Insurance claims workflow datasets for AI training

An insurance claims dataset is a de-identified record of how real claims were handled from first notice of loss (FNOL) to closure: coverage checks, reserve changes, adjuster notes, documents, photos, payments and the final settlement, denial or recovery. SourceX sources claim histories from insurers, managing general agents, third-party administrators and adjusting firms running systems such as Guidewire ClaimCenter or Duck Creek, and licenses them for an agreed permitted use with claimant identities, medical details and photos redacted.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Claim files from first notice of loss to closure, with reserves, notes and outcomes
Typical systems: Guidewire ClaimCenter, Duck Creek Claims, Sapiens, Origami Risk, Xactimate, CCC ONE
Typical history: Varies by partner; long-tail claims can stay open for years
Modality: Structured claim events, adjuster notes, letters, estimates and damage photos
Delivery formats: Agreed per order; JSONL or Parquet claim records with documents and images linked
Preparation: Claimant, insured and witness identities and medical details removed; photos redacted
Licensing: Non-exclusive or snapshot-exclusive; third-party reports excluded unless their terms allow
Availability: Sourced to your spec from carriers, MGAs and TPAs; not guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
claim_id	string	Pseudonymous claim key, stable across exposures, payments, notes, documents and photos.
line_of_business	enum	Personal or commercial auto, homeowners, commercial property, workers' compensation, general liability or another line.
loss	object	Loss and report dates, cause-of-loss code, the state whose rules apply and the FNOL channel.
coverage_check	object	Policy reference, coverages in force, limits and deductibles, open coverage questions and any reservation of rights.
exposures	array	Each coverage and claimant pairing, such as third-party property damage or bodily injury, handled as its own unit with its own status.
exposures[].reserves	array	Dated reserve changes for indemnity and expense on that exposure.
payments	array	Indemnity and expense payments by payee type, plus subrogation and salvage recoveries.
activity	array	The ordered file history — adjuster notes, diary tasks, calls, letters, assignments, authority requests and status changes.
documents	array	Typed attachments such as estimates, recorded statement transcripts, police reports, medical bills, demands and releases.
photos	array	Damage and scene images with subject tags and a record of what was redacted in each.
siu	object	Special investigations unit referral, its reason codes and the investigation result as recorded.
litigation	object	Attorney representation, suit status and how a litigated exposure resolved, without party names.
outcome	object	Closing disposition, total paid, recoveries, reopen count and days from report to closure.

Example record

{
  "claim_id": "clm_4e81b2",
  "line_of_business": "personal_auto",
  "loss": { "loss_date": "2022-11-04", "reported_at": "2022-11-05T08:14:00Z", "fnol_channel": "phone",
            "cause": "rear_end_collision", "state": "[STATE]" },
  "coverage_check": { "policy_ref": "pol_19c7", "in_force": true, "reservation_of_rights": false,
                      "coverages": ["collision", "pd_liability", "bi_liability"] },
  "exposures": [
    { "id": "exp_1", "type": "third_party_property", "claimant": "clmt_01", "status": "closed_paid",
      "reserves": [{ "d": "2022-11-07", "indemnity": 4200 }, { "d": "2022-11-16", "indemnity": 6150 }] },
    { "id": "exp_2", "type": "third_party_bodily_injury", "claimant": "clmt_01", "status": "closed_settled",
      "reserves": [{ "d": "2022-11-21", "indemnity": 15000, "expense": 2500 },
                   { "d": "2023-05-31", "indemnity": 22000, "expense": 3000 }] }
  ],
  "payments": [
    { "exposure": "exp_1", "kind": "indemnity", "payee_type": "repair_shop", "amount": 5930 },
    { "exposure": "exp_2", "kind": "indemnity", "payee_type": "claimant_and_attorney", "amount": 18500 }
  ],
  "activity": [
    { "t": "2022-11-05T08:31:12Z", "type": "note", "role": "fnol_rep",
      "text": "Insured admits looking at nav screen; hit vehicle ahead at low speed. Other driver held neck at scene." },
    { "t": "2022-11-16T11:30:00Z", "type": "note", "role": "pd_adjuster",
      "text": "Supplement after teardown: bumper reinforcement and park sensor. Reserve raised." },
    { "t": "2023-05-30T11:05:18Z", "type": "note", "role": "bi_adjuster",
      "text": "Specials reviewed [MEDICAL_DETAIL_REDACTED]. Countered demand; supervisor approved authority." }
  ],
  "documents": [
    { "id": "doc_03", "type": "repair_estimate", "photos": ["img_07"] },
    { "id": "doc_05", "type": "letter_of_representation", "exposure": "exp_2" },
    { "id": "doc_09", "type": "release", "exposure": "exp_2" }
  ],
  "photos": [{ "id": "img_07", "subject": "rear_bumper_damage", "redacted": ["license_plate"] }],
  "siu": { "referred": "2022-12-14", "reasons": ["early_attorney_rep", "prior_similar_claim"],
           "result": "not_substantiated" },
  "litigation": { "attorney_rep": true, "suit_filed": false },
  "outcome": { "closed": "2023-06-20", "disposition": "paid_and_settled", "total_paid": 24430,
               "recoveries": 0, "reopened": 0, "days_report_to_close": 227 }
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Train claims-handling agents on adjuster decisions

