Cozmo AI’s gradient-boosted claims platform forces Verisk, Guidewire, Allstate to accelerate AI

Gradient-Boosted Trees on Messy Claims Data

Cozmo AI has deployed an AI-native claims platform at its anchor customer, a top U.S. restoration franchisor, using gradient-boosted decision trees on legacy Xactimate data to automate estimates during hurricane season — a live deployment that is forcing Verisk, Guidewire, and Allstate to accelerate their own AI initiatives.

The claims file arrives as a mess: line items shift columns when carriers update pricing tables, supplements drop in with new codes, adjusters type free-text notes into fields built for structured codes. The data is tabular, high-cardinality, and never clean. Cozmo, a five-person team founded in 2025 by Nuha Hashem and Alok Kumar in San Francisco, built its platform on gradient-boosted trees rather than large language models or tabular foundation models. The choice was deliberate. Boosted trees train in sub-second intervals on CPUs, infer in microseconds, require no GPU, and impose no row caps — constraints that matter when a founding team has no dedicated data engineers and needs to deploy on live claims during hurricane season.

"We work almost six days a week, in person in New York, and go-live weeks take the seventh day too," the team says. "We say this plainly because intensity is our structural advantage against incumbents with a hundred times our headcount."

Gradient boosting builds a sequence of weak learners, each correcting the errors of the previous one. XGBoost implements this with L1 and L2 regularization to reduce overfitting. CatBoost adds native categorical feature support that eliminates manual one-hot encoding and uses symmetric decision trees for faster prediction. LightGBM finished as the best classical model in a 2026 benchmark of 19 real-world datasets covering roughly 260,000 samples.

The benchmark, published by AIMultiple in July 2026, tested eight tabular learning models on datasets ranging from 435 to 48,800 rows using five-fold cross-validation with identical splits and known leakage columns removed. Every model ran on the same hardware. The results confirmed the boosted trees' case in 2026: operational, cost, simplicity, and deployment freedom.

On high-cardinality categorical data, the strongest classical models and Google's TabFM foundation model are statistically tied. TabFM won 15 of 19 datasets and sits 218 Elo points above the next model, 538 above the best gradient-boosted tree. No boosted tree won a single dataset.

But TabFM moves the cost to inference. Every prediction batch re-processes the training data as context. Served once per dataset this is fine; for repeated scoring on large data the per-call cost compounds where a trained XGBoost model predicts in microseconds. The full benchmark cost TabFM about $27 in GPU time; TabPFN 3 did the same work for $0.65 and LightGBM for effectively nothing on CPU. Tabular foundation models are pretrained once on hundreds of millions of synthetic datasets generated from structural causal models, then adapt to new datasets at inference time with no gradient training and no hyperparameter tuning. That architecture suits batch analytics. It suits production claims pipelines less.

Cozmo's data (Xactimate line items, carrier supplement codes, adjuster notes, photos mapped to estimate rows) lives in the regime where classical boosting still competes. The datasets are medium-sized, categorical-heavy, and schema-drifting. CatBoost's native categorical handling means the team can ingest new carrier code sets without rebuilding feature pipelines. LightGBM's speed means retraining on the latest hurricane-season claims happens between adjusters' coffee breaks. XGBoost's CPU parallelism means the same model trains on a laptop during a site visit. No GPU procurement. No MLOps team. The model generalizes across classification, regression, and ranking tasks with minimal feature engineering: claim severity, supplement probability, line-item mismatch detection all from the same framework.

The trade-off is accuracy on the largest datasets. On data exceeding 10,000 rows, TabFM averages 99.4 percent against XGBoost's 97.0 percent on large numeric data, and leads 88.4 percent to 87.6 percent on large hybrid data. The old pattern that boosting catches up at scale is gone. But restoration franchisors don't operate at TabFM's scale. They operate at the scale of thousands of claims per storm, each claim a few hundred rows, schema shifting weekly. The team that ships a model this week that adjusters trust next week wins. The team waiting for foundation-model inference costs to amortize loses the season.

That operational constraint — no GPU, no MLOps team, deploy this week — is why a five-person team ships while incumbents plan.

The Team and the Anchor Deployment

The name Cozmo AI carries an inside joke. The company shares its name with Anki's Cozmo 2.0, a palm-sized educational toy robot with a 2MP camera, capacitive touch sensor, and a personality designed to recognize pets and smiles. That Cozmo shipped in 2018. This Cozmo, founded in 2025 as a Y Combinator W22 company, builds AI employees that answer 2 a.m. pipe-burst calls, log loss photos into Xactimate, dispatch crews against SLA, and draft carrier-ready estimates from field photos. The through-line is not the hardware — it is the conviction that an agent must perceive, decide, and act inside the messy constraints of a real workflow.

