AI-native TPA

A third-party administrator is a firm hired to run an insurance program's claims, without itself carrying the underwriting risk. A self-insured employer that funds its own workers' comp losses still needs someone to intake claims, set reserves, authorize treatment and issue payments; rather than build that operation in-house, it hires a TPA to do it under contract. An insurance carrier can do the same thing, outsourcing part of its claims book to a TPA instead of staffing it directly. Either way, the TPA examiner working the file does the same underlying work as a staff adjuster (reading records, applying policy and statute, deciding what is owed and when) but the employer relationship sits one layer removed from the entity actually bearing the loss.

That layer changes incentives in ways worth naming plainly:

A TPA that handles a file inconsistently with its client's instructions risks the contract as well as the claim outcome, which is part of why TPA licensing and oversight runs through a separate regulatory track from individual adjuster licensing.

The Bureau of Labor Statistics does not break its claims-adjuster occupational code (13-1031) out by employer type, so there is no published national count of TPA-specific examiners separate from carrier staff adjusters. What is countable is the demand side: how many organizations choose the self-insured, TPA-administered model over buying a standard insurance policy. That is where this guide starts.

Self-insurance is the demand engine behind the entire TPA industry. An employer large enough to absorb its own claims volatility can fund workers' comp losses directly instead of paying an insurance carrier's premium and margin, and then hire a TPA to administer the claims that funding has to pay. SIIA, the Self-Insurance Institute of America, reports more than 6,000 US corporations currently running a self-insured workers' comp program on this model.

The pattern shows up on the health-benefits side too, where KFF has tracked it longest: the share of covered workers in a self-funded health plan has sat inside a 65-to-69-percent band for a decade rather than trending consistently up or down. Workers' comp self-insurance does not publish an equivalently precise, regularly updated national participation rate; the honest summary available from named sources is that most, and by some counts nearly all, states permit employer self-insurance for workers' comp, without a single confirmed count of exactly how many. This guide is not going to round that up to a specific number no source published.

What matters for a TPA examiner is what self-insurance does to the incentive structure on a file. The employer funding the claim is frequently a named client the TPA's account team talks to directly, sometimes an employer sophisticated enough to have its own risk manager reviewing reserve decisions. That proximity is part of why TPA claims-handling instructions tend to be more explicit, and more frequently audited, than a standard carrier's internal claims manual.

Every TPA in the market now says it uses AI. The phrase that separates them is doing real work, and it is worth pinning down before reading another vendor page. Five Sigma, a claims-platform vendor, published the most rigorous public definition in August 2026, and its test is 1 line long: can the AI complete an action, or only recommend one?

That question sorts the market into 3 groups. A point solution does 1 task, well, and then hands the claim back. A claims system with AI bolted on assumes, in its architecture, that a human executes every action, so the AI can suggest but never finish. An AI-native platform builds the agents and what Five Sigma calls the operational rails together: the workflows, reserve authority rules, payment infrastructure, compliance frameworks and audit trails an action has to run through to actually happen. Without those rails, an agent hits what the same piece names an action ceiling, doing most of a workflow and then stopping because it has no authority to complete the last step.

Five Sigma's 5 criteria are worth carrying into any vendor conversation, because each one is checkable rather than rhetorical: agents and rails built together; domain-specific claims intelligence proven on production volume rather than in a demo; a workspace that works for humans and agents at once as the AI share of a book moves from roughly a fifth to roughly four fifths; monitoring and governance covering accuracy, confidence thresholds, model drift and exception patterns across the whole book; and a learning loop that captures adjuster corrections so the system compounds instead of plateauing.

The definition is not unique to claims, which is a useful signal that it is describing something real. Concirrus published the same split for underwriting in February 2026, arriving at it independently: AI underwriting means point solutions that speed up tasks, AI-native means intelligence built into the architecture so the decisions themselves change. Their line is the cleanest statement of the distinction anyone has published: what defines the category is not branding, but architectural intelligence embedded into operations. Both sources are vendors with an interest in the category existing. The definitions still hold up, because they are falsifiable: you can put a specific product in front of the "complete or recommend" test and get an answer.

The AI-native TPA is not a product category that emerged from inside the insurance industry. It was named from the outside, by venture investors describing a broader shift, and the TPA happened to be one of the clearest examples they could point at.

