One AI Agent Handles 12,000 Claims a Year. Your Adjuster Does 1,300.

The Bet: $6.3 Million and a Structural Thesis

Polygrade emerged from stealth on October 6 with $6.3 million in seed funding, Business Wire reported, and a claim: the warranty industry has waited decades for its core-platform moment. Construct Capital, American Family Ventures, and Nine Four Ventures led the round. Their convergence signals more than a routine insurtech bet.

The founders ran the operations they now automate. David Steckel, chief executive, founded Setter, a home-management platform backed by Sequoia and later acquired by Thumbtack, where he built its home-services businesses. He then joined Sears Home Services as chief product officer, modernizing warranty integrations and direct-to-consumer products. Eui Chung, co-founder and chief technology officer, spent nearly 30 years at Sears and Transformco, most recently leading Sears Home Services' AI transformation. Kaustubh Vongole, co-founder and head of product, led fulfillment products and AI strategy. The trio initially named the company ProPay, reflecting a narrow focus on faster contractor payment, before expanding scope across the full claim lifecycle and adopting the Polygrade name.

Their thesis is structural: property and casualty insurance has Guidewire. Warranty never got an equivalent core platform. Most warranty companies run on systems built to log and route claims, not to decide them. Construct Capital cites $2.27 trillion in U.S. durable-goods spending in 2025 and $30.4 billion in warranty claims paid by U.S. manufacturers — figures that exclude extended and home warranties. That gap is the opening.

Polygrade targets home warranty first, an overlooked software category where the operating pattern is familiar: diagnose what failed, determine coverage, decide what should happen, obtain the right part or replacement, and pay the parties that performed the work. The platform supports five stages — first notice of loss, pre-dispatch triage, authorization, parts procurement, and payment, using specialized AI agents that operate within rules and thresholds each warranty company sets. Exceptions escalate for human review. The system sits above a warranty company's existing technology stack, turning data into actions and building a proprietary intelligence graph across policies, claims, decisions, and outcomes so each claim informs future decisions.

Early deployments are in production. Fidelity National Home Warranty uses the platform across its claims operations. Cinch Home Services works with Polygrade on appliance dispatch triage and parts sourcing. Supplier integrations include Encompass, UED, Sibi, and Marcone. The company reports, DevCuration found, roughly four in five homeowners engage over SMS after claim initiation, approximately 60 percent job-code prediction accuracy, and one in 13 submitted claims flagged as avoidable truck rolls through triage. Parts-order throughput has jumped from six to 60 orders per agent hour, according to DevCuration. The system also flags procurement overspend and claims where replacement may be more economical than repair.

"We built Polygrade to understand the claim and act on it," Steckel said. "The technology that exists is largely designed to keep track of a claim and move it through a workflow." Rachel Holt, cofounder and managing partner at Construct Capital, said the team's combination of domain expertise and technical experience is unusual: "They've run these operations at scale and understand why applying AI to a complex enterprise workflow requires much more than building a model."

The funding will expand forward-deployed engineering, go-to-market, and partner-enablement teams. Polygrade plans to move beyond home warranty into extended warranty, OEM warranty, and equipment-breakdown insurance — categories where the policy and customer relationship change but the operating pattern remains the same. The infrastructure for managing those claims, Steckel said, will look very different in an AI-native world. The first production deployments will determine whether that world arrives on Polygrade's terms or someone else's.

The $147 Billion Bottleneck

In 1966, German engineer Wilfried Reuter stored manufacturer replacement-part catalogs and repair-time tables on punch cards — the first serious attempt to automate collision estimating. The cards ran through mainframes the size of rooms. Today the same fundamental problem persists: a warranty claim still arrives as a pile of PDFs, photos, dealer invoices, and technician notes that a human must read, interpret, and reconcile against a contract written in legalese. The medium changed; the manual bottleneck did not.

Other estimates place the 2025 figure between $165 billion and $189 billion, but every major analyst agrees on the trajectory: double-digit growth driven by products that cost more to fix and fail in more complex ways.

The average new light vehicle in the United States now exceeds $48,000. Battery-electric packs carry replacement costs of $10,000 to $20,000. A mid-range smartphone repair has risen 15 to 20 percent in three years; HVAC repair costs in North America have climbed 25 percent since 2021. Flagship laptops breach $3,000; home automation installations top $5,000. Each price jump expands the financial exposure a warranty covers, and the incentive to game the system.

