Vertical AI for insurance document work

AI Claim Intake you can audit

SXPHIA takes the Thai claim documents that come in, extracts the fields, checks they are complete and consistent, then tells you whether the file can pass straight through, needs a document request, or should go to a person. Every value carries its source and every decision is logged.

Contact usHow it works

Claim file C-24187

9 of 10 pages read

Documents in file

  • Hospital receipt
  • Medical certificate
  • Itemised bill· missing
  • ID card copy
  • Lab results· unreadable

Extracted fields

Source

  • Hospital99%
    Bangkok Hospital

    ↳ Receipt p.1 line 2

  • Date of treatment98%
    14 Jul 2026

    ↳ Medical certificate p.1

  • Total amount99%
    42,850.00

    ↳ Receipt p.2 table footer

  • Receipt number87%
    RC-6907-2213Awaiting staff confirmation

    ↳ Receipt p.1 top right

Itemised bill missing. A follow-up request has been drafted.

Audit record

Every confirmed value is logged with the reviewer and timestamp, exportable as a report.

Policy P-88214

Health rider + 2 endorsements

Case asked

Appendectomy, two nights inpatient. How much is payable?

Answer

Payable. 78,000 THB of benefit remaining.

Room and board capped at 4,000 THB per night; the excess is the policyholder's.

Clauses cited

  • Inpatient medical expenses, limit per confinement

    ↳ Rider, clause 3.2

  • Room and board, per-night cap

    ↳ Benefit schedule, p.2

  • Surgery not falling under exclusions

    ↳ Endorsement 2, clause 1

Waiting period
Cleared, 14 months in force
Benefit used this year
22,000 of 100,000
Pre-existing exclusion
No related history found

Comparing 3 policies

Inpatient health · from source docs

Ins. A

current policy

18,500premium/yr

Ins. B

↑ 4↓ 1

lines better / worse

+5,500premium/yr

Ins. C

↑ 0↓ 1

lines better / worse

−2,600premium/yr

Inpatient limit per stay↳ cl. 3.2100,000+50,000−20,000
Room per night↳ cl. 3.54,000−1,000same
Maximum nights30+30same
Surgeon fee↳ cl. 3.3ScheduleActualsame
Waiting period120 dayssamesame
Cataract↳ sec. 5ExcludedCoveredsame
Annual premium18,500+5,500−2,600

Measured against the current policy. Green is more cover, red is less, and a higher premium is amber because it is cost, not cover. The system does not say which to pick — the recommendation stays with the broker.

Swipe to see the other services

969,221M THB
Direct written premium in Thailand, 2025
15days
Statutory window to settle, from complete documents
12.36%
Growth of health riders, the fastest-growing line, 2025

Source: OIC insurance industry statistics, 2025.

Insurance specialists

SXPHIA puts the whole company into one problem: the document work of the insurance business, from claim files to policy wording to comparing offers. Every model we train, every feature we ship and everyone's time goes into this.

What we specialise in

  • Claim documents across health, accident, travel, motor and property
  • Policy wording across every line, riders and endorsements
  • OIC notifications and registrar orders
  • Benefit schedules and claims adjudication logic

Insurance work is always first in the queue

What people working in insurance ask for goes to the front of the roadmap, because insurance is the only work we take.

Every training round makes the models better at insurance

Handwritten medical certificates, hospital receipts, motor claim forms and policy wording, round after round. The more of this industry's material the models read, the more accurate they get at it.

OIC notifications are our daily work

Every new notification and registrar order gets read and turned into rules in the system, because this is the rulebook we follow day to day.

Why now

01

Cost pressure is pushing AI into claims faster than ever

Medical inflation runs at 8 to 15 percent a year, and in 2025 the industry introduced mandatory copayment. That pressure is what is driving claims adjudication to automate.

02

Since late 2025, claims counts as a high-risk activity

The OIC AI Governance Guideline for insurers, issued 18 November 2025, classifies underwriting, pricing and claims administration as high risk. It calls for a human in the decision, an explanation of how the AI reached its output, and event logging across the system lifecycle.

03

Explainability cannot be bolted on afterwards

A system built purely for speed cannot tell you which line of which document a number came from. We started from that question, and speed follows from it.

Services

All three run on the same models of ours, and return the same shape of answer: a value, its source, a confidence level, and a record you can audit. Start with whichever one you like. The examples on this page come from health, the line with the most complex documents, but all three work the same way on other lines.

01

Claim document review

Claim Intake

On reimbursement cases the policyholder photographs their documents and sends them over LINE or email. Claims staff open each image, work out what it is, key it in, then request missing documents two or three more times, while the 15-day clock keeps running.

  • Drag in the whole pile. Document types are sorted automatically and unreadable images are pulled out first.
  • Fields are extracted and cross-checked across documents, so mismatched totals and dates surface immediately.
  • Everything is matched against your required-document list, so one request covers everything missing.
  • Values are mapped to the codes your core system expects, such as ICD-10 on the health side.
  • Each file comes back as pass straight through, request documents, or send to a person, with the reasoning behind it.
  • Click any value to jump to the exact spot it was read from.

Complete, consistent files pass through untouched. The rest go to a queue for staff, and every value can be traced to the line it came from.

