Owl
Owl Technologies builds AI that reads insurance claim files, for claims, SIU, legal and underwriting teams in life and health, disability, property and casualty, and government programs. OwlVision turns medical records, police reports, recorded calls and handwritten notes, across more than 220 document types, into a cited evidence graph that flags contradictions between documents. OwlAssist answers questions about the file and drafts reservation of rights letters, diary entries and memos in the team's own house style. OwlSignal matches outside data to claimants on open files every day, and OwlEnrich fills in claimant records through APIs.
Unum, Sun Life, The Hartford, Canada Life, Voya, Symetra, OneAmerica and WSIB are among its production customers. On Owl's own blinded benchmarks, OwlVision averaged 96% extraction accuracy against 71% to 73% for general frontier models, and OwlSignal scored an F1 of 0.92 against 0.62 for LexisNexis. A 45 day pilot with an unnamed US insurer estimated about four hours saved per general liability large loss report.
Owl's rule is that AI never denies a claim: an adverse determination needs a licensed human reviewer and a written rationale. Carriers tune which signals fire, at what threshold and to which queue, and connectors reach Guidewire, Duck Creek, Origami and Snapsheet. Each carrier runs in an isolated single tenant environment on AWS, claim files never train shared or foundation models, and fees are set in an enterprise order form. The founders, Sean and Sohrab Merat, lead a team spread across Vancouver, New York and Toronto.
Capability Axes
Capability grades
15 of 15 axes rated · 11 graded A or B
OwlVision turns each claim file into a cited evidence graph across more than 220 document types, OwlAssist answers questions and drafts letters from that graph, and OwlSignal and OwlEnrich match outside claimant data to open files. Take the models away and none of the four modules has work left to do, so the AI is the product itself.
Carriers set which OwlSignal signals fire, at what threshold and to which queue, by line of business, jurisdiction or claim cohort, and identity matching is calibrated to the carrier's tolerance for false positives. Insights arrive marked review indicated in the existing claims queue, adverse determinations go to a licensed human reviewer with a written rationale, and audit logs record every page accessed and every answer produced.
Extracted values carry page and paragraph citations with per field confidence scores calibrated against ground truth, and OwlAssist refuses an answer the file does not support. The published benchmarks use blinded holdout sets labeled by senior examiners: 18,400 documents for OwlVision and 12,800 closed claims for OwlSignal. An independent validation, model version schedule or drift monitoring report is not published.
Owl's own benchmarks put OwlVision at 96% average extraction accuracy against 70.8% and 73.2% for GPT and Claude class models, and OwlSignal at an F1 of 0.92 against 0.62 for LexisNexis. A 45 day pilot with an unnamed US insurer estimated about four hours saved per general liability large loss report. Named customers such as Unum and The Hartford appear without a measured result attached.
Claim files are never used to train shared or foundation models, a limit Owl says it enforces by contract. Any improvement on production data stays inside the customer's own tenant and happens only if the customer opts in. Files are returned or destroyed, with written confirmation, at the end of the agreed retention period.
Claimant files and account data get separate roles: Owl is processor or business associate for the files a carrier submits and controller for account and website data. Retention follows the carrier agreement, with return or destruction at offboarding and 14 months for analytics, and a public subprocessor list names Amazon Web Services for hosting. Claimant rights requests run through the carrier, and a BAA is available to covered customers.
An annual SOC 2 Type II audit backs the security program, with the report released under NDA, and the trust center at trust.owl.co lists it beside a penetration test and security policies. ISO 27001 is named as an assessment standard without a certificate, scope or certifying body, and a BAA is available to covered customers.
The terms limit the service to assisting qualified human adjudicators and leave claim decisions with the carrier, which places the product as claims review software. A license for Owl to adjust or adjudicate claims on an insurer's behalf is not identified, and the materials do not present one as needed.
Owl's stated rule is that AI never denies a claim: an adverse determination needs a licensed human reviewer and a written rationale, a commitment it says is enforced by contract and audited externally. Each OwlSignal insight shows its source, match confidence and the eligibility rule behind it instead of scoring the claimant. Subgroup error rates or an independent bias assessment of the flags are not published.
The public terms supply the service as is, defer enterprise liability and indemnities to the signed agreement, and cap aggregate liability at $100 where no agreement applies. Claimants are pointed to their carrier for privacy requests. Neither route gives a remedy for a wrong extraction, answer or flag.
OwlVision is described as a domain specific system rather than a prompt over a frontier model, built on a custom vision language model, and the public subprocessor list names only Amazon Web Services, for cloud hosting. Each carrier runs in its own tenant. The base model behind the custom build and its version history are not named.
Bidirectional connectors put OwlSignal and OwlAssist inside Guidewire, Duck Creek, Origami and Snapsheet, with OwlAssist also in Microsoft Teams, and insights land as work items in recovery, SIU, supervisor and legal queues. Outside data comes from ISO ClaimSearch, MIB, OFAC sanctions lists, Munich Re Risk Solutions and public record providers, refreshed daily through APIs.
Each carrier's files are processed in an isolated single tenant virtual private cloud on Amazon Web Services, the one entry on the subprocessor list, and Owl operates in the United States, Canada and the United Kingdom. A carrier can contract for claimant data to stay in a specified region. A hosting region for the product itself is not stated.
A carrier books a live walkthrough with a claims engineer on a closed claim file, and enterprise fees are set in the order form. Subscription amounts, usage rates, implementation costs and separate prices for OwlVision, OwlAssist, OwlSignal or OwlEnrich are not published.
Life and health, disability, property and casualty, health and government programs are the lines served, for claims, SIU, legal and underwriting teams. Production customers listed include Unum, Sun Life, The Hartford, Canada Life, Voya, Symetra, OneAmerica and WSIB, and the published pilot covered auto, general liability, property, professional and surety claims.
Alternatives to Owl
Other products that do the same kind of work as Owl, and what each one does.
Elysian handles insurance claims as a licensed third party administrator and sells the AI behind that work as software.
EvolutionIQ sells AI claims software to insurers handling disability, workers' compensation, auto casualty and income protection claims.
Snapsheet sells cloud based claims management software to property and casualty carriers, third party administrators, managing general agents and self insured organizations, and it reached that position from an unusual starting point.
Hi Marley runs the conversations around a property and casualty claim, connecting policyholders, adjusters, agents, repairers and other providers in one thread by text and voice.
CLARA Analytics sells casualty claims intelligence to insurers under the CLARAty.ai platform, applying document intelligence, predictive models and generative AI to medical notes, medical bills and legal demand packages across workers compensation, commercial auto liability and general liability.
ManageMy sells a digital front end to life and health and property and casualty insurers, covering sales, policy servicing and claims without replacing the systems that hold policy and claims records.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | Data Protection Terms | Implementation | Source |
|---|---|---|---|---|
|
Contact sales
|
Enterprise fees are set in an order form after a sales discussion and a live walkthrough on a closed claim file. | A business associate agreement is available to covered customers, with no pricing tier attached. | Implementation and onboarding fees are not published. | Vendor Published |
Enterprise fees are set in an order form. The security page offers a BAA to covered customers, and the public terms describe the business associate role for claim data, without a published tier or price.