Agents Lookup

Recipes / E-commerce ops

Shopify Order-Exception Triage Agent

An agent reviews new Shopify orders for address problems, risk flags, stock gaps, split shipments, and customer notes, then sorts them, routes them, and drafts customer messages for a person to approve.

CategoryE-commerce ops
Steps9
Human sign-offs5

The problem

Most orders ship without trouble, but a handful each day need a human decision before anyone packs them. An address with no unit number, an order Shopify flags as risky, a line item that went out of stock after checkout, an order that has to ship in two boxes from two places, or a note saying it's a gift and to leave the receipt out. In a small store these exceptions get found late, usually by whoever is packing, and the fixes happen in a rush: a quick email to the customer, a hold someone forgets to release, a cancellation nobody wrote down. This recipe adds a triage step between order placement and fulfillment. The agent checks each new order against rules you write down, gives every exception a category and a priority, sends it to the right person, and drafts the customer message. A person still makes every hold, cancellation, and outbound message decision.

Who it's for

Not a fit when

Tools

Shopify adminSource of order, customer, fulfillment, and risk data; where a person applies holds, tags, and cancellations
Zapier or MakeStarts the workflow when a new order arrives through a webhook, a new-order trigger, or a scheduled export, and passes order fields to the agent
Claude or a similar LLMClassifies each order against your exception rules, explains why, and drafts customer messages
Google SheetsFree exception queue and audit log: one row per flagged order with category, priority, status, and approver
Slack or emailAlerts the right owner when a high-priority exception lands in the queue
Gmail or your help deskWhere approved customer messages are sent from by a person, so they stay in the normal support thread
Address validation service (optional)Checks shipping addresses against postal data before the agent decides whether to flag them

Inputs you need first

Step-by-step build

  1. Write the exception rules

    Before touching any tool, list every exception type in a shared doc with a short definition and an example. A good starting list: ADDRESS (missing unit number, PO box with a carrier that can't deliver there, mismatched country and postcode), RISK (a medium or high risk indicator from Shopify or your fraud app, billing and shipping countries that don't match on a first order), STOCK (any line with available quantity at or below zero after the order), SPLIT (lines fulfilled from more than one location), and NOTE (any customer note with words like gift, urgent, change, cancel, or deliver by). Give each one a default priority from P1 to P3 and an owner.

  2. Build the exception queue sheet

    Create a Google Sheet with these columns: order_number, order_date, customer_email, exception_types, priority, agent_reason, suggested_action, draft_message, owner, status, approved_by, approved_at, resolution_note. Status values: New, In review, Approved, Rejected, Resolved. Lock the approved_by and approved_at columns so only reviewers fill them. Add a second tab called Rules that copies the definitions from step 1, so the agent and your team read the same source. A free Google account is enough to start; move to a help desk or database later if volume grows.

  3. Set up the order trigger

    In Zapier or Make, start a workflow on new paid orders. Depending on your plan and setup this could be a native new-order trigger, a Shopify webhook, or a scheduled export every 15 to 30 minutes. Pass only the fields the agent needs: order number, created time, shipping address, line items with SKU and quantity, fulfillment location, available stock per line, risk level, tags, and customer note. Leave out payment details and full card data entirely. Filter out test orders and orders already tagged as reviewed so the same order isn't processed twice.

  4. Write the classification prompt

    Paste the rules from your Rules tab into the agent instructions and ask for structured output: a list of exception types, a priority, a one-sentence reason that quotes the field that triggered it, a suggested action from an allowed list (contact customer, hold for review, substitute, ship partial, no action), and a draft message where customer contact is suggested. Tell the agent to return NONE when no rule matches, and to say UNSURE rather than guessing when a field is missing. Ask for JSON so the automation can write each field to its own column.

  5. Route and alert

    Add a step that writes every flagged order to the queue sheet and skips orders that return NONE. Then route by priority: P1 items, such as high risk or an address that cannot be shipped, post to a Slack channel or email the owner right away; P2 items collect into a digest every two hours; P3 items wait for the daily review. Apply a tag in Shopify such as exception-pending only if your team agrees, and only as a visible marker; the agent should not place fulfillment holds itself.