Fraud leakage is the industry's silent tax. Rising fraud and claims severity have pressured underwriting margins, forcing providers to invest in detection systems that still rely on rules engines and sampling. The OECD estimates counterfeit goods at 3 to 4 percent of global trade, roughly $500 billion annually. Consumer skepticism compounds the problem: roughly one in three extended-warranty buyers report difficulties getting claims fulfilled, suppressing repeat purchases and referrals.

Cycle time is the operational metric that exposes the architecture's limits. Copperberg's 2026 benchmarks show manufacturers who rebuilt the claims layer achieved 70 to 90 percent cycle-time compression and 30 to 50 percent processing-cost reductions, lifting first-pass approval rates from the mid-60s to above 85 percent. Extend, a warranty technology provider, targets a reduction from 72 hours to under four hours for eligible categories. Those numbers are not incremental — they are the signature of a process redesigned for autonomous execution rather than human acceleration.

The market structure amplifies the opportunity. Manufacturers control roughly 60 percent of distribution; individuals represent 74 percent of buyers; automobiles, mobile devices, and PCs together command more than half of revenue. Standard protection plans dominate at roughly 60 percent share. Online channels now capture 42 percent of volume and grow at 11 percent annually versus 7 percent for offline. Commercial and fleet segments are accelerating at 12 percent annually as electrification makes total-cost-of-ownership calculations warranty-sensitive. Every channel feeds claims into the same fragile, labor-intensive adjudication layer.

Legacy administrators (Assurant, Asurion, AIG, Endurance, AXA) sit on decades of claims data but run on platforms built for document routing, not decision automation. The paper trail is no longer paper; it is unstructured digital exhaust that resists the structured inputs an agent stack requires.

The Agent Stack: From First Notice to Contractor Payment

The warranty claims lifecycle has always been a relay race of handoffs — FNOL intake, coverage verification, damage assessment, fraud screening, repair authorization, contractor dispatch, payment release. Each handoff adds latency, error surface, and a human who must context-switch. Polygrade's agent stack treats the lifecycle as a single orchestrated workflow, assigning a specialized agent to each of its five stages while a governance layer maintains the audit trail regulators demand.

The press release and early deployments confirm the five-stage process outlined above. Specialized agents operate within company-defined rules and thresholds; exceptions escalate for human review, while the platform sits above existing technology stacks, converting data into actions and constructing an intelligence graph as previously described.

"We hold new technology to a simple standard: it has to improve the customer's claim experience right away and lower costs," said Jason Manns, EVP and COO at Fidelity National Home Warranty. "Polygrade accomplished both of those goals, while not requiring a wholesale replacement of the systems warranty companies depend on. We take our time, and we validate pilots before we scale. Polygrade stood out for bringing practical AI into our existing workflows, from smarter claims intake to parts procurement."

Early deployments have identified unnecessary service visits, repairs unlikely to resolve an issue, and procurement overspend. Cinch Home Services reports streamlined processes, improved efficiency, and solved operational problems on the parts side.

The stack runs on an architecture that researchers identify as table stakes for agentic maturity: a Data Layer for unified context, an Intelligence Layer for rule-aware reasoning, an Execution Layer for governed action, and a Governance Layer for audit and compliance. Governance added after deployment cannot be trusted. The Execution Layer manages omnichannel intake, routing each case to the appropriate resolution workflow, handling summarization and next-best-action recommendations for human agents, and escalating complex cases with full context pre-populated.

Engineers will recognize the frameworks: LangGraph for stateful graph execution, CrewAI for role-based agent crews, AutoGen for conversational agent patterns. But the differentiation is domain depth — the agents speak warranty. That operational context is the moat.

Why Legacy Administrators Cannot Bolt This On

The legacy TPA model rests on a simple equation: more claims require more people. Capacity is a function of headcount. For thirty years that equation rewarded mass: Carlyle Group acquired majority ownership of Sedgwick in a 2018 transaction valued at $6.7 billion, and the acquirer's playbook that built the industry's current champions will not build the next decade's winners. The TPA market has historically been controlled by large players able to keep entrants out due to their size, growing at 8 percent annually. In North America, TPAs still deliver a structural cost advantage over carrier-owned Administrative Services Only arrangements, with median medical costs roughly 30 percent below Blue Cross ASO comparables on a per-member-per-month basis.

But the architecture underneath that advantage was built for a different moment. Traditional third-party administrators operate on the principle that volume spikes after storms or during seasonal peaks lead to hiring waves or backlog build-up. Both cost money. Processing time stays dependent on human availability. At a classic TPA the priority is to clear cases, not to handle them well — adjudication quality and customer satisfaction are structurally last. The quality of handling fluctuates with onboarding, experience, and turnover. Fraud checking is structurally limited: comprehensive review of every single case is too expensive, so reviews are sampled. Small amounts are waved through because reviewing them isn't economical. That's not a failure of individual employees but a rational response to linear cost structures. Whoever outsources claims handling to a classic TPA ultimately buys an extended workbench: operational relief, but no different architecture.

