Skip to content
Telesurety
Platform · 03 · Contract intelligenceRoadmap

Ask your contracts. Get the clause with the answer.

Reads your corporate documents, summarizes them and answers your questions by pointing to the clause. When the documents don't support an answer, it doesn't invent one; it says there is no such provision.

Request a demo

Highlights

Contract intelligence
Roadmap
Question answering over agreements is not built yet. It is in scope for our planned first research project.
Policy Q&A
Prototype
A research demo tried on policy and quote PDFs, running entirely on local infrastructure.
Where contracts are processed
In-house
Agreements carry confidential commercial terms, so they will be processed on in-house servers.

Agency agreements, web service agreements, data protection (KVKK) notices and regulatory letters sit in different places today, so answering one question means reading each of them in turn. Contract intelligence is on our roadmap: it will read these documents, summarize them and answer with the clause each answer rests on. When a document gives no basis for an answer, it will say there is no such provision.

How it works

Contract intelligence is on the roadmap and builds on the fourth problem in our research agenda.

  1. 01 / 05

    Documents in one place

    Contract intelligence will run on the planned document management module. Customer and corporate documents will be kept there, versioned and access-controlled.

  2. 02 / 05

    Chunked with layout intact

    Each document will be split into passages that follow its page layout. Tables will stay whole.

  3. 03 / 05

    Searched three ways

    Each question will be matched to the relevant passages using keywords, semantic similarity and the coverage ontology together.

  4. 04 / 05

    Answered with a citation

    Every answer will quote the document and clause it relies on. If no source can be shown, no answer will be given.

  5. 05 / 05

    Interpretation stays with you

    The assistant will find and summarize the clause. Legal interpretation and the decision remain yours.

  6. Illustrative. Names and values shown are examples. Step by step. Documents in one place, Chunked with layout intact, Searched three ways, Answered with a citation, Interpretation stays with you.

Today's prototype

Prototype

A question-answering prototype for policies

We have a prototype that answers Turkish questions about an uploaded policy or quote PDF. It runs entirely on local infrastructure: a small open-weight language model, a local embedding model and a self-hosted vector database. It calls no external AI service. It will be one of the reference points the research project compares against.

  • Each PDF is validated, its text is extracted page by page and split into fixed-length, overlapping chunks tagged with page numbers. Each document is indexed under its own ID, so a question searches only that document.
  • Retrieval is hybrid. For make, model, chassis, engine and plate details, and for questions on coverage limits, replacement car, deductible and glass damage, keyword and field matching runs first. Summary questions use the opening passages of the document; other questions use semantic similarity.
  • For coverage-limit questions, when the document has amount lines, the language model does not write the answer. Coverage lines and amounts are taken directly from the document and returned as a list.
  • Answers drawn from the document show their sources as 'Page N' badges in the chat, each carrying the file name. Retrieval, generation and total time are shown with the answer.
  • Its limits are clear. It has been tried only on policy and quote PDFs with a text layer, and it does not accept scanned PDFs whose text cannot be read. It cites at page level; there is no clause-level citation yet. It also cannot abstain yet: when the context falls short, or the question is not about the document, it answers from general insurance knowledge without citing a source.

Research question

Roadmap

Can a small model stay faithful to the document?

Cited question answering over policies and corporate agreements is one of the problems on our research agenda. The open question is whether a small in-house model can stay faithful to the document and learn to stay silent when the document gives no basis. The method will combine layout-aware chunking, mandatory citations, declining unsupported questions and a small model that runs in-house.

  • No model will be trained from scratch, and answers will not come from a hosted model's API. A small open-weight model will be fine-tuned with LoRA to learn insurance terms, schema-conformant output and quoting its source.
  • Facts will always come from the document. In live use, only the in-house small model will run.
  • A large cloud model may act only as a 'teacher', generating training examples from anonymous or synthetic documents. Real customer documents will never reach it.
  • Results will be compared on the same test set against two kinds of baseline: today's rule-based readers, and off-the-shelf OCR with a general language model that has not been fine-tuned. An expert-built Turkish insurance Q&A evaluation set is itself a planned output.
  • Success targets for cited correct answers and fabricated answers were written down in advance and are still proposals. No results have been measured yet.

