Use casesFinancial institutions & fintech

  • Insurance

Insurance underwriting

Brokers and agencies send commercial proposals every day; each needs the proposal form, the survey report, and the loss history read against your guidelines before the underwriter can quote.

With the systemEach proposal and survey read, your guidelines and authority level checked; the underwriter sets terms and price.

Who decides
Underwriter

Broker email · 4 November 2026Illustrative

Risk survey report · plastic packaging plant · §4 Fire protection

The site has hydrants and portable extinguishers; there is no sprinkler system. Plastic granule stock is stored inside the production hall, next to the injection-molding machines.

→ refer to the fire lead · sprinklers or separate stock

The document as it arrives. In red, what the system marks for your expert.

1. The work

Today

  1. ReadProposals arrive from brokers and agencies by email and the agency portal; the underwriter reads the proposal form (teklif formu), the schedule of sums insured for buildings, machinery, and stock, the previous policy, the loss history, and, for larger risks, the risk engineer's survey report (risk keşif raporu).
  2. GatherThey look up the insured's past policies and claims with you, the earthquake and flood zone of the address, and what you already insure at the same site or next to it, screen by screen.
  3. CheckThe guidelines are checked by hand: whether you write the occupancy, which protections it needs, which authority level the sums insured call for, and whether the reinsurance treaty covers it or facultative cover is needed.
  4. WriteThe underwriter writes the referral note in your format, turns the survey's recommendations into conditions, drafts the special conditions and the questions for the broker; answers left blank on the proposal form are easy to miss.
  5. DecideThe underwriter at the right authority level accepts, sets the terms, deductibles, and price, asks for more, or declines; proposals wait in one queue while the broker's deadline runs.

With the system

  1. ReadThe SLM reads the proposal form, the schedule, the previous policy, the loss history, and the survey report with its photos into one structured file, each field cited to its page, and lists the questions left unanswered.SLM
  2. GatherThe agent, with read-only access, pulls the insured's past policies and claims, the hazard zones of the address, what you already insure at the site and next to it, and the sanctions screening result, and lists what is still missing.Agentic AI
  3. CheckYour guidelines and authority table, run as code, not a model, mark the occupancy as one you write, one to refer, or outside your appetite, list the protections it needs, name the authority level, and say whether the treaty covers it.Your rule
  4. WriteThe SLM drafts the referral note in your format: the survey's recommendations as proposed conditions, draft special conditions, and the questions for the broker; the price, the deductibles, and the decision are left to the underwriter.SLM
  5. DecideThe underwriter at the level your authority table names accepts, sets the terms, deductibles, and price, asks the broker for more, or declines, as before; nothing reaches the broker until the underwriter sends it.Person decides

What lands on the desk

The underwriter gets the proposal and the survey already read with their pages, the look-ups already done, and the guideline answer, the authority level, and the treaty answer already set by your own table. The referral note is drafted; the underwriter decides and sends.

FIG. 1 · Underwriting fileIllustrative
Broker proposal · fire and machinery breakdown · 04.11.2026Plastic packaging plant, organized industrial zone

SLM · proposal and survey

  • OccupancyInjection-molded plastic packaging; raw stock in the production hallproposal p. 1
  • ConstructionSteel frame, sandwich-panel roof and wallssurvey §2
  • ProtectionHydrants and extinguishers; no sprinklerssurvey §4
  • UnansweredTwo proposal-form questions: hot work and night watchproposal p. 3

Agent · checks, read-only

  • flaggedLoss history, last five yearsloss historyone fire claim
  • Earthquake zone of the addresshazard toolhigh
  • Other risks you insure at the sitenone
  • metSanctions screeningno hit

Your table · guidelines and authority

  • OccupancyPlastics processing: refer to the fire underwriting leadguidelines §3
  • RequiredSprinklers, or raw stock separated from productionguidelines §5
  • AuthoritySums insured above branch level: head office
  • TreatyWithin the fire treaty

SLM · referral note

Drafted in your format: the survey's recommendations (separate the raw stock, a hot-work permit system) as proposed conditions, and the two unanswered questions for the broker.

  • SLM
  • Agent
  • Code

Why a person decidesAcceptance, terms, deductibles, and price are the underwriter's; the system only prepares the file.

DecidesUnderwriter

2. The system

For your technical team — skip it if you like.

How it works

In

  • Proposal form and schedule of sums insured
  • Previous policy and loss history
  • Survey report and site photos
  1. SLM · understands

    Proposal and survey reader

    Fine-tuned to return one structured file, each field cited

    Reads

    • Everything under In

    Out

    • Insured, address, occupancy, sums by item
    • Construction, protections, hazards
    • Unanswered questions on the proposal form
  2. Agent · prepares

    Underwriting-file agent

    Parallel checks, read-only

    Reads

    • The structured file

    Out

    • Past policies and claims with you
    • Earthquake and flood zone of the address
    • What you already insure at the site
    • Sanctions screening
    • Missing documents listed
  3. Code · calculates

    Guideline and authority table

    Your written guidelines and limits, run as code; no model

    Reads

    • The file and the checks
    • Your guidelines and authority table
    • Your reinsurance treaty terms

