Who will stand behind an AI agent?
Most commercial insurance was written before generative AI and is silent on AI failure: it neither clearly covers a hallucinated output, a mis-executed transaction or a model that drifts, nor clearly excludes it. A market has formed to close that gap — specialist managing general agents, Lloyd’s coverholders, performance-guarantee programmes, independent evaluators and audit firms.
This index tracks that market with the same evidence standard as everything else here. Every claim on a vendor entry cites the document it came from and the date it was checked, and where a vendor does not publish something — its limits, its indemnification terms, its regulatory status — we record that instead of inferring it.
What we track
| Insurance availability | Whether cover can actually be written for this system |
| Carriers and programmes | Who is named as underwriting it, and in what structure |
| Coverage limits | Published policy limits, where a vendor states them |
| Indemnification | What the vendor indemnifies its customer against |
| Liability cap | Contractual limitation of liability |
| Regulatory status | Carrier, MGA, coverholder, broker, or unregulated |
| Underwriting backing | Syndicates, reinsurers, or capital behind the offering |
| Evaluation coverage | Whether an independent evaluator benchmarks it |
| Audit trail | Whether actions produce a tamper-evident, reviewable record |
| Settlement mechanism | Escrow, milestone release, or clawback on failure |
Twenty-one assurance fields in total, alongside the thirty-eight covering agency, safety, practicality, foundation models, ecosystem and impact. See the methodology.
Why nobody else covers this
The large AI agent directories organise by what an agent does: sales, coding, customer support, video, images. That is a useful way to shop and a poor way to assess risk. On insurance specifically — a named carrier, a managing general agent, a Lloyd’s coverholder or a syndicate — neither major directory lists one, checked repeatedly and still true. A handful of individual listings elsewhere touch compliance, audit or attestation function, but as an unclassified single line: no dedicated category, no coverage detail, no source citation behind the claim. Either way, the question a buyer actually has to answer before deployment — what happens when this thing is wrong, and who pays — gets a structured, cited answer only here.
Frequently asked
- Does standard commercial insurance cover AI agent failures?
- Most commercial insurance was written before generative AI and is silent on AI failure: it neither clearly covers a hallucinated output, a mis-executed transaction, or a model that drifts, nor clearly excludes it. A specialist market — managing general agents, Lloyd's coverholders, performance-guarantee programmes, independent evaluators and audit firms — has formed to close that gap.
- What does AI Dispatch track about AI insurance and assurance vendors?
- Insurance availability, named carriers and programmes, published coverage limits, indemnification terms, contractual liability caps, regulatory status (carrier, MGA, coverholder, broker, or unregulated), underwriting backing, independent evaluation coverage, audit-trail tamper-evidence, and settlement mechanism (escrow, milestone release, or clawback on failure) — 21 assurance fields in total, alongside 38 covering agency, safety, practicality, foundation models, ecosystem and impact.
- Why don't the major AI agent directories cover AI insurance?
- The large AI agent directories organise entries by what an agent does — sales, coding, customer support, video, images — which is useful for shopping and silent on risk. On insurance specifically, neither of the two largest directories lists a named carrier, managing general agent, Lloyd's coverholder or syndicate; repeated checks have found that gap total. A small number of individual listings elsewhere touch compliance or audit function, but as an unclassified single line, with no dedicated category, no coverage detail, and no source citation behind the claim — not the risk-transfer question of who pays when the system is wrong.
- How is a vendor's coverage or evaluation status verified before it's published?
- Every claim on a vendor entry cites the document it came from and the date it was checked, with the same evidence standard used across the rest of the index. Where a vendor does not publish something — its limits, its indemnification terms, its regulatory status — that is recorded as no public information rather than inferred.
Vendors
19 entries
For: Customer Support Voice Agents Internal Tools RAG & Search Autonomous Agents CUA Coding Agents
For: Sold into high-volume contracting work including EPC (engineering, procurement, construction) and supply agreements
Langfuse describes itself as an open-source agent evaluation and observability platform. The vendor reports use by 21 of the Fortune 50, more than 100,000 engineers building on Langfuse, and 10+ billion observations processed per month. No insurance, indemnity or outcome-linked settlement terms are published on the pages checked.
For: industries where a perfectly functional AI is not good enough.
