AI data market
Healthcare administration data investment: six workflows scored
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Healthcare administration data is worth backing for AI evaluation where a workflow has an outside grader, a standard format and a permission path that survives HIPAA, not where record counts are largest. Of six nonclinical workflows scored here, remittance and denial follow-up rates best, because the payer's remittance grades the work in a standard transaction. Prior authorization and coding have strong graders but carry clinical material.
Value drivers
Healthcare administration
Key takeaways
- Administrative work is gradable without medical judgment: a payer's response, remittance or authorization decision records whether the work was right.
- HIPAA's adopted X12 transactions give claims, eligibility, claim status, remittance and authorization a shared format across organizations.
- The permission chain is the asset: a business associate may de-identify patient data only as far as its agreement allows, and PHI may not be sold without authorization.
- Safe Harbor keeps only the year of patient-related dates, which strips the deadlines and intervals most administrative tasks are graded on.
- Sector size is not supply: about 200,700 medical records specialist jobs say nothing about which records anyone may license.
What would make healthcare administration data worth investing in?
Three things, all checkable: an outside party grades the work, the records share a standard format across organizations, and someone with authority can permit a de-identified copy to leave. Healthcare administration has unusually strong versions of the first two and an unusually strict version of the third, which is why the opportunity is real but narrow.
The case does not rest on diagnosis. Eligibility checks, prior authorization requests, claim coding, claim follow-up, remittance posting and scheduling are administrative decisions with recorded outcomes. An evaluation built on them asks whether an agent did the paperwork right, a question a payer's response can answer, rather than whether it practiced medicine well, a question it cannot.
Volume is the wrong yardstick. Every hospital and practice produces administrative records, but what an investor needs to know is how much of that output can be graded, joined across systems and licensed. The scorecard below rates six workflows on those terms and names the evidence each one still lacks.
Administrative records can grade an AI without judging medicine
They can, because the grade is usually written down by someone other than the worker. When a claim goes out, a payer answers; when a prior authorization is requested, a decision comes back; when a payment posts, a remittance says what was paid and what was not. Those responses are the answer key an evaluation needs.
HIPAA standardized the formats. CMS's list of adopted standards maps health care claims to the X12 837 transaction, eligibility inquiries and responses to 270 and 271, claim status to 276 and 277, electronic remittance advice to 835, and referral certification and authorization to 278. A test set built from one organization's transactions therefore uses fields that other organizations exchanging the same HIPAA transactions use too, which makes evaluation results unusually portable.
That portability is the investment-grade feature. Formats shared across organizations let a single evaluation design travel from one supplier to the next, and they let a buyer compare suppliers on the same fields. The open question is which workflows produce a grade worth having, and which can be licensed at all.
Six nonclinical workflows, scored on graders, formats and permissions
Remittance and denial follow-up scores best: an outside grader, an adopted standard and permission work limited mostly to dates. Prior authorization and claim coding have strong graders but weaker formats or heavier permission burdens, and scheduling is weak on all three counts.
The public evidence behind the ratings is uneven. On prior authorization, the AMA's physician survey counts nearly 40 requests per physician each week, taking physicians and staff about 13 hours, and a CAQH post on its 2024 Index says only 35 percent of prior authorizations are processed electronically, so most requests still do not move through the standard electronic transaction. On coding, a 2024 NEJM AI benchmark found that the best of four large language models, GPT-4, matched ICD-10-CM codes exactly 33.9 percent of the time, ICD-9-CM 45.9 percent and CPT 49.8 percent on a code-querying benchmark, a measured gap that evaluation data could help close.
Ratings use three criteria defined in the methods note: whether a party other than the worker records the outcome, whether an adopted HIPAA transaction carries the workflow, and how far the task depends on clinical material or patient-level dates. Each row ends with the evidence gap an investor should ask a supplier to fill:
| Workflow | Outside grader | Standard format | Permission and date fit | Evidence gap |
|---|---|---|---|---|
| Eligibility and benefits checks | Partial: the payer's response answers a lookup, but the task is mostly retrieval | Strong: adopted X12 270/271 | Partial: coverage dates belong to a patient | No public rate of eligibility answers later found wrong |
| Prior authorization | Strong: each request ends in a payer decision | Partial: X12 278 exists, but about 35% ran electronically | Weak: clinical attachments travel with the request | Share of decisions separable from clinical attachments |
| Claim coding | Strong: codes can be checked against reviewed or accepted codes | Partial: codes ride in the 837, but the source is clinical documentation | Weak: the task reads clinical text | No public benchmark pairing coding decisions with audit outcomes |
| Claim submission and status | Partial: status replies show where a claim stands; the outcome arrives later | Strong: adopted X12 837 and 276/277 | Partial: service and submission dates drive the task | No public count of status inquiries per claim |
| Remittance and denial follow-up | Strong: the remittance records what was paid; appeals add a second outcome | Strong: adopted X12 835 | Partial: filing and appeal deadlines need intervals | No public cross-payer measure of how denial reasons are recorded |
| Scheduling and appointment changes | Weak: visits kept or missed are recorded, but the right booking is often a preference | Weak: not among the standard transactions above | Partial: appointment dates belong to a patient | No public benchmark of scheduling decisions found |
The permission chain runs through business associate contracts
Administrative records are often held by one party on behalf of another, so the license can depend on a contract the holder did not write alone. When a billing or revenue-cycle vendor holds patient records as a business associate of a provider or plan, its rights over them are set by that agreement, not by possession.