Emergence Capital says it named the category AI-Native Services in 2024, defining it as a firm that uses AI to deliver complete services rather than software, so that the customer buys an outcome rather than a tool. Its published investment criteria for the category are 6:

• Own the outcome.
• Build AI into delivery.
• Prove a measurable advantage.
• Increase AI leverage over time.
• Earn domain trust.
• Choose workflows that compound.

Its portfolio description of Strala, one of the firms covered later in this guide, is the category's thesis in a sentence:

An AI-native TPA where a dozen engineers process claims more accurately than thousand-person incumbents.

Emergence's own economic claim for the model is that it can deliver 5 to 10 times the speed or throughput of a labor-based operation while holding gross margins above 50 percent.

Sequoia Capital published the same argument from a different angle in March 2026, in a piece called "Services: The New Software." Its framing separates selling the tool from selling the work: if you sell the tool, you are in a race against the model, and every capability the model gains erodes what you charge for. If you sell the work, every improvement in the model makes your service faster, cheaper and harder to compete with. The same piece notes that for every dollar spent on software, 6 are spent on services, which is why the revenue pool on the services side is the one worth attacking.

Sequoia's opportunity map sizes claims adjusting, explicitly including TPAs, at 50 to 80 billion dollars, and its description of why is the most direct statement of the opening anyone has published:

• Standard-line claims get settled mechanically. By interpreting policy language against damage schedules and setting reserves from actuarial tables.
• The adjuster workforce is aging out, with nobody replacing it.
• The market is already massively outsourced to independents and TPAs.

That last clause is the strategic point. Sequoia's wedge logic is that replacing an outsourcing contract with an AI-native provider is a vendor swap, while replacing headcount is a reorganization. Vendor swaps get signed by a procurement team. Reorganizations require a board.

For anyone inside a TPA, the capital flowing into competitors is a bet that the administration contract itself changes hands. A TPA reading the category as a technology-adoption question is reading it one level too shallow.

There is no trustworthy single number for the size of the TPA industry, and the guide is going to say that plainly rather than pick the most flattering estimate. 3 published figures circulate, they measure different things, and the gap between them is larger than most national insurance markets.

Reading those 3 rows as a range would be a mistake. They measure different geographies, different line coverage and different revenue definitions, which is exactly why this guide cites Business Insurance's firm-by-firm ranking instead of a market-size number when it needs a fact about scale. What the market-research estimates are useful for is direction rather than magnitude, and on direction they are specific enough to be worth reading.

Mordor's own forecast puts the fastest growth in the industry inside AI-enabled TPA platforms, at 13.75 percent a year through 2031, against 7.36 percent for the market as a whole. It rates the market's concentration as low, which matters for anyone assuming the incumbents are unassailable. And it names its own restraints with estimated drag on the growth rate, which is the most useful part of the report for a TPA operator because each one is a business risk with a name attached.

One geographic note is worth carrying because it contradicts a common assumption. Mordor's own reading of North America, still the largest region at 29.39 percent of the global market, is that its growth is slowing relative to the post-ACA period because large-employer self-funding has effectively saturated. Incremental growth now depends on the mid-market and on smaller employers. A TPA whose commercial model is built on winning large self-insured accounts from other TPAs is competing for a pool that is no longer expanding.

There is no standard TPA fee schedule, and no single reliable published national average. The industry's own reference literature is explicit about this: IRMI, the International Risk Management Institute, uses $1,000 per claim as an illustrative example in its cost-modeling guidance, explicitly not as a survey result or market average, because actual rates are negotiated per contract and vary with claim complexity, service scope and client volume.

Glenn Backus, writing for PRIMA Central (the Public Risk Management Association), put a real number behind that variability with a named side-by-side comparison: one TPA's per-claim fee running at roughly double another's for broadly comparable work. Fee structures generally fall into a few recognizable shapes rather than one standard:

• A flat fee per claim, regardless of size.
• A percentage of paid losses.
• A flat annual retainer covering an expected claim volume.
• A hybrid combining a base fee with per-service add-ons for specialized work like litigation management or bill review.

A buyer evaluating TPAs will find no reliable target number to negotiate toward. The lesson is that fee structure alone tells you little about service quality, and a very low per-claim fee on a lean staffing model is one of the mechanisms that pushes an examiner toward the high end of the caseload range the adjuster guide in this series documents, with the attention-rationing failure mode that comes with it.