Data: Trust center confirms customer data is deleted on request
For: Fortune 500s & Startups
For: testing automation
For: Teams use Arize to understand how their AI agents and applications behave, measure quality with evaluations, monitor production performance, and co…
For: the world's most regulated industries
Fiddler AI is an AI observability and guardrails vendor. Its published material describes guardrails that enforce policy inline on an agent's request and response path in under 80ms, stopping violations before data leaves the customer network rather than flagging them afterwards, and states that every enforcement decision is recorded with who triggered it and what happened. It is sold on three plans, with the Developer plan usage-based at $0.002 per trace, and is deployable as SaaS, VPC or on-premises, including AWS GovCloud. The vendor states its users include Fortune 100 organizations in regulated sectors such as financial services, healthcare, insurance and government, and lists inclusion in four Forrester, Gartner and IDC analyst reports. No compliance certificate, auditor or report date is published on the source page, and no insurance cover, customer indemnity, coverage limit, outcome-linked payment or settlement mechanism is published. All statements above are drawn from the vendor's own homepage, which is a primary source for what the vendor claims about itself.
For: Financial services/banking, healthcare & life sciences, insurance, retail, cybersecurity, biotech/pharma, government, telecom, energy & utilities
For: AI for Lawyers & Law Firms
Provides purpose-built, affirmative AI insurance for generative AI and AI agents, which the vendor states responds directly to AI failure modes. Operates as a coverholder at Lloyd's, with underwriting capacity from certain underwriters at Lloyd's; a published testimonial from a Swiss Re Senior Underwriter endorses Armilla's model-verification work. Offers an "AI Performance Warranty" that the vendor says funds a customer remedy when an insured AI system fails to perform as promised. Serves businesses deploying AI and sells through appointed, licensed surplus-lines brokers; named customers and case studies include WorkTango, SkyHive (now part of Cornerstone OnDemand), Private AI, Accend, hireEZ, and a bank/credit-union lending deployment (MKIII). Frames its coverage around "high-risk" AI system categories referenced in the Colorado AI Act and EU AI Act. Coverage availability is jurisdiction-dependent. No coverage limits are published on the page checked.
Keywords AI supplies an AI observability platform, which the page checked refers to as Respan. Per the vendor's own homepage, every LLM call, tool run, retrieval and agent turn becomes a span in a single trace, with the model, latency, cost, tokens and the exact input and output recorded on every span. The vendor states budgets and rate limits can be set per key, per customer or org-wide, with warnings as they are approached and requests blocked before spend runs over. Outputs can be scored against a bar the customer sets, on a test set and in production. ISO 27001, SOC 2, GDPR and HIPAA compliance with a BAA available are self-declared, with no report, auditor, certificate date or SOC 2 Type published on the page checked. The platform is described as sitting behind 80 trillion or more tokens, and pricing starts free. No insurance, indemnity, liability cap, settlement mechanism, outcome-based pricing or tamper-evident logging claim appears anywhere on the page checked.
Markets cyber insurance combined with security tooling. Its enterprise cyber coverage is described on its own site as backed by Allianz's A+ rated capacity. Its own regulatory status, and whether it writes AI-specific cover, are not stated on the page checked.
Credo AI markets an AI governance platform. It publishes policy packs mapped to the EU AI Act, NIST AI RMF, ISO-42001, OMB M-25, Colorado ADMT and the NAIC AI Playbook — these are frameworks the product maps customers to, not certifications Credo AI itself holds, and no certificate, auditor or report date is published on the page checked. The page describes compliance mapping with evidence recording and audit, names integrations with Snowflake, Databricks, AWS, Azure, ServiceNow, Jira, Confluence, Slack, GitHub and MLflow, and states the platform is sold into insurance, financial services, federal and healthcare. The vendor states it was named a Leader in the Forrester Wave: AI Governance Solutions, Q3 2025. No insurance, indemnity or outcome-linked payment terms are published.
For: Deploying your agents to production? We're here to help.
Integrates: Ingests metrics, logs and traces from observability tools including Grafana, Prometheus, Datadog, Dynatrace, New Relic and Splunk
Audit trail: after stopping the interaction can be saved and archived
If you are a vendor
Entries are compiled from public sources. The fastest way to correct or complete one is to publish the information — we cite what you state, and we record what you do not. A field marked no public information reflects your disclosure, not our effort.