AI-native TPAs invert the architecture. At inca, AI sees every case first. Before a claims expert takes over, AI has already checked coverage, analyzed documents, assessed fraud signals, and in many cases already triggered first communication to the policyholder. This isn't a tool that supports staff — it's a different architecture. The human is not the default but the exception, stepping in where judgment, empathy, or complex liability questions are required. Standard cases run through. Complex cases are escalated deliberately. One claims expert with AI support handles 8,000 to 12,000 cases per year, compared with 1,200 to 1,300 at classic handlers. A 24-hour turnaround per claim action is the standard. Higher adjudication quality emerges because AI reviews every case rather than samples; fraud patterns are detected systematically and claims leakage is reduced. Small amounts that get waved through manually no longer slip through.

Bolting a copilot onto legacy screens through screen scraping or spreadsheet uploads usually creates more risk than value. The assistant feels disconnected and brittle, and adjusters avoid it. In most carriers, the claims tech stack still exists to supervise, correct, and re-key what hundreds or thousands of human adjusters do by hand. Claims software does not remove that fragility — it accelerates it. A new claims core or workbench often becomes a faster interface for the same biologically constrained process. You still need tiered QA because a tired adjuster might approve the wrong estimate. You still need management layers to monitor dashboards and reassign backlogs after a catastrophe. You still absorb 100 percent of the operational risk because error and fatigue remain human. The result is a widening gap between the investment narrative ("We are digitizing claims") and the financial reality. Executives sign multi-year licenses for new claims platforms expecting double-digit improvements in combined ratio, but the gains stall at the margins. The hidden constraint is not the tool. It is that the entire operating model is still organized around human throughput, supervision, and correction.

The economics have flipped. For two decades, TPAs won on labor arbitrage: specialists processed claims cheaper than the carrier could do it itself. Agentic AI collapses that arbitrage. When the marginal cost of an adjudicated simple claim approaches zero, "we outsource because they're cheaper" stops being a rational answer. Independent analyses point to AI-driven cost reductions of up to 20 percent and cycle-time compression of up to 50 percent for well-scoped P&C claims workflows. Independent benchmarks on insurance-specific workloads show that a full triage and policy-limit verification flow can often be completed for around 3 cents in model and infrastructure costs when workloads are optimized across tiers of models. Many carriers still pay external BPO providers $10 to $20 per task for activities such as low-complexity intake, data validation, or reserve recommendations. In an anonymized property and casualty intake engagement, a managed operations provider reported reducing intake turnaround from 42.5 minutes to 4.2 minutes (a 90 percent cycle reduction) while quadrupling throughput per full-time equivalent.

The incentive structure inside most TPA boardrooms works against the urgency. EBITDA is still growing. Contracts are multi-year. Renewals are sticky. Sponsor-owned assets are on or near their hold-period exit windows. Incentives are almost perfectly engineered against the urgency the moment demands. A recent MIT-backed analysis found that roughly 95 percent of generative AI pilots fail to scale. North American insurers are set to spend an estimated $10.5 billion on core IT modernization between 2024 and 2026, and roughly 70 percent fail to meet their goals. Every one of those failure patterns applies to TPAs — in many respects more acutely because most TPAs are materially less mature than the carriers profiled. The most consistent failure pattern inside TPAs today is treating AI as a productivity add-on to an unchanged workflow. Current TPAs use AI to write financial reports faster, but the license-administration and adjuster-deployment processes underneath look identical to 2015.

Genuine leverage comes from redesigning the process around what AI makes possible — deploying adjusters dynamically based on complexity rather than provisioning licenses for every seat. This is an operating-model change, not a technology project. The wrap-do-not-rip approach offers a path: front-door the core with an API gateway and an event stream, emit canonical lifecycle events (fnol.received, claim.triaged, coverage.decided, payment.proposed, settlement.completed), and let AI agents and adjuster copilots subscribe to these events and respond by calling APIs. SageSure's AI-first modernization guide argues that API-first wrapping can cut IT costs per policy by around 40 percent and improve operational productivity when paired with workflow redesign. But even that pattern requires the carrier to contract for measurable outcomes instead of another software rollout.

