Reserv Prices Its Claims AI at Compute Cost Plus 10%

Reserv CEO CJ Przybyl spun the AI-native claims administrator's own technology stack into a standalone licensing entity on July 21, 2026, pricing every model call at compute cost plus a 10% markup with no seat fees, minimums, or multi-year contracts (BusinessWire, July 21, 2026). The new unit, Apeiros Insurance Data Exchange, or AiDE, cleans a bordereaux file, categorizes the errors, and produces a correction plan for about $0.25, against roughly $13 for the legacy tools it displaces.

That is the number that makes this a pricing story rather than a product launch. AiDE is not a new claims model; it is the same stack that already runs Reserv's own third-party administration business, the one processing claims for the carriers, MGAs, brokers, and captives that make up its client base, now sold directly to anyone who wants to run it themselves. Reserv disclosed the metered pricing structure, the underlying model catalog, and per-task cost examples in a joint release with AiDE, LLC, its newly formed licensing affiliate (Finanznachrichten, July 2026, republishing the BusinessWire release). Przybyl framed the decision as a response to carrier demand rather than competitive pressure: "We are continually demoing our internal system to carriers that don't use TPAs, and they all say, 'I want that.' There is clearly something comforting about seeing tech productionalized inside of the globally diverse 600-person adjusting staff of RCA," he said, using Reserv's internal shorthand for its claims administration arm (BusinessWire, July 21, 2026). The launch arrives eleven weeks after Reserv closed a $125 million Series C led by KKR, a round the company said it did not solicit (fintech.global, May 2026).

What the License Actually Includes

AiDE's catalog splits into two tiers. Commodity features, priced purely on compute consumption, cover first-notice-of-loss automation, claim triage, subrogation detection, demand-letter detection, bordereaux ingestion and financial reconciliation (including Lloyd's Financial Conduct requirements), data extraction, and claim file quality assurance. A second, more expensive tier covers ad hoc, claim-by-claim work: large-language-model-driven data mapping and system rollover tools, adjusting logic and automation, natural-language claim interrogation, portfolio analytics, action-plan generation, and contextual guidelines analysis (Finanznachrichten, July 2026). A third layer, custom workflow, UX, and integration work, is billed separately as a capital expense rather than a runtime cost, which matters for how a carrier's finance department books the spend.

The stack underneath that catalog is not experimental. It has been validated across hundreds of thousands of live claims inside Reserv's own operation, which serves nearly 200 insurers, MGAs, brokers, and corporate captives through roughly 600 adjusting staff globally (BusinessWire, July 21, 2026; fintech.global, May 2026). That production history is the pitch: a carrier licensing AiDE is not buying a model with synthetic benchmarks. It is buying the exact software that already processes Reserv's client claims, unbundled from the labor that has historically come attached to it.

Compute Cost Plus 10%: What the Number Actually Discloses

Per-seat SaaS pricing has always obscured the marginal cost of running claims software, because a seat license bundles infrastructure, model inference, support, and margin into one flat number that never moves with usage. Compute-cost-plus pricing does the opposite. It publishes the marginal cost directly, then adds a fixed, disclosed spread on top of it. A 10% markup on token and inference cost is a materially different unit economics claim than the 60% to 80% gross margins typical of enterprise SaaS, and it is only defensible if the underlying compute cost is falling fast enough, and the volume is large enough, that a thin percentage still clears Reserv's cost of capital.

That distinction is not just a vendor-positioning detail. Actuarial Standard of Practice No. 29 directs actuaries to base expense provisions in property casualty ratemaking on the actual anticipated cost of performing the function being priced, not on a bundled industry average (ASOP No. 29, Actuarial Standards Board). A per-seat license obscures that anticipated cost inside a subscription fee that does not vary with claim volume, forcing an actuary building a claims-expense load to fall back on an allocated average. A metered, disclosed unit cost, $0.25 to clean a bordereaux file, a per-token charge for triage or subrogation flagging, is closer to the kind of activity-based cost data ASOP No. 29 contemplates, because it ties the expense directly to the transaction driving it. For a carrier's ULAE study, that changes the input from a fixed allocated overhead figure to a variable cost that scales with claim counts, which is a cleaner fit for the paid-to-paid and open-claim-count methodologies actuaries already use to project unallocated loss adjustment expense (Werner and Modlin, "Basic Ratemaking," Casualty Actuarial Society, 2016).

