AI data market
Insurance workflow data investment: score decisions, not claim counts
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Insurance workflow data investment should rest on decisions a model can be graded on, not on claim counts. A usable record links the evidence on file, the rule applied, the decision, its stated reason and the later outcome. Our tally of the claims format CMS suggests for patient access shows why volume misleads: payment status is required in all ten profiles, a reason code in none. Rate traceability and sensitivity before upside.
Value drivers
Insurance workflow data
Key takeaways
- Claim counts measure events; the scarce asset is a decision trail with a stated reason and a known outcome.
- In the CARIN Blue Button format that CMS suggests for its Patient Access API, all ten explanation-of-benefit profiles require payment status and none requires a reason code.
- Regulation is adding reasons where data is most protected: CMS-0057-F requires specific prior authorization denial reasons from 2026, inside HIPAA-covered health plans.
- EIOPA's April 2026 work on the AI Act places the risk of life and health pricing systems mainly in data selection and governance, not model design.
- On our scorecard, property and casualty claim files and commercial underwriting balance traceability against sensitivity best; health prior authorization ranks highest on both.
- Ask any supplier, SourceX included, for the share of decisions with a coded reason and a linked outcome, not for record counts.
Is insurance workflow data investment a bet on claim volume?
No. Insurance workflow data investment is a bet on decision records: files that show what an insurer knew, which rule or policy wording it applied, what it decided, why, and what happened next. Claim volume tells you how many events a carrier processed. It says nothing about whether those events left a trail a model could learn from or be scored against.
The distinction matters because the insurance work AI developers want to automate is judgment, not arithmetic. O*NET's profile of claims adjusters, examiners and investigators lists tasks such as examining claim forms and other records to determine coverage, checking police reports, medical records and bills to judge liability, interviewing claimants, witnesses and physicians before a settlement or denial, and entering payments and reserves. Every one of those tasks ends in a decision; only some leave a record of the reasoning.
So the investor's first test comes before market size. Ask how traceable the decisions in a workflow are and how sensitive the records that carry them are. Commercial potential means something only after that, because a large book of business whose decisions cannot be traced or licensed gives a buyer very little to use.
Event rows and decision trails: where an insurer's reasoning is kept
Most of an insurer's reasoning lives in the claim file, not in the structured claim row. The row records the event: dates, coverage, amounts and status. The file holds what produced the decision: correspondence, adjuster notes, estimates, statements and reports.
Regulators insist on the file. California's Fair Claims Settlement Practices Regulations require insurers, under section 2695.3, to keep for examination every document, note and work paper that reasonably relates to a claim, copies of correspondence included, and to log when each material document was received and processed. The same regulations forbid denying a claim on the strength of a phone call or interview that the claim file does not document.
That makes a well-kept claim file the nearest thing insurance has to a gradable decision record, and also the hardest to prepare. Notes and letters are free text. They name claimants, witnesses, doctors and repair shops who never signed anything with the insurer, and the reason for a decision is often spread over several entries instead of sitting in one field.
A field tally: what the payer claims format CMS points to requires
In the claims feed standardized for federal patient access rules, the decision reason is optional. We read the explanation-of-benefit profiles in HL7's CARIN Consumer Directed Payer Data Exchange guide, better known as the CARIN IG for Blue Button, which CMS has suggested payers consider for the Patient Access API its interoperability rule requires. CMS says a payer that builds that API by following the suggested guides to specification could comply with the rule.
Across all ten EOB profiles, five non-financial base profiles and the five full profiles built on them, the benefit payment status slice, which marks payment as in or out of network, has a minimum of one, so every conformant record carries it. The five full profiles also require total amounts. Reason codes appear in every profile, as adjustment reasons for non-covered amounts or as reject reasons on pharmacy claims, and their minimum is zero in all ten. So is the minimum for the date the claim was received. The guide marks these elements must support, meaning a system that holds the value must be able to send it, but a record without them still conforms.
