Claims and quoting agents that recover from failures and leave a paper trail.

Your quoting agent pulls rates from six carriers, compares coverage, and generates a recommendation in seconds. In a demo. In production, carrier APIs time out, rate structures change mid-session, and you need a paper trail for every decision. Hatch handles all of it.

The problem.

Multi-carrier quoting agents fail in a specific pattern: the agent calls six carrier APIs in parallel, three respond in 400ms, two respond in 2 seconds, and one times out at 5 seconds. The agent retries the timeout. The carrier returns a rate — but it's from a stale rate table that was updated between the first call and the retry. The quote comparison is now internally inconsistent: five rates from the current table, one from 18 hours ago. The agent presents it as a valid comparison. The customer binds based on a rate that no longer exists. The carrier rejects the bind request.

Claims agents create liability the moment they trigger a disbursement without a documented approval chain. State insurance regulations in most jurisdictions require that claim payments above a threshold be authorized by a licensed adjuster. An agent that calls your payment processor directly — even with accurate fraud detection — is operating outside your E&O coverage if the payment authorization is not logged with an adjuster's explicit approval. This is not a hypothetical; it's the specific question your outside counsel will ask after the first disputed payout.

Workflow state is not durable in standard agent deployments. A claims agent processing document 47 of 200 in a due-diligence batch crashes when the document parser OOMs. On restart, the agent has no record of which documents it already processed — it starts from document 1. Documents 1 through 46 get processed twice: duplicate structured extractions written to the claims system, duplicate flags raised with adjusters, duplicate entries in the audit log. At 200 documents this is embarrassing. At 2,000 it corrupts your claims data.

What Hatch handles.

Hatch stamps each carrier API call with the rate-table version returned in the response header. Before presenting the final comparison, the agent validates that all six responses reference the same rate table version. If versions diverge — because one carrier returned a cached response — the stale call is retried with a cache-busting header and the comparison is rebuilt. The customer never sees a quote built from mixed rate tables.

Hatch persists agent state to a write-ahead log after each document is processed, keyed by document ID and workflow run ID. If the parser crashes at document 47, the agent resumes from document 48 on restart. Document IDs already processed are in the WAL; the agent skips them. No reprocessing, no duplicate extractions, no corrupted claims data.

Human approval gates on disbursements are enforced at the infrastructure level — not in application code that can be bypassed in a hot-fix. When a claims agent reaches the payout step, Hatch pauses the workflow and writes a pending-approval record. The workflow does not proceed until a licensed adjuster calls the approval endpoint with their credential token. The approval, the adjuster ID, the timestamp, and the claim amount are written to the immutable audit log. Payment processor calls are blocked until the approval record exists.

Carrier API failures are isolated per carrier. When one carrier's endpoint returns 503, Hatch marks that carrier's step as failed, continues processing the remaining five, and queues a retry for the failed carrier with exponential backoff capped at your configured SLA window. The partial result is visible in the workflow dashboard. If the carrier does not recover within the SLA, the workflow escalates to a human with the five available rates and a flag indicating which carrier is down.

Agents that run on Hatch.

Multi-carrier quote engine

Fans out rate requests to six carrier APIs in parallel over gRPC, validates rate-table version consistency across responses, ranks by price and coverage fit, and returns a structured comparison. Carrier failures are retried in isolation; the agent presents partial results with explicit flags rather than failing the entire quote on one carrier's outage.

500+ quote requests/day across multiple carriers

Claims processor

Pulls claims from an SQS queue, calls the document parser, extracts structured fields, runs underwriting rules against a rules engine API, writes extractions to the claims database with workflow-run provenance, and blocks at the payout step until a licensed adjuster's approval is recorded via the approval webhook.

200+ claims/day with full audit trail

Underwriting agent

Receives application submissions via a gRPC stream, calls third-party data providers (LexisNexis, ISO, internal actuarial APIs) with idempotent retry, runs the underwriting model, and writes a structured risk assessment with model version, input features, and output scores to the underwriting database for actuary review.

Continuous processing across all product lines

The 2-week PoC.

Take your existing quoting agent. Deploy it on Hatch. In two weeks, it completes multi-carrier quotes at production volume with rate-table version validation, isolated carrier retry, and a compliance-ready audit log. Zero inconsistent quotes from mixed rate tables.

500 quote requests/day with rate-table version consistency validated across all carrier responses

Carrier API failures isolated — one carrier down does not fail the quote; partial results returned with explicit flags

Full structured audit log of every quote: carrier responses, rate versions, ranking logic, and final recommendation

Human approval gates enforced at infrastructure level before any disbursement — adjuster ID and timestamp logged immutably

Why now.

NAIC's model bulletin on AI use in insurance (adopted by 20+ states as of 2024) requires that AI-driven underwriting and claims decisions be explainable and auditable. Several state insurance departments have begun requiring pre-filing of AI model documentation for personal lines. Carriers that can produce a step-level audit trace of every underwriting recommendation are clearing regulatory review in weeks; carriers that cannot are being asked to pause AI-driven decisioning until they can. The documentation burden lands on your ops team if you don't build the logging infrastructure now.

Have an agent stuck in staging?

Tell us what it does and where it's stuck. We'll scope a 2-week PoC and show you what production looks like.

book a call →