under two seconds.
New lawyers ask the assistant about any topic and get oriented immediately — with references to the most relevant documents to read first.
Content-level questions get answered on the spot. The team actually asks them now, because the friction is gone.
Search understands what you're looking for, even when documents use different terminology for the same concept.
"It changed how we work with our document library. Instead of spending time searching, we spend time analyzing and making decisions. The AI doesn't replace our judgment — it removes the busywork that used to stand between a question and an answer."
Legal & Compliance Team
The Results
The platform went live and delivered immediate, quantifiable improvements across Kiwi.com's legal operations.
From hours of manual searching to instant answers across the entire library.
Every contract, NDA, policy, and filing — read, indexed, and instantly accessible.
Fast search and answer generation that returns results in real time.
Complex multi-document queries that once took days now complete in minutes.
Under the hood
Legal teams can't afford "close enough." The system blends meaning-based search with exact keyword matching and a final relevance check — so it surfaces the right clause with high precision, and never invents an answer it can't back up with a source.
Smart token management, intelligent caching layers, and tiered model routing ensure the platform runs at a fraction of the cost of comparable enterprise solutions — without compromising on quality or speed.
End-to-end encryption, role-based access control, comprehensive audit trails, and EU-hosted infrastructure ensure full regulatory compliance — critical for handling sensitive legal documents at this scale.
Plugs directly into existing legal workflows and document management systems. No workflow disruption — the AI assistant enhances how the team already works, not replaces it.
Built with
State-of-the-art transformer models for semantic document representation
High-performance HNSW-indexed storage for sub-second similarity search
Top-tier LLMs with advanced prompt engineering and response grounding
Multi-stage retrieval-augmented generation with re-ranking and filtering
