Solutions · AI Agent Insurance

Claim-grade evidence for AI agent insurance.

GeoClear helps insurers, MGAs, brokers, and enterprises verify what an autonomous agent did, what it was allowed to do, what evidence it used, and when. Denial records, reason codes, outcome records, and escalation outcomes can support future model evaluation, governance, and policy tuning. Verified action outcomes can support future risk and underwriting evaluation.

GeoClear does not verify the AI was right. It provides verifiable evidence that the action followed the approved evidence path before the customer system accepted it.

GeoClear preserves evidence. The insurer determines coverage.

GeoClear does not determine coverage. We preserve the evidence insurers need to review what happened.

▶ Run the interactive demo (~7 min) → Or see the example claim ↓ How the verification layer works →

The problem

Autonomous agents are creating a new liability problem. They can approve refunds, send messages, call tools, move money, access data, and make commitments. When something goes wrong, screenshots, vendor logs, and depositions are not enough. Three things insurers cannot get from those artifacts:

Tamper-evidence. A log can be edited. An operational evidence record cannot, without breaking verification.
Policy-period semantics. A screenshot does not know which version of the policy was active at the moment the action ran.
Independent verification. The insurer's claims team should be able to verify the artifact without depending on the vendor that produced it.

Insurers writing AI-agent liability coverage need evidence that survives disputes, audits, vendor changes, and time.

How GeoClear helps

The same substrate works at three moments in the insurance workflow.

1 · Before coverage

Evidence for underwriting and authorization scope.

A verified Agent Scope Evidence captures what the agent is authorized to do at policy bind: tools, action types, spending limits, data access, human-approval thresholds, agent operating policy version. The insurer underwrites against a machine-checkable scope, not a free-text questionnaire.

2 · During coverage

Operational evidence records for high-risk agent decisions.

When the agent issues a refund, sends a contractual response, calls a financial API, or acts without required human approval, a Action Evidence Record records what it did, under what context, against what agent operating policy, and when. Operational evidence records are the equivalent of telematics for agents.

3 · After an incident

Customer-Held Evidence Bundle for review.

When something goes wrong, the question is whether the agent acted inside or outside scope. The Customer-Held Evidence Bundle gathers the scope certificate, the action evidence records, the policy snapshot, the tool-call record, and the verification report into one portable, customer-held ZIP. Insurer, insured, and broker can each verify it independently.

Example claim

Want to walk through this with real cryptography in your browser? The same scenario is wired into a 5-step interactive demo: ▶ run the AI Agent Claim Review demo → (about 7 minutes; nothing leaves your browser; download a real Evidence Bundle ZIP at the end).

An enterprise deploys a customer-support agent. The agent is authorized to answer billing questions and create support tickets. It is not authorized to issue refunds above $500 or make contractual commitments. A customer asks for a refund and threatens legal action.

The agent does three things: promises a $7,500 refund, tells the customer the company accepts liability, and calls the refund API without the required human approval. Now the insurer has to answer: was this covered behavior or outside authorized scope? Three review steps, one independently verifiable bundle.

Step 1 · at policy bind

Agent Scope Evidence

agentAcme Support Agent

authorizedrefund ≤ $500

restrictedlegal admissions

approval req.refund > $500

agent op policyv1.7

🛡 verified at bind

Step 2 · at action time

Action Evidence Record

customer askrefund + legal threat

agent actionrefund $7,500 + liability admission

tool callrefund_api(7500)

human approvalmissing

timestamp14:32:18 UTC

⚠ outside scope

Step 3 · insurer view

Claim Review

questioninside or outside scope?

findingoutside approved scope

evidencelimit exceeded · approval missing · restricted admission

verificationsignature valid · drift = 0

bundleavailable

🛡 evidence preserved

Tamper test, what a screenshot can't survive

The insurer's reviewer changes the refund amount in the bundle from $7,500 to $500 to test the verification. The operational evidence record no longer matches the recorded payload.

Verification, before edit

payload.refund_amount = 7500 verify(record, retained verification material) → VALID

After reviewer edit

payload.refund_amount = 500 ← edited verify(record, retained verification material) → VERIFICATION FAILED The bundle no longer matches what the agent's runtime actually recorded. A screenshot does not have this property.

🛡 customer-held bundle independent verification policy-period bound tamper-evident

GeoClear does not determine coverage. We preserve the evidence insurers need to review what happened.

The five evidence artifacts

Each artifact is a domain-specific Evidence Bundle or operational evidence record produced on the same platform. The substrate is unchanged; the artifact name maps onto the moment in the insurance workflow.

Artifact	Insurance moment	What it captures
Agent Scope Evidence	Pre-bind	Agent identity · authorized + restricted actions · spending limits · data access · approval thresholds · policy version
Underwriting Evidence Bundle	Pre-pricing	Governance assessment · adversarial probe summary · security posture · tool access map · risk controls · escalation rules
Action Evidence Record	Per high-risk action	Agent identity · tool invoked · arguments · result · policy at action time · approval status · timestamp
Customer-Held Evidence Bundle	Post-incident	Scope certificate · action evidence records · policy snapshot · tool calls · approvals (or absence) · verification report
Renewal Evidence Summary	Pre-renewal	High-risk action count · escalations · exceptions · drift indicators · control improvements · evidence coverage

These are packagings of the existing GeoClear substrate, not new primitives. See the developer glossary →

Buyer value

Insurers · MGAs

Faster claim review
Better underwriting evidence
Less ambiguity on scope
Stronger renewal data
Stronger enterprise trust

AI startups

Easier procurement
Better risk posture
Evidence for enterprise buyers
Stronger insurance application
Faster incident review

Enterprise risk teams

Clear accountability
Reviewable agent behavior
Evidence for governance
Better vendor risk management
More confidence deploying agents

One platform, multiple operational evidence record profiles

GeoClear is real-time operational evidence record infrastructure for machine-driven systems. AI Agent Insurance is the reference implementation of the platform applied to AI-agent liability evidence. The same evidence-record substrate powers the other record profiles:

Location Record Profile, physical-presence claims (drone, logistics, IoT).
Flood & Census Record Profile, mortgage / insurance / secondary-market review.
Decision Record Profile, AI and rules-based decisions (credit, underwriting, agent actions).
AI Agent Insurance, claim-grade evidence for insured autonomous agents (this page).

What GeoClear is not

To make the boundary unambiguous: GeoClear is not an insurer, not a broker, not a claims adjuster, not a coverage-determination service. We do not price risk. We do not write policies. We do not approve or deny claims. We do not make legal conclusions about an agent's actions.

What we do: produce customer-held Evidence Bundles that an insurer, insured, broker, regulator, or auditor can independently verify with retained verification material. The insurer determines coverage. We preserve the evidence.

▶ Run the interactive demo (~7 min) → 📄 2-page buyer brief (PDF) How the verification layer works → Trust posture →

GeoClear preserves evidence. The insurer determines coverage.

Last updated: 2026-05-06