Who uses it
Claims departments at insurers, TPAs, and brokers who handle claims for corporate clients

Claim file C-24188

10 pages · 6 fields

Documents in file

  • Hospital receipt
  • Medical certificate
  • Itemised bill· missing
  • ID card copy

Where this file goes next

  • Pass straight through
  • Request documents
  • Send to a person

Reasoning behind it

  1. 1The itemised bill is missing and it is on your required-document list.
  2. 2Receipt total matches the sum of line items, no conflict found.
  3. 3A follow-up request is drafted and waiting for staff to send.
02

Is this claim payable

Policy Answer

One policy carries riders and endorsements stacked on top of it. A newer staff member answers wrong, the customer believes they are covered, and the rejection turns into a complaint.

  • Type in the case and get an answer quoting the clause from that policyholder's actual policy.
  • Exclusions, sub-limits and waiting periods relevant to the case sit alongside the answer.
  • When the information is not enough it says so rather than guessing, because guessing is where complaints start.

A new hire answers about as accurately as someone with ten years in the role, and every answer points back to a clause.

Who uses it
Customer service teams, claims teams and insurance brokers

Policy P-88214

Health rider + 2 endorsements

Case asked

Appendectomy, two nights inpatient. How much is payable?

Answer

Payable. 78,000 THB of benefit remaining.

Room and board capped at 4,000 THB per night; the excess is the policyholder's.

Clauses cited

  • Inpatient medical expenses, limit per confinement

    ↳ Rider, clause 3.2

  • Room and board, per-night cap

    ↳ Benefit schedule, p.2

  • Surgery not falling under exclusions

    ↳ Endorsement 2, clause 1

Waiting period
Cleared, 14 months in force
Benefit used this year
22,000 of 100,000
Pre-existing exclusion
No related history found
03

Compare policies

Policy Compare

A corporate client asks you to compare offers from three or four insurers. Every one writes its benefit table differently, so you read each document and retype it into a spreadsheet. It takes a day, and one missed sub-limit or exclusion comes back when the client cannot claim.

  • Drop in several documents at once; the system pulls every benefit table into the same set of fields.
  • Lay them side by side on whatever fields that line of business uses: limits, sub-limits, waiting periods, exclusions, premium.
  • Surface only what genuinely differs instead of making you scan the whole table.
  • Every cell shows which clause, page and document it came from.
  • Compare a renewal against last year's policy and see exactly what changed.

A day of work becomes a few minutes, and when the client asks why you recommended this one, there is a table with a source behind every line.

Who uses it
Insurance brokers, and employee benefits teams

Comparing 3 policies

Inpatient health · from source docs

Ins. A

current policy

18,500premium/yr

Ins. B

↑ 4↓ 1

lines better / worse

+5,500premium/yr

Ins. C

↑ 0↓ 1

lines better / worse

−2,600premium/yr

Inpatient limit per stay↳ cl. 3.2100,000+50,000−20,000
Room per night↳ cl. 3.54,000−1,000same
Maximum nights30+30same
Surgeon fee↳ cl. 3.3ScheduleActualsame
Waiting period120 dayssamesame
Cataract↳ sec. 5ExcludedCoveredsame
Annual premium18,500+5,500−2,600

Measured against the current policy. Green is more cover, red is less, and a higher premium is amber because it is cost, not cover. The system does not say which to pick — the recommendation stays with the broker.

Measured on numbers you already track

We are not asking you to take our word for it. Agree the numbers before we start, then measure them on your own work. These four are what we propose; adjust them to whatever your claims team already uses.

01

Share of files that pass without a human touch

Complete, high-confidence, internally consistent files go through. The rest queue for staff. This number converts directly into hours.

02

Number of document-request rounds

The target is one. Every extra round is time lost and a more frustrated policyholder.

03

Time from intake to decision-ready

From documents arriving to the file being complete enough for staff to decide, not how fast the model runs. The second one is not your problem.

04

Accuracy per field, not in aggregate

A wrong total and a wrong hospital name do not carry the same weight, so one blended number hides what matters.

Approach

01

Every value has a source

Each extracted field points to its position in the original document, with a confidence level. When an examiner asks where a number came from, there is an answer, not a shrug at the model.

02

People decide where people must decide

The system approves nothing on its own. Low-confidence or conflicting values go to a queue for staff to settle, which is what the guideline asks of claims work.

03

You can answer every question about where the data sits

It runs on cloud infrastructure where you choose the data region, each organisation's data stays separate, and the whole audit record is exportable at any time. If regulation calls for it, the design moves into your own systems.

About

SXPHIA started from one question. If you put AI into claims work, and one day an examiner asks where a number came from, what do you say. That question determined how the whole system is built.

We chose to go deep on a single industry because insurance paperwork has enough of its own vocabulary, rules and document formats to be a full-time job on its own. The longer we stay with it, the more of that depth accumulates.

We train our own models on Thai insurance documents: handwritten medical certificates, hospital receipts, motor claim forms and policy wording. That is material general-purpose models were never taught.

Built to be audited
Every value carries its source and confidence, not a bare result
Who confirmed which value, when, and on what basis is recorded
You choose the data region, and the whole audit record is exportable at any time

Test it on your own documents

If you sit in claims at an insurer or a TPA, or you are a broker running group cover for corporate clients, and one document process eats your team's time every day, start there. One measurable process first, then the same system extends to the rest.

or email us at hello@sxphia.com

How it starts

  1. 01

    Thirty minutes

    Work out which document process costs your team the most. Nothing to prepare.

  2. 02

    Tested on your real files

    Personal data can be redacted. We measure accuracy and hand you the number.

  3. 03

    See the report you would show an examiner

    Accuracy and the audit trail, both. If it does not clear your bar, there is no next step.