We ran the tally on HL7's published 2.1.0 tag (commit 3fa63d4, February 2025) and on version 2.2.1 in the repository's main branch (commit 09f7bde, September 2026), with identical results. The finding is not that payers leave reasons out. It is that a count of conformant claims tells an investor nothing about how many carry one, so that share has to be measured in the supplier's actual records.
| Field group in the EOB profiles | Profiles declaring it | Profiles requiring it | What counting claims tells you |
|---|---|---|---|
| Benefit payment status (in or out of network) | 10 of 10 | 10 of 10 | Present on every conformant record |
| Total adjudicated amounts | 5 of 5 full profiles | 5 of 5 full profiles | Present on every financial record |
| Reason codes (adjustment or reject reason) | 10 of 10 | 0 of 10 | Nothing: could be missing from every record |
| Date the claim was received | 10 of 10 | 0 of 10 | Nothing: when the claim arrived may be missing |
Regulators are adding reasons where records are most sensitive
The direction of travel is toward more traceable decisions, but it is fastest in health coverage, where records are most protected. Under CMS's Interoperability and Prior Authorization final rule, CMS-0057-F, impacted payers must give a specific reason when they deny a prior authorization request, starting in 2026, and must post prior authorization metrics on their websites every year, with the first set due by March 31, 2026. Drugs are outside the requirement. The impacted payers are Medicare Advantage organizations, state Medicaid and CHIP programs and their managed care plans, and qualified health plan issuers on the federally facilitated exchanges, the only commercial payers the rule reaches.
All of those records sit inside HIPAA. HHS's summary of the Privacy Rule counts health insurers, HMOs and government programs that pay for care as health plans, which makes them covered entities. The same summary leaves out plans that cover only workers' compensation, auto insurance or property and casualty risk, so the injury details in an auto claim file are not health plan records under HIPAA, although other privacy and insurance rules can still apply.
Europe puts the risk in the data. The AI Act lists AI systems used for risk assessment and pricing of natural persons in life and health insurance as high-risk. In an April 2026 letter, EIOPA asked the co-legislators and the European Commission to take generalized linear and additive models out of that category, and its technical annex concluded that the risks come mainly from data selection, data governance and underwriting or pricing policies rather than from the model architecture. The request is a proposal, and we found no record of its adoption. For an investor the signal is that the EU's insurance and pensions authority treats the data, not the model, as the thing to watch.
The insurance workflow scorecard: traceability against sensitivity
Read across the rows and the tension is plain. Where regulation forces reasons onto the record, health prior authorization above all, the strictest privacy law applies. Commercial underwriting carries less personal data, and policy servicing systems record what changed far more reliably than why. Property and casualty claim files sit between the two: the reasoning is there because examiners can demand it, but it is buried in notes about people who are not customers.
Score each workflow on decision traceability and on sensitivity, then name the single evidence gap that would most change its rating. The ratings are editorial judgments built from the sources in this article as of October 27, 2026, and they rate workflows, not companies.
| Workflow and buyer task | Decision traceability | Sensitivity | Rights and preparation burden | Evidence gap to close |
|---|---|---|---|---|
| Health prior authorization: approve, deny or ask for more | High and rising: specific denial reasons required from 2026 for impacted payers | Very high: health plan records under HIPAA | Heavy: clinical documents, provider and member identities | Whether reasons are coded or free text, and how appeals ended |
| Health claim adjudication: pay, adjust or deny lines | Low in the exchange format: reason codes optional in all ten CARIN profiles | Very high: covered entity data under HIPAA | Heavy: de-identification at claim-line scale | Share of claims in the actual records with a coded reason |
| Property and casualty claims: coverage, liability, settlement | Medium to high in the file: documents, notes and dates kept for examination | High: claimant and third-party data; injury details outside HIPAA's health plan scope | Heavy: police reports, statements and notes about non-customers | Whether notes link to the final payment and any reopening |
| Life underwriting: accept, rate or decline an applicant | Medium: the decision is recorded, the reasoning often only in notes | Very high: medical evidence; high-risk under the EU AI Act for natural persons | Heavy: consent terms attached to medical evidence | How often a decline reason is captured as a field |
| Commercial underwriting: refer, quote or decline a business risk | Medium: referral and decline notes alongside quote versions | Medium: mostly business data, some named individuals | Moderate: broker and client confidentiality terms | Buyer demand evidence and decline-reason coverage |
| Policy servicing: endorsements, cancellations, changes | Low for why, high for what changed | Medium to high: policyholder personal data | Moderate: system change logs plus correspondence | Whether a change reason exists outside free text |
Demand evidence: insurers already apply AI to these workflows
The clearest dated demand signal comes from regulators surveying insurers. According to an article in the NAIC's Journal of Insurance Regulation, the NAIC's surveys found that 88% of personal auto insurers, 70% of homeowners insurers, 58% of life insurers and 84% of health insurers use, intend to use or are exploring AI and machine learning; the four survey reports were issued between December 2022 and May 2025. The life report counts 94 of 161 responding companies, and its figures cover predictive models including generalized linear and additive models, a broader class than the name AI suggests.