The founding team is five people, backed by Y Combinator and Dubai Future District Fund, with headquarters in San Francisco and an EMEA hub in Dubai. Their hiring profile is specific: forward-deployed engineers (FDEs) who embed in customer operations, map how humans work today, re-engineer the process around AI, and deploy agents on top of existing systems (Xactimate, Zendesk, carrier portals, telephony) without demanding a migration.

The FDE role is three jobs in one. First, shadow adjusters and dispatchers to document the actual workflow — not the SOP, the reality. Second, redesign that workflow around what the model can reliably do: intake, triage, document chase, estimate assembly, settlement reconciliation. Third, ship the integration. Cozmo's Agent Protocol Engine (APE) sits above the agents, enforcing deterministic guardrails, shared state, and an audit trail on every action. The goal is not a copilot; it is a programmable workforce that runs the claim end to end while the human team works a live worklist organized around the one thing that needs a person next.

Recent hires signal the engineering gravity. Alek Mechkaroski joined as Staff Software Engineer in November 2025, bringing production-systems experience across international teams. Youssef B. joined the same month as Lead Engineer after building large-scale ad-tech infrastructure at TF1 Group handling millions of daily requests. Both are building the core infrastructure for voice-native AI employees that speak, think, and act in regulated enterprise environments, including sub-1.5-second latency across telephony, 110-plus languages, and compliance-ready for local sovereignty laws.

The team's origin in voice automation for Fortune 500 policy calls ("zero wait time and full compliance" in one global deployment) gave them a hard constraint: the agent must work inside the customer's existing telephony, CRM, and core platforms. No rip and replace. Carrier-grade SIP with in-country numbers across 80-plus countries. Native connectors. Live in days. That constraint forced them to reverse-engineer the data contracts of Xactimate estimate files, Zendesk ticket schemas, and carrier portal APIs, the same legacy surfaces the rest of the industry treats as immutable. They treat them as programmable interfaces.

The result is an operating system that runs both halves of a claims operation: everything the customer touches (intake, dispatch, field tracking) and everything behind the desk (estimate, settle, audit). Restoration franchisors (the anchor customer is one of the largest restoration franchisors in the United States) get a single console where every open claim shows the next human action required. The FDEs who deploy it don't hand off a dashboard; they stay until the workflow runs without them.

The US property restoration market sits at $52.79 billion in 2025 and growing, projected to reach $55.8 billion in 2026 at a 5.7 percent compound annual rate, per The Business Research Company. But the addressable surface for an AI-native claims platform is narrower and more concentrated than the top-line number suggests. A few hundred restoration franchisors, third-party administrators, and adjusting firms control the bulk of claim volume. Their executives know each other, convene at the same handful of conferences, and share collective scar tissue: most have been burned by software that promised automation and delivered a chatbot.

That network density makes the anchor customer decisive. Cozmo's anchor is one of the largest restoration franchisors in the United States, a single entity whose adoption signals permission for the rest of the cohort. In a market where boards have mandated "AI-native" transformation but internal teams lack the capacity to build it, a reference deployment at this scale functions as a de facto standard. The Y Combinator job posting for Cozmo's founding GTM role states it plainly: "Our anchor is that franchisor."

Cozmo reports 10-plus enterprise deployments, including Fortune Global 500 companies like Nestlé, with over 1 million production calls and 10 million interactions processed. Restoration teams go live in roughly seven business days (three weeks from pilot to production) without ripping out Xactimate, Zendesk, or existing phone systems. The platform captures 100 percent of actions with an audit trail, answers first notice of loss in any language 24/7, and triages and routes inbound claims in 12 seconds. Cost per successful resolution drops up to 70 percent. Outcome-based pricing means a call that goes nowhere costs nothing.

These numbers translate directly into the language the anchor's peers speak: cycle time, supplement approval rates, adjuster capacity during hurricane season. Cozmo's platform captures context from every claim it touches and gets sharper with each one. On the carrier side, the same work cuts cycle time to hours and leaves a line-item record of every loss, so risk can be priced by property instead of by ZIP code. For the franchisor, operators run more jobs per adjuster and send estimates that clear review the first time.

"Most AI tools are just prompt trees with a UI," Cozmo's YC profile notes. "They're fine for neat, repetitive tasks; but they fall apart the moment operations get messy, ending in half-filled tickets, off-script collection calls and disputes scattered across systems." The anchor customer's willingness to put Cozmo's agent system — autonomous workers that answer intake across channels, coordinate crews, and write up losses from first notice through supplement approval — into live production across the full claim lifecycle is the strongest signal the market has seen. It converts a board mandate into a working reference.

