Agentic insurance · Offer 3

Policy and claims agents: your policies read on site, your handlers alerted

An open AI model runs on a machine installed on your premises. It reads policies and claim files, extracts covers, exclusions, deductibles, limits and deadlines, and written rules trigger alerts to the handler, with the source page. Once the agent is installed, no document leaves your network.

In short

Policy and claims agents read your documents on site, with no upload to an online service, and turn what they read into verifiable alerts: deadline and termination notice, missing document in a file, inconsistency between a declaration and the policy, unusual clause. The handler decides; the agent accepts and refuses nothing. Actual status: tested in-house on our own supplier contracts (local reading, deadline alerts); insurance policies, claim files and the connection to a management tool will be assessed during the trial on your documents; no delivery at an insurer to date. Price set after scoping.

Typical situations for a portfolio handler

Four typical situations for insurers, mutuals and brokers managing a portfolio. For each one: the problem, then what the agent would do.

Policies sleep in PDFs and scans

Special conditions, endorsements and old policies are files, sometimes scanned, that may exist in Arabic and French. Finding a cover or an exclusion takes time and depends on one person's memory.

What the agent would do : The agent reads every document, scanned ones included, and extracts a table of clauses reviewed by the handler, with the page number of each value.

A deadline or a notice period goes unnoticed

Tacit renewal, termination notice, end of a cover period: the date is in the policy, not in the handler's calendar.

What the agent would do : Written rules compute the deadlines from the extracted clauses and warn the handler at set milestones, quoting the clause.

A claim file arrives incomplete

Declaration without a report, invoice without an estimate, unreadable document: the file goes back and forth, and the handling time grows.

What the agent would do : The agent compares the documents received with the list expected for that type of claim and flags what is missing or inconsistent with the policy, before a handler opens the file.

Documents must not leave

Policies and claims contain personal and sometimes health data. Sending them to an online AI service raises a compliance question nobody wants to settle in a hurry.

What the agent would do : Everything runs on a machine installed on your premises, with an open model. Documents stay in your network; the people allowed to access them are named in the contract.

What the agent does, and what it leaves to the handler

The AI model is only used to read. Alerts come from written, dated and verifiable rules, never from the model's impression.

What the agent does (tested) and would do (to assess during the trial)
  • Tested on our contracts: it reads contracts in native or scanned text and extracts the useful clauses into a table reviewed by a person; written rules compute deadlines and notice periods and produce alerts that state their cause and quote the clause.
  • To assess on your documents: the reading of insurance policies and claim files, the documents expected per claim type and the consistency between declaration and policy.
  • It would send the alerts to handlers by e-mail; the connection to your management tool is built at scoping, it does not exist off the shelf.
  • It keeps a log of every reading and every alert, for review and for internal control.
What the agent does not do
  • It accepts, refuses and assesses no claim: the settlement decision stays with the handler.
  • It writes nothing into your management system without validation: the clause table is reviewed before being used.
  • It does not replace the lawyer: an unusual clause is flagged, not interpreted.

The deliverables

An installation on your premises, readable rules, alerts that can be checked.

Machine and model installed on your premises

An open model installed on a machine in your network, sized at scoping according to your volume. We propose a trial on an existing workstation before any hardware purchase.

Clause table per policy

Parties, subject, covers, exclusions, deductibles, limits, dates, notice periods, renewal, each value with its source page. Reviewed and validated by the handler.

Written alert rules

The milestones, the documents expected per claim type and the consistency checks, written in plain language and editable by your teams, with the reference of each rule.

Alerts and log

One e-mail or notification per alert, with its cause, what to do and the clause quoted; a log of all readings and alerts.

When these agents are not useful

Three cases where we point elsewhere.

Your management system already does it

If your policies are entered as structured data and your system already manages deadlines, the agent would duplicate it. It is only useful where information sleeps in documents.

A very small portfolio

A few dozen policies are followed in a spreadsheet. Scoping an agent would take longer than the reading itself.

Your concern is the system itself

If the problem is an old management system to modernise, see the insurance industry page: the agents come afterwards.

How it works

Three steps, from the trial on your real policies to scoping.

  1. 1

    Trial on a batch of policies

    A batch of real policies and files, anonymised if needed, is read on a workstation in your network or on a machine we bring and that you keep under your control; nothing is copied off your premises. We compare the clause table with what your handlers know of these policies.

  2. 2

    Review with your handlers

    Correct clauses, partial clauses, missed clauses: we review the result with the people who manage these policies, and we set the useful alert rules together.

  3. 3

    Scoping

    Free first scoping conversation. Volume, formats, languages, machine, rules, alert recipients, access rights: everything is written down. The price and the timeline are fixed at this stage.

Frequently asked questions

Questions before installing policy agents on your premises

Do the documents really stay with us?

Yes. The AI model runs on a machine in your network and calls no online service to read. In Tunisia, organic law 2004-63 (article 52) requires the INPDP's authorisation for any transfer of personal data abroad; on-site processing avoids that transfer for the reading. The overall compliance of the processing remains for your data protection officer to qualify, in particular for health data.

How mature is this offer?

Tested in-house: we run local reading and deadline alerts on our own contracts, with an open model on an office machine. No delivery at an insurer to date; the trial on your real policies measures the reading quality on your documents before you commit.

Are scans and Arabic read?

Scanned documents go through text recognition before reading. Our in-house trials covered contracts in French and English; Arabic and your policy formats are assessed during the trial on your documents, as quality depends on fonts and scans. We promise no reading rate before the trial.

Which machine is needed?

It depends on the volume and the model chosen. We start with a trial on an existing workstation or a machine we bring; any hardware purchase is decided afterwards, with your IT department, never before.

What about a broker?

The need is the same, at the scale of a multi-insurer portfolio: deadlines, notice periods and covers of policies written by different companies. The alert rules are adapted to portfolio monitoring.

How does this relate to your supplier contract agents?

Same engine: local reading of documents, written rules, alerts that quote their source. It is presented for companies' supplier contracts in the Documents and contracts family of our agentic solutions.

Page reviewed and updated on · Access International, a Franco-Tunisian IT group founded in Tunis in 1999, with a subsidiary in Paris.

Try the reading on a batch of your real policies

Tell us the type of policies and the volume: we propose an on-site trial, then scoping with your handlers.

Discuss your project

See all agentic insurance offers →