That is demand inside insurers, and it cuts both ways for a data thesis. It shows the buyer tasks are real: claims, underwriting and servicing decisions are being automated. It also shows carriers training on their own books, which can make them builders rather than licensors. The thesis depends on a second group, outside developers who need decision records from many carriers to build tools that work across them, and the surveys do not measure that group.
Why a bigger book of business can mean less usable supply
The strongest objection to this thesis is that volume wins anyway: with enough claims, a model can learn patterns without being told the reasons. For tasks scored on amounts or outcomes alone, such as predicting severity, that can be true. For tasks scored on judgment, such as whether coverage applied or a denial was justified, volume has to pass three filters.
Reasons come first. If reason codes are optional in the exchange format and scattered through notes in the file, more claims mostly mean more unexplained outcomes. Rights come second. A claim file is crowded with third parties, from claimants and witnesses to treating physicians, and each one adds review to every record. Regime comes third. The larger the health or life share of a book, the more of it sits under HIPAA or, in Europe, inside the AI Act's high-risk category, which adds scrutiny to any proposed use. A large carrier can therefore hold more records and less licensable, gradable supply than a smaller specialist with clean files.
Five ways the traceability thesis could fail
Name the failure modes before the upside. Each is live, and each comes with a signal an investor can track.
- Carriers keep the data. The NAIC surveys show most insurers already use or are exploring AI; if they train in-house and never license, outside supply stays thin however good the records are.
- Regulatory scope narrows. EIOPA's request to drop generalized linear and additive models from the AI Act's high-risk list shows the category can shrink as well as grow, which would weaken any premium that well-documented data earns from governance demands.
- Reasons stay unstructured. If decision reasons remain in free-text notes, preparation cost per record may outrun what buyers will pay, and the usable share stays small.
- De-identification strips the reasoning. The notes that explain a decision usually name the people involved, and removing identifiers can remove the context that made the record worth licensing.
- Buyers concentrate. If a handful of developers buy most insurance decision data, suppliers depend on a few renewals, and pricing follows those buyers' plans rather than the quality of the files.
Diligence requests for any insurance data supplier, SourceX included
SourceX should answer the same requests. It turns business records that a company chooses to offer into licensed AI datasets and runs the whole chain itself, from finding and qualifying the records through rights review, contracting and preparation to delivery and payment, and no data reaches it until the company has signed off on scope and terms and executed the agreement. Its supply ratios and commercial figures are not public; investors who ask receive company and commercial metrics from SourceX directly. The company's current raise, under Rule 506(c), is limited to accredited investors, and materials are available on request at sourcex.si/investors.
Ask every supplier for ratios rather than record counts, and fix the definitions yourself so that each one answers the same question.
- Reason coverage: the share of decisions in the offered records with a coded reason, and separately the share with a reason only in notes.
- Outcome linkage: the share of decisions tied to a later event such as payment, appeal, reopening or lapse, with the cutoff date used for unfinished cases.
- Rule versioning: whether each record carries the policy form, guideline or rule version in force on the decision date.