HHS guidance is direct on the first gate. In an FAQ about health information organizations, it says a business associate may de-identify protected health information it holds on a covered entity's behalf only to the extent its business associate agreement authorizes, and that information properly de-identified is no longer PHI and may be used for any purpose, subject to other applicable law. HHS's sample contract provisions suggest the parties state whether the business associate may de-identify, how, and what it may then do with the result.
The second gate is the sale rule. Under 45 CFR 164.502(a)(5)(ii), a covered entity or business associate may not sell protected health information without the individual's authorization, and a sale means a disclosure in exchange for payment from the recipient. Data that has not been de-identified to the rule's standard therefore cannot simply be licensed for a price.
Owning a database settles none of this. Access to the records, permission under the privacy rule, the contract with the covered entity and the right to grant a license are four separate questions, and an investor should see documents for each. Counsel reviews the specific facts; this page describes the rule's text, not its application to any company.
Patient dates: where de-identification and evaluation collide
Most administrative tasks are graded against deadlines and intervals, and HIPAA's simplest de-identification method removes them. The Safe Harbor list in 45 CFR 164.514 requires removing all elements of dates except the year for dates directly related to an individual, such as admission and discharge dates.
That cuts into the workflows that score best. Whether an appeal was filed in time, how long an authorization took or whether a claim met a filing window are questions about days, not years. A Safe Harbor copy of a denial workflow keeps the payer's answer and loses the clock that made it right or wrong.
The rule's other route is a determination by a person with appropriate knowledge of generally accepted statistical and scientific methods that the risk of re-identification is very small, with the methods and results documented. For an investor, the practical test is whether a supplier can show such a determination for a date-preserving representation, and what it cost in time and review. A supplier that can only offer Safe Harbor copies of deadline-driven work is offering a weaker product than its record count suggests.
Does a big sector mean big supply? The strongest counterargument
No, and the gap between the two is the main reason to be cautious. The Bureau of Labor Statistics counts about 200,700 jobs for medical records specialists in 2025, people who use classification systems to assign codes for diagnoses and procedures, and the AMA counts hours of prior authorization work per physician every week. Neither figure says how many of the resulting records anyone may license.
Three filters cut the sector down. Contracts: where records sit with a business associate, its agreement may be silent on de-identification. Channels: with only about a third of prior authorizations processed electronically, by CAQH's account, the decisive evidence often sits outside the standard transaction, where it is expensive to capture and review. Content: coding and authorization records carry clinical material that a nonclinical evaluation must separate before anything leaves.
The counterargument also cuts the other way on demand. The BLS notes that AI tools for medical coding could affect demand for these workers, a sign that buyers are building systems for exactly these tasks. Labor intensity and model weakness argue for evaluation data; permission and capture cost argue that only a fraction of the sector will ever supply it.
Risks that could break the healthcare administration thesis
Four risks could shrink the opportunity, and each has a public marker an investor can watch.
- Permission stays shut. If business associate agreements in a supplier's book do not authorize de-identification, and covered entities decline to amend them, the gradable records exist but cannot be licensed. HHS's guidance makes the agreement the deciding document.
- Security incidents raise the cost of every copy. HHS's Office for Civil Rights opened an investigation of the Change Healthcare cyberattack in March 2024; incidents at administrative intermediaries invite closer scrutiny of any data leaving them, de-identified or not.
- Formats move. CMS's administrative simplification news page reports a December 2025 listening session on possibly replacing the X12 278 prior authorization transaction with HL7 FHIR. Evaluation sets built on one format may need mapping when workflows change.
- Buyers build from their own logs. Software vendors that already process administrative transactions for their customers may assemble evaluation data under their existing agreements, leaving less room for third-party supply. This is an editorial judgment, not a measured trend.
Seven questions before backing a healthcare admin data supplier
Ask for documents, not descriptions, and apply the same list to every supplier in the space, SourceX included.
SourceX sits between companies that hold operating records and the AI developers who license them: it finds and vets the supplying company, checks the rights, puts the contract in place, prepares and delivers the records and pays the supplier. No record moves until the supplying company has approved the scope and terms and signed. None of its supply, deal or financial figures are published; the company provides them to investors who request materials through its investor page, and any offering is limited to verified accredited investors under Rule 506(c). The questions below are the ones to put to it, or to anyone else:
- For each workflow you supply, who records the outcome that grades the work, and in which transaction or system?