The absence of minimums or seat counts also removes a distortion that per-seat AI licensing otherwise introduces into expense studies: a carrier that licenses ten seats but only actively uses three during a slow claims quarter still pays for ten, inflating its measured expense ratio relative to actual claims activity. Metered pricing collapses that gap, so a quarter with fewer claims produces a proportionally smaller AI expense line rather than a fixed one, which is the behavior an expense provision is supposed to model in the first place.

The Book Behind the Product

Reserv's Series C gives the licensing pitch its scale argument. The company told investors it processes roughly 500,000 complex claims annually today and is targeting 30 million within four years, an expansion it says will let it automate a substantial share of the non-field-based commercial property casualty claims market (fintech.global, May 2026). Annual recurring revenue stood at $100 million at the time of the raise. KKR partner Patrick Devine, whose firm led the round alongside existing backers Bain Capital Ventures and Flourish Ventures, credited the company's execution directly: "What Reserv has done from an AI and operational perspective to deliver faster and better quality outcomes for its customers is truly differentiated in the market" (fintech.global, May 2026).

Reserv was founded in 2022 and has doubled its claims-processing capacity annually since, according to the company (fintech.global, May 2026). AiDE inherits that operating history as its main credibility asset, and as its main conflict. The stack was built to run Reserv's own adjusting book, not as a general-purpose product, which is why the company frames it as production-validated software rather than a platform designed from a blank page for third-party licensing. That is also why the launch reads less like a typical insurtech product announcement, most of which lead with model accuracy claims, and more like a TPA disclosing its own cost structure to the market it competes in.

Subrogation Detection and Bordereaux Reconciliation Meet Reserve Timing

Two of AiDE's commodity-tier features, automated subrogation detection and bordereaux reconciliation, sit directly upstream of reserve adequacy and loss-development data, which is where the licensing story becomes an actuarial one rather than a technology one. Salvage and subrogation recovery has grown from roughly 11% of claims paid in 1996 to about 20% in 2021, meaning insurers now recoup close to one dollar for every five paid in auto physical damage claims, and total industry recoveries across auto physical damage, commercial auto liability, and personal auto liability reached an estimated $51.6 billion in 2021 ("How's the Recovery? Salvage and Subrogation," NAIC Journal of Insurance Regulation, 2023). Against that base, roughly 15% of property casualty claims are closed without a valid subrogation opportunity ever being identified, an estimated $15 billion to $20 billion of missed recoveries industry-wide each year (NAIC Journal of Insurance Regulation, 2023). A model that flags subrogation potential earlier in the claim lifecycle, before a file closes, is not just a claims-efficiency story; it is a direct input to how much of a carrier's held reserve should be offset by anticipated recoveries, and how quickly that offset should be recognized in the reserve triangle rather than surfacing as a favorable development surprise years later.

Bordereaux reconciliation carries a parallel effect on the MGA and program-business side of the ledger. Reinsurers and carriers managing delegated-authority relationships depend on bordereaux data to track exposure and loss activity across program partners, and inconsistent formats and manual normalization routinely delay the reconciliation that catches reporting errors before they compound across a treaty year. An AI layer that standardizes and error-checks that data at the point of ingestion, rather than during a quarterly audit, shortens the lag between when a loss event occurs in an MGA's book and when it shows up correctly coded in the ceding carrier's own loss triangle. That lag is exactly the kind of reporting-pattern shift that can distort a loss development factor selection if an actuary is not aware the underlying data pipeline changed mid-triangle, a risk this site has flagged in the context of agentic claims tools reshaping ULAE and reserve data more broadly (see actuary.info's earlier coverage of agentic claims AI and ULAE reserve uncertainty).