- Third-party exposure: how many people besides the policyholder appear per record, and how the rights review treated them.
- Regime map: the share of records that are HIPAA health plan records, the share inside the EU AI Act's high-risk category, and the share under neither.
- Buyer task: the specific decision a buyer would grade, and the acceptance test agreed before delivery.
Methods note: how the scorecard and the field tally were built
The scorecard rates six insurance workflows on criteria defined in this article, using sources checked on October 27, 2026: California's claims regulations, the CARIN Blue Button guide, CMS guidance on the Patient Access API and the prior authorization rule, HHS's Privacy Rule summary, EIOPA's April 2026 letter and technical annex, NAIC survey reporting and O*NET's occupation profile. Ratings are editorial and describe workflows in general, not any company.
The field tally parsed the FHIR Shorthand source of the ten explanation-of-benefit profiles in HL7's carin-bb repository, at tag 2.1.0 and at main commit 09f7bde, and recorded the minimum cardinality of each named slice. Commented-out slices were ignored, and each full profile inherits its base profile's slices. The script needs only the Python standard library and is filed with this article's review materials.
Not measured: the share of real claims that carry reason codes, the share of claim files with complete notes, buyer demand by workflow, preparation cost or any price. No SourceX supply, pipeline or deal figure is used or implied.
Frequently asked questions
Why is claims volume a weak signal for insurance AI data?
Volume counts events, not explained decisions. In the claims format CMS suggests for the Patient Access API, payment status is required on every record while reason codes are optional in all ten profiles. A large claims count can therefore hold few records a model could be graded on.
Which insurance workflows leave the most traceable decisions?
Health prior authorization is gaining fastest, because CMS-0057-F requires impacted payers to give specific denial reasons from 2026. Property and casualty claim files also keep detailed documentation because regulators examine them. Both carry high sensitivity, so traceability alone does not make a workflow investable.
Does HIPAA cover auto and property insurance claim files?
HHS's Privacy Rule summary excludes plans that cover only workers' compensation, auto insurance or property and casualty risk from its definition of a health plan. Medical details in those files are therefore not health plan records under HIPAA, though other rules may apply. This is general information; counsel should review any specific dataset.
How does the EU AI Act classify insurance AI systems?
It lists AI used for risk assessment and pricing of natural persons in life and health insurance as high-risk. In April 2026 EIOPA asked EU institutions to exclude generalized linear and additive models from that category, saying the risks come mainly from data selection, governance and pricing policy. When this article was checked, that request was a proposal, not adopted law.
What should an investor ask an insurance data supplier first?
Ask what share of the offered decisions carry a coded reason and a linked outcome, with definitions you fix in advance. Then ask how third parties named in each record were handled in the rights review. Record counts and premium volume come last.
Sources
- O*NET lists claims adjuster tasks that end in decisions: determining coverage, judging liability from reports and records, interviewing before settlement or denial, and entering payments and reserves. Source
- California's Fair Claims Settlement Practices Regulations require claim files to hold all documents, notes and work papers for each claim, with dates of receipt and processing. Source
- The CARIN Blue Button profiles require benefit payment status on every EOB but leave reason codes and the claim received date optional. Source
- CMS has suggested payers consider the CARIN IG for Blue Button for the Patient Access API. Source
- CMS-0057-F requires impacted payers to give specific reasons for prior authorization denials from 2026 and to publish metrics, first by March 31, 2026. Source
- HIPAA health plans include health insurers and HMOs; plans covering only workers' compensation, auto or property and casualty risk are excluded. Source
- EIOPA's technical annex notes the AI Act's high-risk listing for life and health insurance pricing and locates the main risks in data selection, data governance and pricing policy. Source
- In April 2026 EIOPA asked EU institutions to exclude generalized linear and additive models from the AI Act's insurance high-risk category. Source
- NAIC surveys found 88% of personal auto, 70% of homeowners, 58% of life and 84% of health insurers use, intend to use or are exploring AI. Source
- In the NAIC life survey, 94 of 161 responding companies used, planned to use or planned to explore AI or machine learning. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.