- Which business associate agreements in your book authorize de-identification, and which uses of the result do they allow?
- Which de-identification method applies to each dataset, and for date-preserving formats, who made the determination and when?
- How do you separate administrative decisions from clinical attachments and notes before anything leaves the source?
- What share of prior authorization evidence in your datasets came through the X12 278 transaction rather than faxes, portals or calls?
- How do you handle a decision that was later appealed or overturned when it serves as a label?
- Which change, a format migration, a security incident or a contract amendment, would remove a workflow from your supply?
Methods note: criteria, sources and gaps behind the scorecard
The scorecard was built for this article on 2026-10-27. Outside grader: Strong when a party other than the worker records an outcome that decides whether the work was right; Partial when a recorded response grades only part of the task; Weak when the right answer is often a preference. Standard format: Strong when a HIPAA-adopted X12 transaction carries the workflow; Partial when one exists but public evidence shows much of the work moves outside it, or the source material sits outside it; Weak when none of the transactions named here applies. Permission and date fit: Weak when the task reads clinical text or attachments; Partial when it depends on patient-level dates; Strong was not assigned.
Sources: CMS's adopted-standards page and administrative simplification news, the eCFR text of 45 CFR 164.502 and 164.514, two HHS guidance pages on business associates and de-identification, an HHS news release, a CAQH post on its 2024 Index, the AMA's prior authorization survey page, a 2024 NEJM AI benchmark and the BLS Occupational Outlook Handbook. All were checked on their own domains on 2026-10-27.
Not measured: the volume of licensable records in any workflow, how often business associate agreements authorize de-identification, the cost of an expert determination, label error rates in payer responses, and any SourceX figures. The ratings are judgments on stated criteria; a reader with different weights should rescore the table, not trust the order.
Frequently asked questions
Is healthcare administration data the same as medical records?
No. Administrative data records eligibility checks, authorizations, claims, payments and scheduling rather than diagnosis and treatment, although some tasks, such as coding, read clinical documentation. When it relates to identifiable patients it is still protected health information under HIPAA, so the permission rules apply either way.
Can a billing company license de-identified data it processes for providers?
Only as far as its business associate agreement allows. HHS guidance says a business associate may de-identify protected health information only to the extent its agreement authorizes, and properly de-identified data is no longer PHI. Counsel should review the specific agreements and facts.
Why does remittance score higher than prior authorization?
Both have a payer decision to grade against, but remittance runs through an adopted X12 835 transaction and carries mostly administrative content. Prior authorization evidence often travels outside the 278 transaction, with only about 35 percent processed electronically by CAQH's account, and clinical attachments come with it.
What does the NEJM AI coding benchmark tell an investor?
That a measurable gap exists: in 2024 the best of four models matched ICD-10-CM codes exactly 33.9 percent of the time on a code-querying benchmark. It tests code querying, not coding from full patient charts, so treat it as evidence that evaluation data has a job to do, not as a forecast.
What should an investor ask a healthcare administration data supplier first?
Ask which business associate agreements authorize de-identification and for which uses, then which de-identification method each dataset relies on. Without those two answers, a large record count says little about what can be licensed.
Sources
- CMS's adopted HIPAA standards map claims to X12 837, eligibility to 270/271, claim status to 276/277, remittance advice to 835 and referral certification and authorization to 278. Source
- CMS's administrative simplification news page reports a December 10, 2025 listening session on possibly replacing the X12 278 prior authorization transaction with HL7 FHIR. Source
- A CAQH post on its 2024 Index says only 35 percent of prior authorizations are processed electronically. Source
- The AMA's prior authorization survey reports nearly 40 prior authorizations per physician each week, taking physicians and staff about 13 hours. Source
- A 2024 NEJM AI benchmark found GPT-4, the best of four models tested, matched ICD-10-CM codes exactly 33.9 percent of the time, ICD-9-CM 45.9 percent and CPT 49.8 percent. Source
- Under 45 CFR 164.502(a)(5)(ii), a covered entity or business associate may not sell protected health information without authorization, a sale being a disclosure in exchange for remuneration from the recipient. Source
- HHS says a business associate may de-identify protected health information only as far as its business associate agreement authorizes, and de-identified information is no longer PHI. Source
- HHS's sample business associate provisions suggest specifying whether the business associate may de-identify, how, and the permitted uses of the de-identified information. Source
- 45 CFR 164.514 sets two de-identification routes: removal of listed identifiers, including all date elements except year for dates directly related to an individual, or a documented expert determination that re-identification risk is very small. Source
- The BLS counts about 200,700 medical records specialist jobs in 2025, describes their coding work and notes AI coding tools could affect demand for these workers. Source
- HHS's Office for Civil Rights issued a letter and opened an investigation of the Change Healthcare cyberattack on March 13, 2024. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.