Activity logs show what an experienced adjuster did next at each point in a claim — which document to request, when to raise a reserve, when to ask for settlement authority — and how the claim ended.

Predict severity and reserves from first notice of loss

Dated reserve histories paired with final paid amounts give a labeled target for early severity estimates, including how far the first reserve sat from the eventual cost.

Extract and check claim documents and photos

Repair estimates, police reports, medical bills, demand letters and damage photos, each tied to the decision it informed, make training and test material for document and vision models.

Triage special investigation referrals

Referral reason codes and recorded investigation results support fraud-triage research, as long as a referral and a confirmed finding are kept apart as labels.

Evaluate coverage reasoning

Coverage questions, reservation-of-rights letters and denial letters, with the facts behind them, make held-out test items for checking whether a model reads policy conditions the way adjusters did.

Use-case guides: Enterprise and computer-use agents, Healthcare administration AI, Private evaluation sets

What makes this data valuable

Reserve trajectories

Each reserve change is a dated estimate of final cost that the closing payment later scores.

Exposure-level outcomes

Separate results for each coverage and claimant show which parts of a claim paid, settled or were denied.

Adjuster reasoning

Notes record coverage doubts, liability calls and negotiating positions that never appear in letters.

Linked evidence

Estimates, statements and photos sit at the step where they changed a decision.

Investigation and litigation results

SIU findings, attorney involvement and suit outcomes mark the claims that became contested.

Line and state spread

Several lines of business and states widen the policy forms and handling rules a model sees.

What a claim file captures that nothing public does

A claim file captures how one claim was handled, which published insurance data never shows. What insurers, regulators and courts release is mostly aggregate loss experience, rate filings and the occasional opinion in a disputed case. It describes what claims cost in total, not the calls, document requests, coverage questions and authority requests that moved each claim to its result.

The hardest part of that handling is matching untidy facts to policy language: whether late notice affects coverage, whether a driver had permission, whether water damage was sudden or gradual. Claim files record how adjusters answered those questions in notes and letters written before anyone knew how the claim would end. They also carry each carrier's house rules, such as authority limits, referral triggers and documentation standards, which appear in no public source. A model built for claims work needs both the policy reasoning and the house rules, and only real files hold the two together.

What to watch for in claims data

Most problems with claims data come from time — claims stay open, handling practices change and claim systems get replaced — and one comes from copyright.

Open claims have no outcome yet. In a recent snapshot, the closed files are mostly fast, simple claims, while long-tail bodily injury and workers' compensation exposures are still open. Scope by report year and require closure, or treat open claims as censored labels.
Reserves move for reasons outside the claim. New authority thresholds, revised reserving guidelines and changes of supervisor all show up in reserve histories. Compare reserve behavior across years before pooling them.
Legacy conversions leave marks. Claims migrated from an older system can show notes collapsed onto one date, remapped codes and missing attachments. Ask for the migration date, and flag or exclude converted files.
Codes drift while notes carry the detail. Cause-of-loss and closing codes get revised and applied unevenly, so the narrative notes are often the more reliable record of what happened.
Standard forms belong to someone else. Much policy wording comes from forms published by advisory organizations that hold copyright in them, so the claim data may be licensable while the form text is not.

