Skip to content
Telesurety
Platform · 06 · AI Assistant and AgentsRoadmap

An assistant that will know insurance, understand you and work with your approval.

The AI Assistant will run on an in-house language model of at least 12 billion parameters. It will understand the requests you type in Turkish and answer from the policy and the agency's own records, with the source. Its agents will use the panel's own endpoints with your permissions to do the work or prepare it for your approval. Every irreversible action will wait on an approval card.

Request a demo

Highlights

Assistant and agents
Roadmap
The panel and the staff app contain no language-model, speech-recognition or agent code yet. The staff app and policy Q&A are prototypes.
In-house model target: parameter count
At least 12 billion
It will be built on an open-weight model and adapted to insurance. No model of this size runs today.
Server-side permissions
Live
Access is checked on the server for every menu area; the agents' tools will build on it.

AI Assistant and Agents is being designed so agency staff can start panel work with a single sentence. The assistant will run on an in-house language model of at least 12 billion parameters: it will know insurance terms and workflow, understand requests in Turkish, and answer from the policy and the agency's own records, with the source. The policy and regulation assistant will answer questions. The quote and renewal agent and the inbox and claims agent will use the panel's own endpoints with your permissions to do the work or prepare it for your approval. Insurer selection, premiums and permissions will stay with rules, and every irreversible action will wait for an approval card. All of this is on the roadmap. The foundation running today: the staff app prototype, the question-answering prototype over policy PDFs, and the panel's live server-side permission checks.

How it works

The AI Assistant and its agents are on the roadmap. The operations panel and the staff app contain no language-model, speech-recognition or agent code today. The foundation running today: the staff app prototype, a prototype for question answering with sources over policy PDFs, and the panel's live server-side permission checks.

  1. 01 / 04

    It will understand your request in Turkish

    Type a question or ask for a task: “When does this customer's own-damage policy expire?” or “Prepare an own-damage quote request for this plate.” The assistant will work out the request and the record it refers to; if it fits more than one record, it will ask instead of guessing. It will work by text first; in-house Turkish speech-to-text is a planned research topic.

  2. 02 / 04

    It will find the answer at its source

    The in-house model of at least 12 billion parameters will know insurance terms; the facts themselves will come from the policy, the coverage terms and the agency's own records. Every answer will come with the page or record it rests on; if the data is not there, the assistant will say it could not find it.

  3. 03 / 04

    Agents will work through the panel's endpoints

    Every job the panel does is already an API endpoint. The agents will use those endpoints as tools, with your permissions, so no separate business logic is written for them. The panel's rules, not the model, will decide which insurers to call; premiums will come only from the insurer's response.

  4. 04 / 04

    Irreversible work will wait on an approval card

    Sending a quote, issuing a policy or cancelling will first show a card stating who it is for, which insurer and which premium. Nothing will happen until you tap it, and the action will be logged as done through the assistant with your approval.

  5. Illustrative. Names and values shown are examples. Step by step. It will understand your request in Turkish, It will find the answer at its source, Agents will work through the panel's endpoints, Irreversible work will wait on an approval card.

Today

Prototype

Two prototypes running today

Two pieces the assistant will build on run today as prototypes. The staff app has no backend of its own; it works on the operations panel's REST API, with the same data and the same server-side permissions as the panel. The policy Q&A prototype runs entirely on local infrastructure and calls no outside AI service.

  • Staff app (iOS and Android, Turkish and English): customer search, the month's production and upcoming renewals, requests awaiting approval, comparing insurer offers and turning one into a quote, adding vehicles from the registration, and read-only claim files.
  • Q&A prototype: answers Turkish questions about an uploaded policy or quote PDF. It runs on a smaller open-weight language model, a local embedding model and a vector database we host ourselves.
  • Answers based on the document show their sources by page number. For coverage-limit questions, when the document has an amount line, the amounts are taken straight from the document.
  • Its limits are clear: sources are page-level, with no clause-level citation yet. It does not yet decline when the context falls short and may answer from general insurance knowledge without a source. In the assistant, the goal is to say “I could not find it” in that case; teaching the model to do so is what the research is about.
  • The operations panel and the staff app contain no language-model, speech-recognition or agent code today.