The legacy administrator's operational context (venue-specific judicial considerations, regulatory nuance, decades of claims data) is a genuine asset. Allianz just transferred its entire commercial cyber book to Coalition, an AI-native specialist, because the specialist could execute on that context faster. The asset becomes an anchor when the organization tries to preserve the labor-arbitrage pricing model while layering AI on top. The threat to Polygrade's legacy counterparts is not that they lack data or domain knowledge. It is that their cost structure, their governance cadence, and their boardroom incentives are calibrated to a world where headcount was the moat. That world is ending. The moat is draining.

The Adjuster: Re-Skilled, Not Replaced

Every wave of automation in claims sparks the same worry: that the adjuster's job disappears. What actually happens is the job moves up, into the work that needs real expertise. Before AI, much of the adjuster's day went to work that had nothing to do with adjusting: keying data off documents, hunting down the police report, stitching a file together from a dozen attachments. AI-assisted tools took part of that load off with recommendations, auto-filled fields, and fraud flags. The adjuster still opened every file and made every call, just with less grunt work to get there.

Agentic AI flips the default. The administrative layer is now essentially gone in every operation that has deployed AI at scale. Carriers deploying agentic claims report 70 to 80 percent reductions in processing time on routine claims, with adjusters shifting to complex investigations and customer-sensitive decisions. At Resorts World Las Vegas, that shift came with a 33 percent efficiency gain in claims handling. Straight-through processing rates hit 30 to 50 percent of eligible claim volume. Loss adjustment expense per claim drops 30 to 80 percent on in-scope claims.

The work that vanishes is the routine file load: data entry, document keying, coverage checks, first notice of loss intake, standard document review. The work that remains is the exception queue: disputed coverage, tangled liability, policyholders who need to be heard. The volume an adjuster can carry goes up because the routine file load is gone, and the files that remain are the ones that need an expert. The adjuster's job gets harder in a good way: every claim on the desk is one the AI correctly decided it shouldn't close alone.

This is not a layoff story. It is a hiring-freeze story. Morgan Stanley estimates that 37 percent of industry roles, or about 2.2 million U.S. jobs, face agentic-displacement risk. Salesforce cut roughly 4,000 customer-service positions after AI agents began handling about half of customer interactions. IBM eliminated 200 HR roles after its agentic "AskHR" system automated high-volume workflows. McKinsey surveys show 32 percent of companies expect AI to decrease their employee base by at least 3 percent within the next year. Hiring has slowed to levels last seen in 2010, when unemployment was nearly 10 percent. Economists call it a "big freeze": companies are not firing, but they are not hiring either. Employment looks stable. Opportunity is not.

The impact shows up first in entry-level pathways. A November 2025 study by Erik Brynjolfsson and researchers at Stanford's Digital Economy Lab found a 16 percent decline in early-career employment across the most AI-exposed occupations since late 2022. Employment among developers aged 22 to 25 has fallen nearly 20 percent from its late-2022 peak. Software development job postings have fallen 53 percent from the same starting point. The first jobs to disappear are often outsourced: call centers, agencies, offshore support. Unemployment among recent graduates has climbed to nearly 6 percent, rising twice as fast as the rest of the workforce since 2022.

Meanwhile, 400,000 U.S. insurance professionals retire by end of 2026. The industry cannot backfill those roles through traditional hiring. The "100x claims team" model (a lean core of senior people orchestrating AI agents rather than adding headcount) becomes one of the most realistic goals to plan toward. McKinsey estimates agentic AI could deliver up to 90 percent productivity gains in core-system modernization.

The skill premium shifts. The old unit of human work was a caseload of claims you own end to end, measured by throughput, administrative accuracy, and process knowledge. The new unit is judgment and authority decisions plus ownership of defined claim segments, measured by judgment, negotiation, complex investigation, and oversight of the AI. Management and QA move from claims per adjuster and sample closed files after the fact to automation coverage, exception rate, override outcomes, live monitoring, and audit trail. Capacity models shift from headcount times claims per day to automation coverage plus human capacity for exceptions, sized for peak.

Governance becomes a dedicated function, not a side duty. Enterprise deployment needs separated roles: an operational owner who sets autonomy thresholds and runs day-to-day operations in claims operations; independent validation and challenge that tests models and outcomes without owning them, in model risk or a second line; technology and platform ownership that runs the system, integrations, and security; executive accountability — a named owner of claim outcomes and regulatory exposure. Combining rule-setting, monitoring, and assurance in one seat removes the check. Only one in five companies has a mature model for governance of autonomous AI agents.

