Action class: Refund / credit
How this side effect is governed
Default policy. Escalate over threshold; allow under policy.
Required evidence. Customer-action receipt, amount, policy, evidence.
MINDBURN LABS
Inspect surfaceAI software
Govern what your AI support agent can change
Put a fail-closed boundary in front of AI support agents. Refunds, credits, and account changes get a policy verdict, an approval path, and a verifiable receipt.
The problem
The action is the risk, not the model.
A support agent that can move money or alter an account is one prompt away from an unauthorized refund. Security reviewers ask who approved it and what evidence exists.
Granting this agent access means it can issue refunds, apply credits, and change customer accounts. Each of those is a side effect someone has to answer for.
What HELM does
A verdict before the side effect runs.
HELM checks the proposed action against policy before any effect runs, then records a signed receipt. Here is one path for ai customer support agents.
Agent proposes
Agent proposes a $480 refund outside policy
HELM checks policy
Checks refund policy, amount, and customer context
Verdict
ESCALATE
Proof
Refund decision receipt + approval EvidencePack
Proof you can hand to a reviewer
Evidence that survives outside the dashboard.
Each governed action leaves a record your security reviewer can verify offline.
- Customer-action receipt
- Refund-approval EvidencePack
- Replay report
Questions
What reviewers ask first.
Can the agent act without approval?
Not for this action class. HELM checks the proposed action against policy before any side effect runs. For refund / credit the default is "Escalate over threshold; allow under policy", so anything unknown or unapproved stops by default.
What evidence do I show a security reviewer?
Every decision records customer-action receipt, amount, policy, evidence. You hand the reviewer the signed receipt and EvidencePack, which they verify offline without access to your dashboard.
Does this slow the agent down?
The check sits in the action path, not the conversation. Low-risk actions that policy already permits proceed; only consequential side effects pause for a verdict or an approver.
Keep reading
Follow the proof path.
Hand your reviewer proof, not promises.
Bring one refund / credit action to the boundary and see the verdict and the receipt.