Business Insurance publishes the industry's most-cited annual ranking, "Largest Third-Party Administrators," ranked by prior-year gross revenue. In its 2026 report, Sedgwick Claims Management Services ranked first by a wide margin, with confirmed 2025 gross revenue of $5,085,596,881, up from $4,808,852,319 the year before. The other names that come up repeatedly in claims-industry coverage, alongside Sedgwick, are named below.

This guide confirmed Sedgwick's revenue figures directly against Business Insurance's own published ranking. It did not confirm a comparably precise, dated revenue figure for the other named firms from a source specific enough to stand behind: Arthur J. Gallagher & Co.'s FY2024 "Management" segment revenue ($11,384.3 million) almost certainly blends other business lines beyond Gallagher Bassett's claims work, so citing it as "Gallagher Bassett's revenue" would overstate precision this guide does not have. Where the ranking beyond Sedgwick's row cannot be confirmed against a source this guide trusts, it says so rather than estimating.

A short list of firms now describe themselves, in their own words, as AI-native TPAs or as AI-native claims operations. They are small, they are recent, and almost every number they publish is self-reported with no methodology attached. This chapter lists what each one actually says, marks what is missing, and does not convert any of it into a claim this guide would stand behind.

2 patterns run through that table:

• The outcome figures cluster around cycle time, review speed and coverage, not accuracy. Every firm publishes how much faster it is. Almost none publishes how often it is right, which is the number a claims organization actually needs, and its absence across an entire category is the single most useful thing this chapter has to report.
• The strongest claim in the table is about a workflow stage rather than a decision. Veltha's positioning, that a horizontal platform can summarize a claim file but cannot tell you which provision triggered which adjustment with citations, is the same argument this guide makes throughout: the defensible use of AI on a claims file is locating and citing, with a human holding the decision. It arrives at that position from a product-differentiation motive rather than a compliance one, and lands in the same place.

Among the named TPAs, the depth of public confirmation varies sharply, and this guide is explicit about the difference rather than treating every vendor claim as equally solid.

The pattern across the confirmed rows is consistent: document intake, summarization and correspondence triage are the workflow stages every named adopter has automated first, the same mechanics-not-judgment stages the workflow chapter later in this guide recommends starting with. None of the 6 firms' public materials describe AI as making a compensability or denial decision unsupervised.

The case against the AI-native challenge is worth reading in its strongest published form rather than its weakest. Hexaware, an IT services firm that sells into insurance, published the fullest version in April 2026 under the title "Why Third-Party Administrators Remain Indispensable Even in the Agentic AI Era," authored by its VP of insurance solutions, Shailendra Deo.

Its central argument has real force and does not depend on any of its numbers. A TPA sees claims across many clients at once. That multi-client exposure produces a data position on frequency, severity, fraud patterns and provider behavior that no single carrier can assemble from its own book, which is a network effect running in the incumbent's favor rather than the challenger's. Hexaware builds a 9-part value stack on top of that:

• A strategic AI posture.
• A layered automation architecture, applying rules engines, robotic process automation, predictive models, generative AI and agentic orchestration selectively rather than uniformly.
• Human-in-the-loop orchestration.
• Specialized capability, such as computer vision and telematics.
• Demonstrated return at scale.
• The data moat itself.
• Services beyond claims, including risk-management information systems.
• Regulatory depth across jurisdictions.
• Partnership rather than replacement.

Its prediction is a hybrid: TPAs keep high-volume, low-to-mid-complexity claims, carriers retain complex and strategic files in-house, and carriers use TPA data and platforms to accelerate their own AI roadmaps.

The same chapter is the right place to record what the largest incumbents do not publish. Crawford & Company markets CoverAI, an AI system that automates the policy-coverage review step and can run inside Crawford's own Digital Desk platform, inside another system, or standalone. Accuracy, volume handled and savings are all unpublished. Crawford's own TPA marketing carries no AI content and no quantified outcomes at all. That gap sits in the public record and says nothing against the product. It is also the reverse of the challengers' pattern in the previous chapter, where young firms publish speed figures with no accuracy behind them while large firms publish neither.

