The situation
A customer says they were billed after canceling, cannot log in, and already contacted support. An operator needs account identity, subscription events, invoice status, product access, and the earlier ticket before deciding what to do next.
Example request
“Investigate Acme’s cancellation complaint. Build a timeline, tell me what probably happened, and prepare—but do not issue—any refund.”
Separate investigation from action
Read phase
Build the account timeline
Resolve the account, retrieve plan and cancellation events, list invoices, search tickets, and inspect login events. No state changes.
Write phase
Prepare and approve the remedy
Draft a refund proposal and customer reply. Show amount, reason, recipient, and effect before a person approves separate execution tools.
A safe tool surface
| Tool | Role | Control |
|---|---|---|
find_customer | Resolve a name, email, or company to a stable customer ID. | Return disambiguation choices; never guess silently. |
get_account_timeline | Combine selected read events into one ordered view. | Return sources and visibility labels for each event. |
prepare_refund | Validate eligibility and calculate a proposed amount. | No money moves; return a proposal ID and warnings. |
issue_refund | Execute one previously prepared refund. | Require approval, proposal ID, idempotency key, current eligibility recheck. |
send_customer_reply | Send an approved response through the support system. | Show recipient and final text; keep separate from refund execution. |
Why one “resolve complaint” tool is not enough
A single tool could be convenient, but it would blur investigation, judgment, money movement, and external communication. Separating those risk boundaries lets the host show meaningful approvals, the server apply different permissions, and operators audit exactly what happened.
What the operator should receive
- A chronological account timeline with a source system for every event.
- A clear distinction between observed facts and the assistant’s likely explanation.
- Missing or inconsistent data that prevents a confident recommendation.
- A proposed remedy with policy basis, amount, and side effects.
- Separate controls to approve or reject the financial and communication actions.
Business outcome
A faster investigation without collapsing authority.
The AI can assemble evidence and prepare the next move. The organization keeps permissions, policy, confirmation, and final accountability in the systems that own them.