Design

Roadmap

A model that will know insurance, agents that will do the work

The assistant's information side, sourced answers and summaries, will build on our research on question answering with sources. The side that will do the work will consist of agents that use panel endpoints as tools. The agents are planned as separate research topics, and the staff app will be where they are tried and measured in the field.

  • Model: the target is a language model of at least 12 billion parameters. It will not be trained from scratch, and answers will not come from an outside model's API. It will be built on an open-weight model and adapted with LoRA to insurance terms, schema-conformant output and quoting its sources.
  • Facts will always come from documents and records. The model will know the terms and the workflow; it will not state from memory what a policy pays for.
  • Policy and regulation assistant: it will give page-cited answers over policies and coverage terms. The assistant will also answer report questions such as “Why did own-damage production drop this month?” from the agency's own data.
  • Quote and renewal agent: it will list due renewals, prepare quote requests for the panel's rule-based quote flow, flag customers at risk of leaving and offer coverage suggestions as drafts. You will start the request; rules will decide which insurers it goes to.
  • Inbox and claims agent: it will combine multi-document cases into one file, classify the intent of Turkish messages, find missing documents and draft replies.
  • Sequence: the first release will open with questions, summaries and approvals. The agents' ability to do work will follow in a later phase.

Security and data

Partly live

In-house, with your permissions, with your approval

Part of the foundation is live today. The panel API checks access on the server for every menu area and narrows report and claim data to the user's department scope. The agents will use the same endpoints, so they will pass the same checks. Phone security is built into the staff app prototype; the assistant's own safeguards are on the roadmap.

  • Live: server-side access checks per menu area; the quote endpoints, for example, require quote access.
  • Prototype: device-bound sessions, biometric or device-lock unlock, automatic lock after inactivity, and remote sign-out from one or all devices.
  • Roadmap: the model will run in-house on servers in Türkiye; no model will run on the phone. Cloud models will be used only with anonymized or synthetic data.
  • Roadmap: model output will never be an input to a permission check, and it will pass rule-based validation. Checks such as ID-number check digits, plate format and date consistency will run independently of the model.
  • Roadmap: every action taken through the assistant will be written to the audit trail as performed through the assistant with your approval. No model version will go live before passing test sets prepared by experts.

Use cases

From an intermediary's day.

  • 01

    A customer on the phone: does this policy pay out for flood?

    A sales rep will type the customer's question into the assistant. The policy and regulation assistant will show the relevant page of the policy; if the policy gives no basis, it will say it could not find one. The rep will decide what to tell the customer. Today, questions like this are tried with page sources in the prototype over policy PDFs.

  • 02

    A day that starts with renewals

    An agency manager will ask “What's waiting today?” in the morning. The assistant will return due renewals, open claim files and quotes awaiting approval. The quote and renewal agent will flag customers at risk of leaving and prepare renewal quote requests. The manager will start the requests; before a quote goes to a customer, the approval card will show the customer, the insurer and the premium.

  • 03

    An email for a claim file

    A customer emails one of the documents requested for a claim file. The inbox and claims agent will link the email and its attachment to the right claim file, check the expected-documents list for what is still missing and draft a reminder to the customer. If several customers share the name, it will first ask which one is meant, because a message to the wrong person can count as a personal-data breach. The draft will not be sent until the claims handler approves it.

What it will do

  • Roadmap

    An in-house model that will know insurance

    This model of at least 12 billion parameters will be built on an open-weight model and adapted to insurance terms and workflow. It will understand requests in Turkish and take the facts from documents and records.

  • Roadmap

    Policy and regulation assistant

    It will answer questions such as “Does this policy pay out for flood?” by showing the relevant page. If it finds no basis, it will say it could not find one.

  • Roadmap

    Quote and renewal agent

    It will gather due renewals, prepare quote requests and flag customers at risk of leaving. Coverage suggestions will arrive as drafts; the decision will stay with you.

  • Roadmap

    Inbox and claims agent

    It will link incoming emails and documents to the right claim file, find missing documents and draft replies. The decision to send will stay with you.