There is a 3rd posture beyond arm-the-incumbent and replace-the-incumbent, and it is the one a mid-sized TPA is most likely to end up in. Mazecare, operating in Hong Kong and Singapore, argues the vertical-integration case: if a TPA also owns the clinics, the clinical system and the claims system can be one system, so the clinical event and the claims event stop being 2 records with a handoff between them. No other source in this guide's corpus makes that argument. Its published numbers, all vendor-stated with no named customer or methodology, run to:

• 85 percent auto-adjudication.
• Routine-processing headcount down 70 percent.
• Processing time from 60 days to 3.
• 30 percent of fraud, waste and abuse blocked before payment.
• 99 percent claim-data accuracy.
• A 4-to-8-week initial deployment, plus a 2-to-4-week parallel run.

The existential question for the TPA model is whether the contract exists at all in 5 years, because the same AI that lets a 12-engineer firm process claims also lets a carrier bring administration back in-house.

V7 Labs framed the tension directly in May 2025: AI lowers the barrier for carriers to insource claims administration, which threatens the outsourcing model the entire industry rests on. The counter-argument is that a TPA reaches an efficiency and a data position a single carrier cannot. Both halves are true. What has been missing is a number.

The sharpest published one comes from a customer, not a vendor, and it points the wrong way for TPAs. Quentin Colmant, CEO and co-founder of Qover, on the record in Five Sigma's materials: "Thanks to insourcing the claim management solution at Qover and the platform of Five Sigma, I estimate that we've been able to reduce the unit claim cost by about 35% compared to the moment where it was outsourced to a TPA."

Read what that sentence actually is before using it. It is an estimate by the executive who made the insourcing decision, not a measured, audited figure, and it describes one company in 1 line of business. It is also the only public, named, quantified statement this guide found of a carrier claiming a specific unit-cost improvement from taking work away from a TPA, and it comes attached to a platform vendor that sells the capability to do it again. A TPA whose pitch is cost per claim needs an answer to that sentence, and the answer cannot be that the number is unverified, because the buyer on the other side of the table will have read it the same way.

Mordor's forecast names the same pressure independently, estimating that fee-margin pressure from insurers administering in-house drags roughly 0.7 percentage points off the industry's long-term growth rate. Hexaware's hybrid prediction concedes the same ground from the other direction: complex and strategic claims retained in-house, high-volume routine work outsourced. Every serious source in this guide's corpus agrees that some administration is moving back inside carriers. They disagree only about how much and how fast.

The strategic conclusion for a TPA is uncomfortable and specific. Competing on cost per claim against a carrier's own AI-equipped in-house operation is competing against an operation with no margin to earn and no client-acquisition cost to carry. The defensible ground is 2 things:

• The multi-client data position Hexaware describes.
• The audit trail a carrier's own compliance function would otherwise have to build itself.

Emergence Capital's "AI-Native Services Playbook," updated in March 2026, is the most useful operating document published about this category, and its central warning applies to an incumbent TPA adding AI just as much as to a startup.

The warning is a failure pattern it calls mirage product-market fit: revenue growing on human labor rather than on AI leverage. Its listed warning signs are all measurable from inside a business, which is what makes them useful:

• Gross margin flat or declining as revenue grows.
• Revenue per employee flat.
• Delivery still human-heavy.
• Bespoke work expanding rather than consolidating into product.
• No north-star product metric anyone tracks.

A TPA that added AI tooling in 2025 and grew revenue in 2026 by hiring examiners has mirage fit, whatever its press release says.

2 operating disciplines from the same playbook are worth importing directly:

• Honest cost accounting. Inference costs, model API spend and human-in-the-loop labor all belong in cost of goods sold, not in operating expense. Booking the human review that makes an AI output safe as overhead rather than as delivery cost is the single easiest way to make an unprofitable claims operation look like a software business.
• A north-star metric measured in minutes of human labor. The playbook's worked example, from the legal firm Crosby, is a measure it calls human review time: the minutes of human work required per document after the AI has processed it. As that number approaches zero, margins approach software margins. For a TPA the equivalent is minutes of examiner time per claim after automated intake, triage and drafting, tracked per client program and watched over quarters.

The playbook's other guidance is blunt and specific:

• Automate tasks rather than people, which is Strala's CEO Timon Gregg on the record in the same document.
• Put the people who do the work next to the people building the system, with feedback measured in hours rather than in batched evaluation cycles.
• Stay narrow, which is why Emergence describes Strala as focused specifically on claims processing for carriers, captives and MGAs rather than on insurance generally.
• Secure the right to learn from delivery data in the engagement letter itself, before the first claim is processed, because a data flywheel that starts on day one is the only moat the model has.