AiDE feature tier	Example task	Cost signal
Commodity (compute-metered)	Bordereaux cleaning, categorization, correction plan	~$0.25 per file, vs. ~$13 for legacy tools
Commodity (compute-metered)	FNOL automation, claim triage, subrogation detection	Per-token compute cost plus 10%
Complex (claim-by-claim)	Data mapping, portfolio analytics, guidelines analysis	Higher per-task compute cost plus 10%
Custom services	Workflow, UX, and integration build-out	Billed as one-time capital expense

Selling to Your Own Competitors

The harder question is what it means that the vendor is Reserv's own third-party administration arm, selling the exact stack it uses to compete for the same claims-handling business. A carrier weighing AiDE against building an in-house pipeline is not just comparing a build cost to a license fee; it is deciding whether to hand its claims-expense data infrastructure to a company that also bids for its outsourced claims work. Przybyl addressed the tension directly, saying AiDE was structured to license Reserv's intellectual property "without grant-back data rights to Reserv," so that a licensee's claims data stays proprietary to the licensee rather than flowing back to strengthen Reserv's own competing TPA operation (BusinessWire, July 21, 2026). He also took a swipe at the incumbent AI vendor model: "The industry is currently being bombarded by 'workflow' engines, 'intelligence' layers, and so on that want minimums, SaaS fees, and multi-year contracts," locking buyers into commitments in a market changing faster than the contract terms anticipate (BusinessWire, July 21, 2026).

That data-rights structure is the detail a carrier's own governance program has to verify, not take on faith, because the NAIC's Model Bulletin on the Use of Artificial Intelligence Systems places the burden of third-party AI oversight on the insurer, not the vendor. The bulletin, adopted with little to no material change by 24 states as of March 2025, requires insurers to maintain a written program governing AI systems that includes assessing data and models supplied by third parties and securing contractual protections such as audit rights (Holland & Knight, May 2025). A compute-metered license with no data grant-back may satisfy that requirement more cleanly than a bundled SaaS contract that never separates model logic from data flow in the first place, but it still needs the same audit-rights and documentation review any third-party model vendor gets under an insurer's AI governance program, a point this site has raised in the context of vendor-patented governance controls limiting a carrier's own compliance flexibility (see when AI governance controls are themselves vendor-patented).

Reserv's pitch also lands inside a broader capital pattern actuary.info has tracked across 2026: insurtech funding has concentrated heavily in submission intake, underwriting, and claims workflow automation, with comparatively little going to standalone pricing or reserving engines (see the H1 2026 insurtech funding concentration analysis). AiDE is a variant on that pattern with a twist: rather than a pure-play startup selling only software, it is an operating TPA monetizing the infrastructure behind its own labor-heavy business line, the same move CCC and Sedgwick have each made in different forms as they scale claims AI inside existing adjusting operations (see actuary.info's coverage of CCC and EvolutionIQ's workers compensation claims AI and Sedgwick Omni's claims AI data advantage). The build-versus-buy calculus for a carrier evaluating any of these vendors is no longer simply cost versus control. It now includes whether the vendor's own claims operation is a customer, a competitor, or both at once.

What This Reprices for Reserving Actuaries

The immediate effect of compute-metered claims AI pricing is that a carrier's AI expense line becomes variable rather than fixed, which is a cleaner fit for existing ULAE ratemaking methods than the flat SaaS fees it replaces. The second effect is less obvious: automated subrogation flagging and bordereaux reconciliation change the timing of information that feeds case reserves and loss triangles, not just the labor cost of producing that information. A reserving actuary who treats a newly licensed claims AI stack as a pure cost-reduction event, without checking whether faster subrogation identification or cleaner bordereaux data has shifted the reporting pattern underneath a loss development triangle, risks misreading a data-pipeline change as a genuine improvement in underlying loss experience. AiDE's pricing model makes the expense side of that trade unusually transparent. It does nothing to make the reserving side of it automatic.