Override rate is an ambiguous signal: a low rate can mean the recommendations are excellent, or it can mean people are rubber-stamping. Only sampling the non-overridden decisions tells you which. Build that review in from the start, because a regulator will ask how you know your oversight is real. Exception handling does not scale automatically. Bursts of correlated decisions after a catastrophe, an outage, or a model change can swamp the human queue. A decision queue with no capacity plan turns a resilient process into a fragile one.

The workable model is hybrid: automated progression and decision-based routing on top of persistent case ownership for defined segments. Specialists can see isolated questions without the claim's strategy, so decisions become locally correct but globally inconsistent. Accountability fragments when no one owns the outcome. Policyholders lose a stable point of contact, and files bounce between specialists. The carrier still owns the outcome legally and operationally, whatever the AI did.

For the experienced adjuster who worries about being automated away, the honest read is the opposite: the model raises the floor on what their time is worth. The job shifts from operating the process to owning the decisions and supervising the AI. The roles that emerge (AI operations managers, human-AI interaction specialists, quality stewards) signal a deeper shift: AI is now a structural component of how work is organized. The biggest impact of agentic AI on jobs will not be the layoffs we can see. It will be the opportunities that never materialize — the first steps into the workforce that quietly disappear before anyone notices.

Countermoves and the Next Milestone

The warranty claims arena is not waiting for Polygrade to prove its thesis. Two days after Polygrade's stealth exit, Asurion closed its acquisition of Domestic & General, instantly adding 4,000 employees, 22 million protected appliances, and 2.6 million annual repairs to a portfolio that now spans 12 markets across the UK, Europe, and the US — a $150 billion connected-home play that dwarfs Polygrade's $6.3 million war chest, according to DevCuration. The same week, Pace announced a $46 million Series B led by Sequoia and Thrive to build AI agents for insurers, and Klaimee raised $5.5 million in seed funding for an insurance-backed warranty product aimed at autonomous AI agents themselves. Armilla, which has been selling AI model warranties since late 2023 backed by Swiss Re, Greenlight Re, and Chaucer, brought its total raise to $7 million in February 2024. The capital intensity of the category is escalating faster than Polygrade's seed round can sustain alone.

Legacy administrators are not standing still either. CARS Protection Plus, a vehicle service contract provider, went live with Circuitry.ai's Warranty Decision Intelligence platform in May 2026 — just three months after pilot development began. The deployment is instructive: CARS didn't rip out its administration system. It layered an AI agent on top to solve a specific bottleneck (authorization notifications to repair shops) that its legacy vendor had failed to deliver for months. Circuitry's founder Ashok Kartham, who designed earlier warranty platforms at 4CS (acquired by PTC) and Mize (merged with Syncron), delivered the agent in a week. "For months, we had been asking our administration system to build us an easy way to send a notification out to a repair shop," said CARS' LaCoe. "We took that same problem to Circuitry, and a week later, we had an agent that creates this authorization notification." The pattern suggests incumbents will absorb agent capabilities as modular add-ons rather than suffer full-platform displacement — at least in the near term.

Polygrade's architecture bets on a different endgame. Those early deployments have flagged unnecessary service visits, unlikely repairs, and procurement overspend. But the company's stated plan to expand into those categories will test whether that intelligence graph compounds across fundamentally different claim types, contract structures, and regulatory regimes.

The production-readiness bar is specific and measurable. Gartner's 2025 benchmarks put legacy straight-through processing at 12 percent; manufacturers with mature AI deployments (18-plus months in production across most claim types) report 55 to 75 percent. Polygrade has not disclosed its STP rate. MarketsandMarkets found AI-native platforms now capture 34 percent of new enterprise warranty software purchases, up from 18 percent in 2023, but adoption skews heavily to manufacturers above $500 million in revenue (58 percent deployment rate) versus 14 percent below $50 million. Polygrade's home warranty beachhead sits in a segment where DartWG reports 40 to 50 percent resolution speed gains and 30 to 50 percent call volume drops, yet only 14 percent of major home builders use AI. The next milestone isn't another pilot. It's demonstrating sustained STP above 50 percent across a multi-tenant production environment that includes OEM and extended warranty claim types, with supplier chargeback recovery rates moving from the industry's 35 to 55 percent baseline toward the 58 to 72 percent that PwC's 2025 study found achievable with AI-powered recovery tools. LaCoe and Circuitry.ai's Ted Danner present their nine-month results at the MapConnected Service & Warranty Lifecycle Summit next week. That data (not Polygrade's seed announcement) will be the first independent signal of whether agent-native claims can clear the architectural threshold the legacy vendors have so far avoided.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at OpenAI and Anthropic, and the people building the field.