Against that, the base rate for enterprise AI programs is bad, and the honest version of it belongs in this chapter rather than in a footnote.

The number that reframes all of them comes from McKinsey, which estimates generative AI could add 50 to 70 billion dollars of value in insurance and, more usefully, that an end-to-end claims transformation delivers up to 14 times the impact of isolated applications.

14 separate pilots on 14 workflow stages is not one fourteenth of the value of a coherent rebuild. It is closer to none of it.

That is the mechanism behind the 95 percent figure above, rather than a separate finding.

Alvarez & Marsal published the sharpest strategic read on the industry in May 2026, under a title that does not hedge: "The TPA Reckoning: 24 Months To Reveal Who Remains Relevant in an AI Native Market," by managing director Tamseel Butt and director Chris Taylor.

Its argument overturns the industry's own growth playbook. For decades scale defined success in the TPA industry, headcount created leverage, and acquisitions rewarded size. Their claim is that this model is now outdated, and their reasoning is 1 sentence long: as carriers automate claims or adopt AI-native, outcome-based providers, labor-heavy operating structures shift from a competitive advantage to a margin liability. The section heading states it flatly. The roll-up is dead, because AI economics make bigger a liability.

That is worth sitting with, because private-equity-backed consolidation appears in Mordor's report as a named growth driver for the same industry in the same year. 2 credible sources, one treating roll-ups as fuel and the other as a trap. The difference is the time horizon: acquisitions still add revenue today and still add fixed labor cost against a falling price per claim tomorrow. A&M's closing line puts a clock on it. The old mantra was acquire or be acquired; the new one is thrive or try to survive, measured in the next 4 quarters rather than the next 4 years.

Their diagnosis of why AI programs inside TPAs fail is the most operationally useful part of the paper, and it is not a technology diagnosis. Most pilots fail to scale because of execution gaps. Their 6-stage lifecycle is a checklist a TPA leadership team can run against its own current pilots this quarter.

1. Validate the problem against client RFPs. Not against an internal wish list. If the capability is not something clients are already asking for in writing, the pilot has no commercial landing zone.
2. Select problem-led, not technology-led. Start from the claim workflow that hurts, rather than from the model that is available.
3. Empower a named operator. One person, named, with authority, not a steering committee.
4. Cap the proof of concept at 60 days. A&M's own number. A pilot that cannot show something in 60 days usually has too little scope, not too little time.
5. Commit before you scale. Tie commercial terms to key performance indicators before the rollout, so scaling is a contractual decision rather than an enthusiasm-driven one.
6. Hold accountability after launch. Post-launch measurement against those same indicators, which is the stage almost every failed program skips.

The paper closes with 10 lessons, a single strategic question per TPA archetype, and a 100-day readiness plan for TPA chief executives. Anyone running a TPA should read the original rather than this summary of it.

The clearest documented failure mode, and one that applies to a TPA's claims files exactly as much as it applies to a carrier's, is hallucination in generated summaries. A 2025 peer-reviewed study published in PLOS Digital Health found that 42 percent of GPT-4-generated summaries of emergency department encounters contained at least one hallucinated detail when checked against the source record, absent human review.

Nothing about routing a claim through a TPA instead of a carrier's own staff changes how a large language model behaves; the failure mode travels with the technology, not the employer relationship.

It bites hardest at the exact decision point regulators are watching most closely, the denial or payment-reduction letter, which is why NAIC's 2023 model bulletin on AI governance and several state bills, Florida's HB 527 among them, treat a documented human check at that step as the baseline. A generated summary can:

• Drop a qualifier.
• Misread a date.
• Infer a diagnosis from adjacent text.

That misread then propagates unread into a letter that goes out to a claimant, which is the median failure mode the hallucination literature describes, not an edge case. For a TPA specifically, there is a second-order risk worth naming: a TPA answers to a client relationship as well as to the claimant, and an AI-assisted error that damages that relationship threatens the contract on top of whatever the claim-level consequence is. The fix is the same one that applies everywhere else in this guide: a generated summary is a starting point an examiner checks against the source page, never a substitute for reading it.