    Out

    • Occupancy: write, refer, or outside appetite
    • Protections required
    • Authority level for the sums insured
    • Within the treaty, or facultative needed
  4. SLM · understands

    Referral-note writer

    The same SLM, in your referral and conditions templates

    Reads

    • The file, the checks, and the table
    • Your templates

    Out

    • Referral note
    • Survey recommendations as conditions
    • Draft special conditions
    • Questions for the broker
  5. Decision

    Underwriter at the authority level

    Reads

    • The referral note and the file

    Out

    • Accepts and sets the terms, deductibles, and price
    • Asks the broker for more, or declines
    • Refers what is above their authority
    • Has facultative cover placed when needed
How the agent's work is arrangedParallel

Parallel

The look-ups are independent of one another and every proposal needs all of them, so they run in parallel and merge into one file.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Small language model (SLM)

  1. Test set

    200–400 past commercial proposals, labeled by your senior underwriters with the fields, the survey findings, and the unanswered questions, plus 100–200 referral notes they grade against your format.

  2. Dataset

    Past proposals, survey reports, referral notes, and the conditions written; personal data masked in training copies.

  3. Continued pre-training (CPT)If needed

    Only if the base model fails the test on survey and fire-protection vocabulary.

  4. Supervised fine-tuning (SFT)

    On your proposal forms, survey report layouts, and referral and conditions templates.

  5. Reinforcement learning (RL)Skipped

    Skipped: reading and writing to a template are taught by fine-tuning, and no further reward is needed.

  6. Retrieval

    Retrieval supplies the current guidelines, clauses, and treaty terms; the authority level and the treaty answer come from the table, never from the model.

  7. Monitoring

    The underwriter reviews every note; each new edition of the guidelines or the treaty triggers a re-test of the table and the notes before it applies.

Where the agentic system stops

The underwriter at the level your authority table names decides every proposal: accept, decline, or ask for more, and sets the terms, the deductibles, and the price.

Show its tools, limits, and record
Tools and access
Read-only policy administration, claims, and document systems, your hazard-zone and accumulation tools, and the sanctions screening service; writes only the underwriting file.
May not
Accept, decline, or bind a risk, set a price, a deductible, or a condition, issue a quote or a policy, contact the broker or the insured, or grant itself any access.
Who decides and the written limit
UnderwriterThe underwriter at the level your authority table names decides every proposal: accept, decline, or ask for more, and sets the terms, the deductibles, and the price. The system never accepts or declines a risk, never sets a price, a deductible, or a condition, never issues a quote or a policy, and never contacts the broker or the insured; an occupancy outside your appetite, a sanctions hit, or sums beyond the treaty go to the underwriter, marked.
Stops when
A sanctions hit, an occupancy outside your appetite, sums insured beyond the treaty, a survey it cannot read with confidence, or a proposal-form answer the survey contradicts: the file goes to the underwriter, marked.
Record
Every document read, each check and its source, the guideline and authority-table versions, the drafts, and who decided and what was sent.

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. Quotes back to brokers sooner.
  2. Every guideline, authority, and treaty check made the same way on every proposal.
  3. Underwriter hours freed for the large and unusual risks.

What your institution measures

  • Time from proposal to quote
  • Quotes issued above the authority level
  • Policies bound with proposal-form questions unanswered
  • Referral notes sent back for rework
  • Underwriter time per proposal

Measured in the Decision Blueprint

Today's time to quote, the referrals above authority, and underwriter time per proposal, from your proposal log, set the baseline; your institution sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • The policyholder must disclose every material fact it knows or should know when the contract is made; a fact the insurer asked about is presumed material, and where the insurer gave a written question list, the policyholder answers only for what it asks, bad faith aside (Turkish Commercial Code, TTK arts. 1435–1436). An insurer that makes the contract while some of its questions were left unanswered loses the right to withdraw (art. 1442), so unanswered questions are listed before anything is bound.
  • If a material fact was not disclosed, the insurer may withdraw, within fifteen days of learning of it, or ask for the premium difference (TTK arts. 1439–1440). Policy terms that differ from the proposal to the policyholder's disadvantage are invalid (art. 1425/2), and a signed policy is due within twenty-four hours where the insurer or its agency made the contract, fifteen days otherwise (art. 1424).
  • Each insurer sets its tariffs freely under actuarial practice, compulsory insurance aside, and special conditions are written clearly under their own heading (Law 5684 arts. 11–12): the price and the conditions are your underwriter's, never the model's.
  • No machine-learning model is needed here: commercial fire and engineering losses are rare and large, so the guidelines, the authority table, and the treaty are written rules run as code, and the price comes from your tariff and the underwriter. A model estimating claim frequency could follow for a small-business book where your own history is deep enough.

Start here

We take a recent month of commercial proposals with your underwriters, time each step from proposal to quote, collect your guidelines, authority table, and treaty terms, and test the reader on your own proposals and survey reports before anything is built.

Who builds it Every engagement is led by an engineer who has built machine-learning models that real decisions run on since 2012, and language models and agentic systems in production since 2024. About eqlx

Equalize the gap.

Bring work that takes expertise. Leave knowing the first step.

A short description of the work is enough; no confidential records.

https://eqlx.ai/use-cases/insurance-underwriting · Printed copy; the website is the current version.