What to check before licensing

Trace each file to its owner. Claims handled by a third-party administrator, MGA or independent adjusting firm usually belong to the carrier or self-insured employer, and the claims service or delegated-authority agreement governs what may be done with them.
Ask how health information in injury and workers' compensation files is handled, including medical bills, diagnoses repeated in adjuster notes and recorded statements, and which state rules apply to it.
Review photo redaction on a sample, covering faces, license plates, VINs, house numbers, documents in frame and home interiors.
Keep SIU referrals apart from confirmed findings, and get each year's referral criteria so you know what made a claim eligible for investigation.
List third-party content in the files, such as claims-database search reports, motor vehicle records and vendor appraisals, and agree what is excluded.
Exclude privileged material from litigated claims, such as defense counsel reports and coverage opinions.
Check on the sample that claim, exposure, payment, document and photo IDs join cleanly.
Ask which catastrophe events fall inside the period. Claims from a hurricane, wildfire or hailstorm are often handled by deployed or independent adjusters under catastrophe procedures, so they behave differently from everyday claims.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can insurers' claim files be licensed for AI training?

Yes, with the consent of the files' owner, which is usually the carrier or a self-insured employer rather than whoever handled the claims. Third-party administrators and adjusting firms often hold the files on a client's behalf, so that client must authorize licensing too. Personal and medical details are removed and photos redacted before delivery. Whether files exist for your lines, states and years depends on which partners agree to license.

Do injury claims contain health information?

Yes. Bodily injury and workers' compensation files hold medical bills, treatment records, diagnoses and adjusters' summaries of them. Property-casualty and workers' compensation insurers generally fall outside HIPAA's definition of a covered entity, but many states' insurance privacy rules set stricter conditions for health information than for other claim data. Expect medical detail to be removed or generalized into injury categories, as agreed in scope.

Do claims datasets include fraud labels?

Some do, as special investigations unit (SIU) referrals with reason codes and a recorded result. Treat them with care. A referral is a suspicion, not a finding; unreferred claims are unlabeled rather than clean; and results such as "not substantiated" are ambiguous. SIU files can also fall under state fraud-reporting and confidentiality rules that limit what a partner can deliver.

Which lines of business does this data type cover?

Property-casualty claims, such as personal and commercial auto, homeowners, commercial property and general liability, plus workers' compensation. Auto and property claims close quickly and carry many photos, while liability and workers' compensation claims run longer and gather medical and legal detail. Health insurance claims fall under HIPAA and are covered, from the provider side, under healthcare revenue cycle records.

Can damage photos and repair estimates be included?

Usually, after redaction and a check of third-party terms. Photos are reviewed for faces, license plates, VINs, house numbers and documents in frame, and interior shots of homes may be excluded. Estimates from estimating software are delivered as line items and totals where the software terms allow, because the parts prices and labor times in them typically come from the vendor's databases.

How do state insurance rules affect claims data licensing?

They shape what a carrier will release and how. Claims handling, privacy and data security are regulated state by state, so files from different states can carry different notice requirements and limits on sharing health or financial details. Several states have also issued guidance on insurers' own use of AI, which matters if a model trained on the data will be used in claims handling.

Why would a carrier restrict how I use its claims data?

Because claim files reveal its losses, reserving and settlement practices. A carrier may permit training for claims automation or document processing but exclude pricing, underwriting or competitive benchmarking, or require that outputs never identify it. These limits are written into the license as permitted use, alongside exclusivity, duration and deletion terms, and they can affect price.

Healthcare revenue cycle and prior authorization records
Linked prior authorizations, claims, remittances, denials and appeals for real encounters
Enterprise workflow and task execution histories
Linked task trajectories from request to outcome, across every tool the work touched
Contact center call recordings and transcripts
Recorded service and support calls with diarized transcripts, dispositions and QA scores
SOPs, playbooks and internal knowledge bases
Written procedures with page history, ownership and links to execution records
Human feedback and QA-scored work
Work items with scores, verdicts and corrections from the people who reviewed them

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.