Data and access

Roadmap

Your contracts will stay in-house

Most corporate agreements contain no personal data, yet they carry confidential commercial information. They will therefore be processed in-house and never sent to an outside model. The platform's AI principles will apply to contract intelligence as well.

  • Documents with personal data will be processed only by in-house models on servers in Türkiye.
  • The assistant will work with the user's own permissions; a document the user cannot see stays out of its reach too.
  • An organization's data may be used for model training only with its explicit consent and after anonymization.
  • AI suggests; rules and people decide.
  • A dedicated architecture decision for document management and contract intelligence has not been written yet.

Use cases

From an intermediary's day.

  • 01

    Before opening online sales

    An operations manager wants to know what the agency's insurer web service agreements say about distance selling before going live. Today that means reading every agreement end to end. With the planned service, one question will return the relevant clause from each agreement and list the agreements that have no such provision.

  • 02

    A data protection check for a new request

    The agency is about to ask customers for a new piece of information. The compliance lead asks whether the current privacy notice covers this use. The answer will point to the relevant paragraph, or say the notice gives no basis. The compliance lead makes the call.

  • 03

    Preparing for an agency agreement meeting

    Before a meeting with an insurer, the agency owner wants to see the obligations, term and termination conditions in the agency agreement. The service will summarize the agreement and set the same headings against the agency's agreements with other insurers. Every line of the summary will link to its clause.

What it does

  • Answers with clauses

    Every answer comes with the document and clause number it relies on.

  • Document summaries

    Obligations, terms and termination at a glance for long agreements.

  • No source, no answer

    If the documents contain no provision, it says so plainly.

  • Processed in-house

    Contracts carry confidential business information; they are never sent to an outside model.

What AI does · What you decide

What AI does

Splits documents into passages, finds the relevant clauses and writes the answer with citations.

What you decide

Legal interpretation and decisions are yours. The assistant shows the clause and leaves the reading to you.

Example question

  • “Which of our insurer web service agreements contain provisions on distance selling?”

    The answer lists the relevant clause from each agreement with a short summary, and names the agreements that have no such provision.

Details

Technical details

Document types
Agency agreements, insurer web service agreements, data protection (KVKK) notices, regulatory correspondence
Planned outputs
Clause-level cited answers, agreement summaries, comparison across agreements, measurement of fabricated answers
Model
Planned: a small open-weight model, fine-tuned for insurance with LoRA, running in-house
Data handling
In-house; nothing sent to outside models
Status today
Roadmap. A question-answering prototype has been tried on policy and quote PDFs; the operations panel has no language model component.
Timeline
Planned for the AI core phase, starting with our first research project. Document management and the first mobile assistant release follow in the next phase.

Questions

Frequently asked questions

Can I use contract intelligence today?

No. Contract intelligence is on the roadmap, and the operations panel has no language model or question-answering component today. What works is a prototype tried on policy and quote PDFs; it has not been run on agreements.

What happens when the document has no answer?

The planned service will say there is no such provision and will not produce an answer. Today's prototype does not do this yet. Teaching the model to stay silent is one of the main things the research project sets out to solve.

Will our agreements go to an outside AI service?

No. Agreements will be processed in-house. A cloud model may be used only to generate training examples from anonymous or synthetic documents; your documents never reach it.

How will accuracy be measured?

With gold test sets labelled by experts and never used in training. The rate of cited correct answers and the rate of fabricated answers will be measured separately. Targets have been proposed; there are no measured results yet.

Will I be able to ask from the mobile app?

It is planned. The AI assistant will answer questions such as which insurer's web service agreement contains a distance-selling clause, and link the answer to its source. Its first release is planned for the next phase, together with document management.

Other parts

Back to platform