Errors of the decision agent.
AI Agent E&O is the line we underwrite where a software agent — a decision agent, a retrieval-augmented system, an autonomous copilot, a sales or support agent acting on behalf of the insured — takes an action or makes a recommendation that produces financial loss to a third party. Coverage attaches to the erroneous or unauthorised act of the agent itself.
This is the form a buyer reaches for when the cyber tower's professional services endorsement has the new AI carve-out. It is claims-made, written affirmatively, and indifferent to whether the agent was first- or second-party hosted, model-provider-agnostic, or built in-house.
What we mean by a decision agent.
- Customer-facing chat and support agents
- Sales agents, outbound and inbound
- Retrieval-augmented systems serving regulated answers
- Code-generation copilots integrated into production
- Underwriting, eligibility, pricing, and KYC agents
- Procurement, scheduling, and operations agents
- Compliance and legal-research agents
- Cyber-physical autonomy → Autonomous Systems
- First-party model drift on declared evals → Model Warranty
- EU AI Act & regulatory inquiry → Regulatory Defense
- Intentional disclosure of confidential information → cyber tower
- Trademark, copyright, or IP infringement → existing IP placement
- Bodily injury from a third-party action → existing GL
Where the form attaches and excludes.
| Agent context | Primary limit | Excess available | Standard retention | Telemetry tier |
|---|---|---|---|---|
| Customer-facing chat & supportExternal users, transactional or advisory. | $3M | to $15M | $25K | Tier A |
| Sales agents & outbound automationLead qualification, quote issuance, contracting. | $2M | to $10M | $25K | Tier A |
| Code-generation copilot in productionWhere output ships to customer systems. | $2M | to $10M | $50K | Tier A |
| Underwriting, pricing & eligibilityFinancial services, insurance, lending. | $2M | to $15M | $100K | Tier S |
| Clinical & regulated advisoryHealthcare triage, legal Q&A, compliance. | $2M | to $15M | $100K | Tier S |
| Internal operations agentsScheduling, procurement, IT. | $1M | to $5M | $10K | Tier B |
| Multi-agent orchestrationsTwo or more agents in a hand-off loop. | $2M | to $10M | $50K | Tier A |
| Exclusion | Form reference | Buy-back available | Notes |
|---|---|---|---|
| Known prior actsActs known to the insured before retroactive date. | § 5.1 | No | Retroactive date set at bind by negotiation. |
| Undeclared model substitutionModel swapped post-bind without notice. | § 5.2 | No | Notify within 5 business days to preserve cover. |
| Intellectual-property infringementTrademark, copyright, patent. | § 5.3 | By endorsement | Limited training-data IP sublimit available. |
| Bodily injury & property damageWhere the loss is physical, not financial. | § 5.4 | No | Cyber-physical → Autonomous Systems. |
| Privacy & data-protection penaltiesStatutory penalties where uninsurable. | § 5.5 | By referral | Defense costs of underlying still covered. |
| Intentional bias or discriminatory trainingWhere the insured directed the model. | § 5.6 | No | Inadvertent bias remains covered. |
| Failure to disclose AI useWhere mandatory disclosure was required and skipped. | § 5.7 | By endorsement | Carries through to Regulatory Defense line. |
What we ingest, and on what cadence.
This line binds against a telemetry contract that records the agent's actions and version history. Castra does not ingest prompts, user inputs, agent outputs, model weights, training data, or end-user personal information. The streams are metadata about the agent — not the content the agent handled.
Where the agent operates in a regulated regime, an evaluation set is also required. The eval set is described in detail in the Model Performance Warranty line, where it doubles as the warranty anchor.
The methodology and the three instruments are described in detail on the Underwriting page.
A sample loss, worked end to end.
The agent that mispriced the cohort.
A small-business lender deployed an eligibility-and-pricing agent that scored applicants against a model the insured had declared in the application. The agent quoted rates in real time. After a routine version bump to the underlying model, the agent's price band on one applicant cohort shifted by 220 basis points without the insured's knowledge. Affected borrowers received quotes priced as if their default probability had halved. Loans were issued at those rates.
The error surfaced four weeks later when a downstream reconciliation flagged the mispricing. Affected borrowers were notified, and a class of subrogation claims emerged for the difference between the issued rate and the rate the insured's prior agent had quoted on the same cohort the week before.
The agent did not break. The version it ran did. The form does not care about the difference.
Coverage attached under § 2.1.a for financial loss to the third-party borrowers. The exclusion at § 5.2 (undeclared model substitution) was the live argument: was the version bump declared? The change had been emailed to the broker on the day of the upgrade and recorded in the dependency-graph stream — the exclusion did not apply. The retroactive-date exclusion at § 5.1 did not apply, the act post-dated bind.
Indemnity paid: $2.1M (rate make-whole), $310K (third-party subrogation). Defense paid: $480K. Total within limit, applied retention $50K. The placement was renewed at +9.8% on the telemetry-priced base; the eval set was tightened at renewal to include version-bump regression checks as a fail-fast condition.
Note. This claim example is a composite, drawn from sample patterns. It is not based on any single insured. Amounts are illustrative.