Every exception queue is a different problem wearing one label
Orders that leave the normal path all arrive in a single queue, which treats them as a single work type. They are not. An address that failed validation, a payment held for review, a line item that is out of stock, and a duplicate submission require four different resolutions and four different people.
Working out which is which means reading the order against what happened to it, and often against the customer's history. A first order failing validation and a regular customer's tenth failing it are different situations.
Time behaves differently per type as well. A stock exception can wait a day. A payment hold on a fulfilled order cannot.
The agent classifies, investigates, and proposes a resolution with its reasoning. It refunds nothing, cancels nothing, and charges nothing. Anything touching money stays with a person.
How the agent is built in Agent Studio
Exception types live in prompt skills, with each one defined in your systems, because the same status code covers several situations in most order platforms.
A second prompt skill holds your resolution policy: what may be fixed under your rules, what needs a human, and what goes to finance regardless of size.
The system prompt requires every classification to cite the order data supporting it. The agent must say when the data does not distinguish between two possibilities.
It forbids any action touching money or fulfillment. The agent proposes. A person runs the resolution, or a person triggers a flow that runs it.
A Shopify node or an HTTP Request node supplies the order, its history, and the customer's record, and the Agent Chat node reads them together.
Proposals go to a Monday or Linear queue, grouped by exception type, which turns one undifferentiated queue into four workable ones.
A Human in the Loop node sits on every resolution, and Google Sheets logs the proposal against what was actually done, so the policy improves from real outcomes.
What this agent is built from
- Prompt skills: what each exception type actually means in your systems.
- Prompt skills: your resolution policy, including what must go to finance.
- System prompt: every classification cites the order data behind it.
- System prompt: no refunds, cancellations, or charges. Money stays with the person.
- Shopify (Integration) with HTTP Request (Util): the order, its history, and the customer record.
- Monday (Integration): the queue is split by exception type. Linear and Jira fit the same slot.
- When an ActionFlow is enough: If you route exceptions by their status code, build the workflow. That is a fair first split.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.