That reference cascades. In a network where executives compare notes at the same conferences, a single validated deployment at the top of the franchisor hierarchy rewrites the procurement calculus for the next tier. The market figure becomes a pipeline of identifiable accounts with a shared problem, a shared vocabulary, and now a shared proof point. Cozmo's founding GTM team is building directly into that concentration. The anchor isn't just a logo slide — it's the wedge that opens the rest of the market.

But the model is only half the battle. The adjuster still has to sign the estimate.

Trust, Not Accuracy: The Adjuster's Psychological Shift

The 2023 Atlantic hurricane season forced the issue. NOAA upgraded its prediction in early August from "near normal" to "above normal," and the claims machine shuddered under the volume. First Notice of Loss automation using chatbots and voice recognition cut initiation times, but the bottleneck moved downstream to the estimate. Adjusters who built careers on reading Xactimate line items, spotting missed damage on a roof walk, and negotiating scope with contractors suddenly faced AI-generated estimates that arrived carrier-ready in minutes. The technology worked. The trust didn't follow.

Negotiation research explains why. Canonical theory holds that success depends on balancing empathy and assertion — soft on the people, hard on the problem. Human negotiators perceive a tradeoff and behave as if they must choose. AI agents don't. In an international competition that ran over 180,000 automated negotiations, warmth (friendliness, sympathy, question-asking, gratitude) consistently produced superior outcomes across every metric: more value created, more value claimed, higher counterpart satisfaction. Dominance (assertiveness, firmness, longer conversations) claimed more value but produced more impasses. The warm agents reached deals at significantly higher rates. When deals closed, they actually claimed less value per deal, but the volume of agreements more than compensated. The research suggests a new theory of AI negotiation is needed, one that integrates classic interpersonal dynamics with AI-specific tactics like chain-of-thought reasoning and prompt injection.

Adjusters know this instinctively. Their leverage has always been the ability to say "I saw it" — to stand by the file when challenged. A LinkedIn post from Bluegrass Damage Appraisal in April 2026 put the fear bluntly: "BDA has officially replaced all field adjusters with AI. Our new system reviews losses using satellite imagery, auto-generates estimates, and closes claims in under 3 minutes, with no inspection required. Accuracy is… flexible. But efficiency is through the roof." The company later clarified they were building tools to support adjusters, not replace them: "No software is climbing a roof, catching missed damage, or standing behind a file when it gets challenged. Shortcuts don't win claims. People do." But the signal had landed.

During hurricane surges, the psychological shift required is not "trust the model." It's trust the model enough to sign your name to it when claim volume spikes. The adjuster's identity is wrapped in the negotiation: the back-and-forth with the contractor, the concession on one line to hold firm on another, the relationship that gets the next claim settled faster. AI that outputs a number without the dance feels like abdication. Research on personality engineering shows AI agents can be parameterized along warmth and dominance dimensions with exactness humans can't match, and they maintain their assigned profile across an entire negotiation regardless of counterpart behavior. But adjusters don't negotiate with parameterized agents. They negotiate with contractors, homeowners, and carriers. The warmth that wins in AI-AI negotiations (positivity, gratitude, questions) reads differently when a human adjuster deploys it against a seasoned roofer who knows Xactimate better than they do.

The operational shift Cozmo AI demands is subtle: adjusters become validators, not originators. They review the AI's estimate, flag the misses the satellite couldn't see, and add the line items that carry their professional signature. The estimate becomes a collaboration. The negotiation stays human. But that only works if the AI's baseline is credible enough that the adjuster's edits are surgical, not structural. During CAT season, when a desk adjuster handles heavy volume, credibility is measured in minutes saved versus minutes spent fixing. The psychological contract breaks the first time the AI misses a code upgrade requirement that costs the franchisor significantly on a single claim. Trust isn't built on accuracy metrics. It's built on the adjuster's willingness to stake their license on the output.

The incumbents know this. They're not waiting for trust to build.

Incumbents React: Verisk, Guidewire, Allstate

Guidewire took a platform bet. At its Connections conference, the company positioned agentic AI as a capability "embedded across underwriting, claims, pricing, and even developer productivity," according to its own materials. The new Agent Studio lets Guidewire Cloud customers "build, train, and deploy custom AI agents across underwriting, claims, and development, using the large language models (LLMs) of their choice." Guidewire's play is not a point solution but an orchestration layer: carriers already on Guidewire ClaimCenter get agent tooling without a new vendor. That neutralizes Cozmo's integration pitch for the installed base.

Zendesk, already entrenched in carrier support centers, adds generative ticket resolution to its seat base. For restoration franchisors using Zendesk for FNOL intake, the upgrade path is frictionless, with no new contract and no data migration. Cozmo's founding GTM hires (SDR headcount doubled year-on-year at AI-native firms, Growth Unhinged reported) now sell into a market where every major vendor has a credible AI claims narrative. The messy claims file that once choked adjusters is now the training ground Cozmo's competitors must catch up to.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.