Agentic Certification & Coverage
Available through: your broker for the insurance, and directly for certification
Who it is for
AI vendors
Companies that build and sell agentic software: support agents, finance and operations agents, coding and IT agents, sales agents.
What is insured
The vendor's own liability for the agents it supplies: a customer's loss when an agent fails, and the indemnities the vendor gives in its contracts.
What is certified
Each agent's deployed configuration, tested under adversarial pressure and scored on how it actually behaved. The certificate is the underwriting evidence.
Coverage
| Coverage | What it covers | Basis |
|---|---|---|
| Technology errors and omissions | A customer's financial loss when your agent fails to perform as represented: a wrong action, a missed task, or faulty output acted on. | Third party |
| Contractual indemnity | The indemnities you give enterprise customers for your agent's conduct, within the scope scheduled on the policy. | Third party |
| Data disclosed through output | Claims that your agent mishandled or disclosed personal or confidential data through what it said or did, with no network intrusion required. | Third party |
| Product liability | Bodily injury and property damage caused by an AI product you supplied. | Third party |
| Regulatory defense | Defense costs in a regulatory proceeding over an AI product you supplied, including provider obligations under AI-specific law. | Regulatory |
| AI exclusions removed | No claim is excluded because AI, autonomy or a learned model was involved. Cover is affirmative where technology E&O and general liability forms now carve it out. | Third party |
One defective model, release or configuration across your customer base counts as one event, with limits capped per vendor and per product. That is how a vendor is written without carrying correlated loss. Indicative cover for a line in development, not a schedule of insurance and not an offer to quote. Agreement names, triggers, sublimits and exclusions are subject to the filed wording, and the wording governs in every respect.
Certification
Agents run through a simulated operating range, with live tool calls, adversarial users and poisoned documents, and are scored across eight dimensions of agentic liability: scope violation, unauthorized action, data exfiltration, injection susceptibility, output integrity, behavioral instability, over-refusal and operational control.
A range, not a checklist
Each scenario is run repeatedly, and an agent is graded on whether it holds every time rather than on its best attempt.
Bound to the configuration
A certificate is tied to the model, tools and prompt that earned it, valid for twelve months and void on a material change. Anyone can verify it publicly.
Free to assess
Assessment is free. Accreditation, the official certificate and full report, is an annual fee, with a vendor plan for a fleet of agents.
Underwriting
Rated
On revenue from AI products, adjusted by certification grade, with accumulation capped per vendor and per release.
What we read
Read-only- Your certification profile for each agent: grades across the eight dimensions under repeated trials
What we ask for
7- 01The agents you ship and what each can do inside a customer's systems
- 02Customer count, sectors and deployments per agent
- 03Representations in your contracts and marketing
- 04Indemnity and limitation of liability terms
- 05Evaluation and red-team evidence at release
- 06Release, update and rollback process
- 07Claims and incident history
What moves price
6- 01The authority your agents hold: moving money, changing records, communicating with the public
- 02Certification grades
- 03Customers per agent and per release, which sets accumulation
- 04Contractual indemnity exposure
- 05Regulated versus unregulated sectors served
- 06Release cadence and the ability to roll back
Standards we reference
Review
No audit of your codebase. Certification is the review.