  • Roadmap

    Turkish voice commands

    The goal is to give requests by voice as well. Converting Turkish speech that contains insurance terms and plates to text in-house is a separate research topic. Typing will not wait for that research.

  • Prototype

    Policy Q&A prototype

    A trial that answers Turkish questions over policy and quote PDFs, shows the page number behind each document-based answer and runs entirely in-house. It does not yet know how to stay silent when there is no basis.

  • Prototype

    Staff app

    The iOS and Android app that will be the assistant's face runs today as a prototype, on the panel's API and permissions.

  • Live

    Server-side permission checks

    Access is checked on the server for every menu area. The agents' tools will build on these checks; what you cannot do in the panel, an agent will not be able to do either.

What AI does · What you decide

What AI does

It will understand the request, find the answer with its source, and do the work through the panel's endpoints with your permissions, or prepare it for approval.

What you decide

Which quote to recommend, what goes to the customer and the policy decision are yours. Insurer selection, premiums and permissions will stay with rules; no irreversible action will happen until you tap the approval card.

Rules set from the start

  • Roadmap

    Rules will decide

    Which insurer to call, premiums, sums insured, commission and permissions will not be left to the model. The model will not estimate prices; premiums will come only from the insurer's response.

  • Roadmap

    Approval card

    Sending a quote, issuing a policy and cancelling will first open a card showing who it is for, which insurer and which premium.

  • Roadmap

    A source, or “not found”

    Every answer will come with the page or record it rests on. If the data is not there, the assistant will say it could not find it.

  • Roadmap

    It will ask when unsure

    If a request fits more than one customer or policy, the assistant will not guess; it will ask which one is meant.

  • Roadmap

    In-house, in Türkiye

    The model will run on in-house servers in Türkiye; no model will run on the phone. Personal data will not be sent to an AI service outside the company.

  • Roadmap

    Every action on the record

    Every action taken through the assistant will be written to the audit trail as performed through the assistant with your approval.

Details

Technical details

Status
Assistant and agents on the roadmap; policy Q&A and the staff app in prototype; server-side permission checks live.
Model
Target: an in-house language model of at least 12 billion parameters, built on an open-weight model and adapted to insurance with LoRA. Today's prototype runs on a smaller model.
Where the model will run
In-house, on servers in Türkiye. No model will run on the phone; personal data will not be sent to an AI service outside the company.
Agents
A quote and renewal agent and an inbox and claims agent; on the information side, a policy and regulation assistant. Their tools will be the panel's own API endpoints; no separate business logic will be written.
Pricing, insurers and permissions
Will stay with rules. Premiums will come only from insurer responses; web service connections to 14 insurers are live in the panel today.
Input and interface
Text first; Turkish voice commands are a planned research topic. Document and photo upload secondary. Interface: the iOS and Android staff app.

Questions

Frequently asked questions

Can I use the AI Assistant today?

No. The operations panel and the staff app contain no language-model, speech-recognition or agent code today. What runs today: the staff app prototype, a prototype for question answering with sources over policy and quote PDFs, and the panel's live server-side permission checks.

What kind of model will the assistant use?

The target is a language model of at least 12 billion parameters. It will not be trained from scratch, and answers will not come from an outside model's API. It will be built on an open-weight model and adapted to insurance terms, schema-conformant output and quoting its sources. It will run in-house, on servers in Türkiye. No model version will go live before passing test sets prepared by experts.

Will the agents issue policies for me?

Not at first. The first release will open with questions, summaries and approvals. Doing work, such as starting and sending quotes or issuing policies, will come in a later phase, and even then no policy will be issued until you tap the approval card.

Can an agent choose an insurer or quote a price?

No. Rules will decide which insurers to call, the premium, commission and permissions. The agent will prepare the quote request; you will start it, and the panel's rule-based quote flow will send it to the insurers. Premiums will come only from insurer responses; the model will not estimate prices. Web service connections to 14 insurers are live in the panel today.

Does customer data leave the company?

No, by design. Personal data will be processed only by models running on in-house servers in Türkiye. Cloud models may be used only with anonymized or synthetic data, to generate training examples.

Other parts

Back to platform