Use case

Approve refunds and charges an AI support agent wants to issue

An AI support agent that can look up an order, understand the complaint, and decide a refund is warranted is genuinely useful — most refund requests are routine and the reasoning is easy to check. The part worth gating isn't the judgment call, it's the money actually moving.

Try it free Self-host instead
The problem

Problem

A support agent that's right 98% of the time about refund amounts is still wrong roughly once in fifty. At volume, "wrong" here means real charge-backs, real disputes, real reconciliation work, and a customer who got refunded twice or not at all. Refunds also invite a specific failure mode agents are bad at resisting: a customer arguing forcefully for a bigger refund than policy allows, phrased as if it were already agreed.

How it runs

Workflow

  1. The agent decides a refund is warranted, computes the amount, and calls impri_push_action with the order ID, amount, and reason as the preview — not the refund API call itself.
  2. Impri can route by amount: small refunds under a threshold auto-approve via a rule, while anything above it — or anything the agent flags as a policy exception — reaches a human.
  3. A human reviewer sees the order, the customer's message, and the proposed amount side by side, and approves, reduces the amount by editing the field, or rejects with a note.
  4. impri_await_decision returns the (possibly edited) amount; the agent calls the payment processor's actual refund endpoint only with that approved figure.
  5. impri_report_result closes the loop with the processor's transaction ID, so the audit log ties every refund back to the decision that authorized it.
What it looks like
The Impri inbox showing a pending refund action alongside other queued approvals
The Impri inbox showing a pending refund action alongside other queued approvals
For the agent

MCP example

impri_push_action({
  kind: "payment.refund",
  title: "Refund $84.00 to order #10432 — damaged item, customer provided photos",
  preview: {
    format: "plain",
    body: "Order #10432, $84.00 full refund. Customer: item arrived cracked, photo attached to ticket #6021."
  },
  payload: { order_id: "10432", amount_cents: 8400, currency: "USD" },
  target_url: "https://support.acme.com/tickets/6021",
  editable: ["payload.amount_cents"]
})
// -> { action_id: "act_5jf8...", status: "pending", inbox_url: "https://app.impri.dev/inbox/act_5jf8..." }

impri_await_decision({ action_id: "act_5jf8...", timeout_s: 1800 })
// -> { action_id: "act_5jf8...", status: "approved", decision_at: 1757830200, preview: { format: "diff", body: "..." }, edited_by_human: false }

impri_report_result({
  action_id: "act_5jf8...",
  status: "executed",
  detail: "Refunded via Stripe, txn re_1P..."
})

Read the MCP docs

What changes

Result

Routine, in-policy refunds still flow at the speed a support agent promises. Every refund that's large, unusual, or contested gets a human looking at the actual amount and reason before money moves — and a paper trail that shows exactly who authorized it.

Related: Approve refunds issued by an AI support agent and Human approval for fintech AI agents. Try Impri free.

Try it free Self-host instead