Market briefing
Market briefing · September 2026
AI risk & insurance
Prepare for a conversation about a real workflow, a plausible loss, and evidence that helps someone decide.
A useful opening
“Which AI loss scenario is hardest to assess today, and what evidence would change your decision?”
Glacis is participating in Lloyd’s Lab Cohort 17. The market test starts with a defined exposure and a decision owner. The insurer retains underwriting and claims judgment.
Official cohort announcement ↗Frame the exposure
One workflow. One loss scenario.
Identify the action, the decision owner, the plausible harm, and the policy question.
Interrogate the evidence
Separate a record from a conclusion.
Ask what was observed, which boundary was tested, and which gaps remain.
Make the next step concrete
Design a bounded market test.
Agree the evidence package, recipient, decision, and stopping conditions.
The workflow review
From risk to evidence
Choose a scenario to see the control, the record, and the limits.
Agentic operations
An agent reads a supplier email and prepares a payment instruction.
- Failure mode
- Untrusted content changes the destination, or the agent uses authority beyond the intended task.
- Control to test
- Limit tool permissions and credentials; require approval before a consequential action; test the bypass routes.
- Evidence to request
- The task and policy versions, approval identity, attempted tool call, execution result, and record from the receiving system.
- What remains unknown
- A signed request alone does not establish that the payment executed, that the approver understood it, or that every route was captured.
First underwriting question
Which actions can create a loss, who authorizes them, and where can the agent bypass the control?
Illustrative learning scenarios. Controls and evidence must be evaluated in the actual deployment; this page makes no claim about a customer’s system or policy response.
Keep the distinctions clear
Four boundaries worth remembering
- Access ≠ reliance
- Read-only access can still produce an answer someone acts on.
- Request ≠ effect
- An authorized tool call needs a result and, where possible, evidence from the receiving system.
- Signature ≠ truth
- Integrity and linkage claims need separate support for correctness, completeness, and attribution.
- Evidence ≠ coverage
- Policy wording and the authorized insurer determine underwriting and claims decisions.
Leave with a testable next step
Capture these five things
Use synthetic or authorized sample data for the first review.
- A named workflow and loss scenario.
- The control boundary and known bypass routes.
- The evidence package and its capture gaps.
- The person who will use it and the decision it could inform.
- A pilot scope, success criterion, and stopping condition.