1. Intake. Every incoming document gets logged with an arrival timestamp, ordered by date, deduplicated and paged before anyone reads it. Pure mechanics, no judgment, which is why it is the safest stage to automate first, and the stage every named TPA this guide confirmed an AI deployment for has already automated.
2. Triage flag. The file is checked against cost-risk signals (mechanism of injury, body part, early treatment intensity, litigation indicators) so the examiner's attention goes to the highest-risk file in the caseload first, not the one that happened to arrive first.
3. Record review, human-led. The examiner reads the flagged sections directly, using AI to locate and surface material rather than to conclude, before forming a view on compensability, causation or reserve adequacy.
4. Draft, not decision. A generated summary or draft correspondence is a starting point the examiner edits against the source page. It never goes out to a claimant or a client unread.
5. Sign-off and documentation. The person with authority to deny or reduce a claim signs it personally, and the file records which AI tool touched it and for what step.
6. Client reporting. Because a TPA answers to a client relationship on top of the claim itself, the same documentation trail that would satisfy a market-conduct exam doubles as the record a client's own risk manager can review, which is a TPA-specific reason to build the habit even where no regulator currently requires it.

The same 2-category split applies inside a TPA as it does at a carrier.

3 questions worth asking any claims AI vendor before a pilot, TPA-specific version:

• Does it cite the exact page and line it is summarizing? So a file can be checked in seconds rather than minutes.
• Does it distinguish between what the record states and what it infers?
• Can its output and audit trail be segmented cleanly per client? Because a TPA serves multiple client programs at once, so one client's data and documentation never bleed into another's reporting.

A vendor that cannot answer all 3 is not ready for anything that touches a denial letter or a client-facing report.

FurtherAI, an AI claims vendor, published the most specific statement of what a TPA needs that a carrier does not, in a 2026 buying guide updated in August. 4 requirements, and every one of them is a direct consequence of serving many clients instead of one.

1. Multi-client configuration in a single platform. Different rules, coverage structures, approval thresholds, correspondence templates and branding per client, with no cross-client data exposure. A carrier's system needs one configuration. A TPA's needs dozens that cannot see each other.
2. Client-level reporting as a product, not an export. Every client program needs its own performance view, on the client's terms rather than on the TPA's internal schema.
3. Proof rather than performance. Their line is worth quoting exactly: a TPA's audit trail is a commercial asset as well as a compliance record. The trail that satisfies a market-conduct exam is the same trail that wins a renewal, which is the argument this guide makes in its own quality chapter from the regulatory side.
4. Deployment speed as a sales weapon. A TPA that can stand up a new client program in weeks rather than quarters can bid on work a slower competitor cannot service in time.

The same guide is unusually honest about the state of the evidence in its own market, and the line is worth adopting as a standard: of the vendors it profiled, only 2 published quantified customer outcomes at all, and both were vendor-reported. Its own profile set is a useful map of who is actually selling into TPAs, with the caveats attached.

Set against that, the buying questions a carrier or self-insured employer should actually ask are shorter than any vendor's feature list.

1. Show me an action completing. Five Sigma's 1-line test, run live: can the system complete the action, or only recommend it? Watch the payment or the letter, not the recommendation.
2. Show me the citation. For any generated summary, the exact page and line it came from, and a visible distinction between what the record states and what the system inferred.
3. Show me the client wall. 2 client programs in one platform, with configuration, reporting and audit trail separated, demonstrated rather than described.
4. Show me the accuracy number, or tell me you do not have one. Almost nobody in this market publishes one. A vendor that says so plainly is telling you more than one that changes the subject to cycle time.
5. Show me where the human signs. Which step, which authority level, and what the file records about the tool that touched it beforehand.

A TPA's file answers to more scrutiny than a carrier's own staff-adjusted file, because it sits under 2 separate relationships at once: the claimant's, and the client's. 3 things put it under outside review:

• Litigation.
• A state market-conduct exam.
• The client's own audit of the program it is paying the TPA to run.

All 3 ask a version of the same question: what did the examiner know, when did they know it, and what did they do about it, with the client audit adding a 4th question specific to the TPA relationship: did the examiner follow this client's specific claims-handling instructions as well as general good practice.

The defensible version of an AI-assisted TPA file looks almost administrative:

• A timestamp showing when records arrived.
• A timestamp showing when they were reviewed.
• A note on what tool touched the file for what step.
• A human signature on any decision that carries legal or financial weight.

A client conducting its own program audit is typically looking for exactly this trail, which means the documentation habit a market-conduct exam would reward is the same one a client renewal conversation rewards. None of it is expensive to build; most of it is a habit, not a purchase.

This guide stops at the firm. The individual career track inside claims work, what the licensing and certification ladder costs, what the wage data says and what a first 90 days on a caseload looks like, is covered in the companion AI-native workers' comp adjuster guide rather than repeated here.

Copy this checklist into your own file-review process. Every item traces to a chapter above.

• Client-specific claims-handling instructions pulled before the caseload is touched
• Every incoming record logged with an arrival timestamp
• Records ordered by date and duplicates removed before anyone reads them
• High-cost-risk signals checked within the 1st review, not the 3rd
• Any AI-generated summary or draft checked against the source page before use
• AI tool use documented per client: what touched the file, for what step, reviewed by whom
• Denial or reduction decisions signed personally by the person with authority to make them
• Entity-level TPA registration and individual adjuster licensing both current in every state you handle claims for
• Every AI vendor put through the 1-line test: can it complete the action, or only recommend it
• Client configuration, reporting and audit trail demonstrably separated per client program, in practice
• Inference spend and human review labor booked to cost of delivery, not to overhead
• Examiner minutes per claim after automated intake tracked as a number, per client, over quarters
• Every AI proof of concept carrying a named owner and a hard 60-day decision date

Download this guide as a worksheet (CSV), with Done and Notes columns.

Questions readers ask

How is a TPA different from a staff adjuster at an insurance carrier?

A staff adjuster works directly for one carrier or self-insured employer. A TPA examiner works for a third-party administration firm the carrier or employer has hired to run its claims program instead of handling it in-house. Both do the same underlying work; the employer relationship, the entity-level licensing the firm carries, and often the caseload economics differ.

Is a TPA regulated the same way an insurance carrier is?

No. A carrier is licensed to underwrite risk. A TPA is licensed or registered as an administrator under a state's adoption of NAIC's Third-Party Administrators Model Act, a separate track that governs claims handling on someone else's risk rather than risk-bearing itself. The exact filing name and issuing authority vary by state.

Does a TPA need its own AI governance program, or can it rely on its clients'?

A TPA operating across states needs its own, because AI governance expectations are set at the state regulatory level, not delegated by an individual client contract. Rules vary enough by state that a TPA licensed nationally is effectively reading several different rulebooks for the same underlying AI use, so a TPA's compliance obligation is set by where it operates rather than by what any one client happens to require.

Will AI replace TPA claims examiners?

Every named TPA this guide confirmed an AI deployment for (Sedgwick, CorVel, Gallagher Bassett) has automated document intake, summarization and correspondence triage rather than the compensability or denial decision itself. That pattern, automating mechanics while keeping a human sign-off on judgment calls, is what every confirmed deployment shows today, whatever any individual vendor's roadmap claims.

Can an AI tool legally deny a workers' comp claim by itself inside a TPA?

No state currently requires a human sign-off on an AI-assisted denial by statute, though several have floated it. A TPA administering a denial on a client's behalf carries the same exposure a carrier would if a market-conduct exam or bad-faith claim later asks who actually made the decision, which is reason enough to keep a human signature on every one regardless of what the law currently requires.

What's the single biggest documented risk of using AI on a TPA claims file?

Hallucination in generated summaries. A 2025 peer-reviewed study in PLOS Digital Health found 42 percent of GPT-4-generated clinical encounter summaries contained at least one hallucinated detail absent human review. A TPA file carries an added exposure a carrier's own staff-adjusted file does not: an AI-assisted error damages a client relationship on top of the claim-level consequence, since a TPA answers to a client as well as to the claimant.

What makes a TPA AI-native rather than a TPA that uses AI?

The published test that separates them is 1 question: can the AI complete an action, or only recommend one? Five Sigma's August 2026 definition calls an AI-native platform one where the agents and the operational rails, meaning workflows, reserve authority rules, payment infrastructure, compliance frameworks and audit trails, were designed together, so an action can finish inside the system. A claims system with AI added assumes in its architecture that a human executes every action, so the agent hits what that piece calls an action ceiling: it does most of the workflow and then stops because it has no authority to complete the last step.

Which firms actually call themselves AI-native TPAs?

Strala, Elysian, inca and Veltha all use the phrase in their own materials, and Titan describes itself as an AI-native claims platform serving TPAs. Pace is adjacent rather than a TPA: it sells an AI operations platform to carriers, brokers, MGAs and TPAs. Every outcome figure any of them publishes is self-reported with no methodology attached, and none of them publishes a per-claim accuracy figure, which is the number a claims organization actually needs. The challengers chapter lists what each one publishes and what it does not.

Are carriers going to insource claims administration and drop TPAs?

Some already are, and there is one named, quantified public example: Qover's CEO Quentin Colmant estimates the company reduced unit claim cost by about 35 percent after insourcing claim management that had been outsourced to a TPA. That is his estimate rather than an audited measurement, and it describes one company. Mordor Intelligence independently estimates that fee-margin pressure from insurers administering in-house drags roughly 0.7 percentage points off the industry's long-term growth rate. The consistent reading across sources is that some administration moves back in-house while high-volume routine work stays outsourced; nobody credible claims the TPA disappears.

Does an AI-native TPA still need state TPA licensing?

Yes. Entity-level registration attaches to administering claims on someone else's risk, not to how the administration is performed, so automating the work does not remove the underlying filing in each state a firm operates.

How big is the TPA industry, in dollars?

No single figure holds up, and this guide names the disagreement rather than picking one. Mordor Intelligence estimates the global insurance TPA market at USD 592.52 billion in 2026 rising to USD 845.30 billion by 2031, and states plainly that the figures come from its own proprietary estimation framework rather than a census. Sequoia Capital sizes US claims-adjusting spend including TPAs at USD 50 to 80 billion, which measures one function rather than TPA revenue. A projection of more than USD 500 billion by 2030 for the US industry is attributed to Investopedia second-hand. They measure different things, so averaging them would be wrong. What is directly confirmed is Business Insurance's firm-level ranking, led by Sedgwick at $5,085,596,881 in 2025.

We did not survey TPAs, and we did not commission any research to write this manual. Every figure traces to a source we can name, and every one of those sources is public. A large share of them are vendors and investors writing about a category they have money in, which is stated on the line every time one is used.

What we did not do: we did not estimate a figure no source published. We did not settle on a single TPA industry market-size figure. Chapter 5 names the 3 published estimates side by side, says what each one actually measures, and explains why averaging them would be wrong; where this manual needs a hard fact about scale it uses Business Insurance's firm-level ranking instead. We did not restate any vendor's marketing percentages as findings: Hexaware's family of 40-to-80-percent ranges, Mazecare's auto-adjudication and accuracy figures, and every outcome number published by Strala, Elysian, inca, Veltha, Pace and FurtherAI are labeled on the line as the company's own claim. We did not convert Mordor Intelligence's estimates into facts, because Mordor itself says they come from a proprietary estimation framework rather than a census. We did not write anything from the LinkedIn posts, the YouTube video, the search-results page or the redirect links circulating about this category, because we could not retrieve them; where a company publishes no number, the challengers chapter says so rather than filling the gap. We did not cite an exact count of how many states permit workers' comp self-insurance; named sources confirm "most" or "nearly all" without a single agreed number. We did not attribute a generative-AI product to Broadspire that actually belongs to a different Crawford & Company business unit in the UK, after finding it during research and confirming the mismatch before writing this. We did not treat Helmsman's brief, search-indexed AI messaging as equivalent in strength to Sedgwick's, CorVel's or Gallagher Bassett's first-party product confirmations, and we said so directly in Chapter 7. We did not find a named ESIS generative-AI product to cite, and we are not implying one exists. We did not present the Business Insurance ranking beyond Sedgwick's confirmed row as verified, because we could not retrieve the full table from a source we trusted.

NAIC, KFF, SIIA, IRMI, PRIMA Central, Business Insurance, Sedgwick, Microsoft, CorVel, Gallagher Bassett, Crawford & Company, the Florida Legislature, Five Sigma, Concirrus, Emergence Capital, Sequoia Capital, Alvarez & Marsal, Mordor Intelligence, Hexaware, FurtherAI, V7 Labs, Strala, Elysian, inca, Veltha, Y Combinator, Pace, Titan, Mazecare, Qover, Spear Technologies, Riskonnect, DataGenix and Kognitos are not affiliated with Medrecords AI, have not reviewed this document, and do not endorse it. Their work is public and you should read it.