  6. Draft customer messages from templates

    For exceptions that need customer contact, the agent fills your approved template with order details: first name, order number, the specific question (for example, please confirm your apartment or unit number), and a reply-by time. It must not promise refunds, discounts, delivery dates, or substitutions you have not approved. Drafts go into the draft_message column, or into Gmail or your help desk as unsent drafts. A person reads, edits, and sends each one from the normal support inbox so the thread stays with the customer record.

  7. Test on past orders

    Export two to four weeks of past orders, including ones you know had problems. Run them through the workflow with alerts and drafts turned off, and compare the agent's labels with what actually happened. Track misses (real exceptions it didn't flag) and false alarms (clean orders it flagged). Tighten rules where it confuses categories, such as treating every gift note as urgent. Keep adjusting until your reviewer trusts the queue enough to work from it, then test a few live orders while still checking everything by hand.

  8. Go live with a review rhythm

    Turn on the live trigger and set clear review times, for example a check of P1 items within 30 minutes during business hours and a full queue sweep twice a day. The reviewer opens each row, checks the order in Shopify, and either approves the suggested action, changes it, or rejects it, filling approved_by and approved_at. Holds, cancellations, partial shipments, and customer messages happen only after that sign-off, done by a person in Shopify or the help desk. Fulfillment staff should not pack any order still tagged exception-pending.

  9. Review and maintain monthly

    Once a month, filter the sheet for rejected suggestions and missed exceptions found later at packing or through complaints. Update the Rules tab and the prompt together so they stay in sync. Retire rules that never fire and add new ones when your catalog, carriers, or locations change. Check that the automation still receives every field after any Shopify app or theme change, since a missing field can quietly turn into missed flags. Keep the sheet as your audit trail, and archive old rows each quarter.

Example agent instructions

Adapt this to your business. Placeholders in braces are filled by your automation tool.

You are an order-exception triage assistant for {store_name}. You review one Shopify order at a time and decide whether it needs human attention before fulfillment. You never cancel, refund, edit, hold, or fulfill orders, and you never send messages. You only classify and draft.

Exception rules:
{exception_rules}

Order data:
{order_data}

Approved message templates:
{message_templates}

Instructions:
1. Check the order against every rule. List each matching exception type: ADDRESS, RISK, STOCK, SPLIT, NOTE.
2. If none match, return {"exceptions": "NONE"}.
3. For each match, quote the exact field value that triggered it.
4. Assign priority P1, P2, or P3 using the rules. If two rules apply, use the higher priority.
5. Choose one suggested_action from: contact_customer, hold_for_review, offer_substitute, ship_partial, no_action.
6. If the action is contact_customer or offer_substitute, fill the matching template. Do not promise refunds, discounts, delivery dates, or substitutes not listed in {approved_substitutes}.
7. If a required field is missing, mark it UNSURE and explain what is missing. Do not guess.

Return JSON with: order_number, exceptions, priority, reason, suggested_action, draft_message, owner.

Human sign-off points

Guardrails

Expected time saved

Estimate: if your team checks about 150 orders a week and spends about 2 minutes on each to spot problems, plus about 10 minutes per real exception to write a customer email, a triage queue with drafted messages might save 3 to 5 hours a week, mostly from not scanning clean orders and from starting with a draft. This assumes your rules are written down and order data is complete. Your results depend on volume and data quality.

How to tell it's working

FAQ

Can the agent place holds or cancel orders automatically?

This recipe deliberately keeps those actions with a person. Holds and cancellations affect revenue, customer trust, and sometimes payment disputes, so a reviewer makes the call in Shopify after reading the agent's reason. Once your team trusts the queue, you can reduce effort by batching approvals, but a human approval step stays in place for every hold, cancellation, and outbound message.

Does this replace a fraud screening app?

No. The agent reads whatever risk indicators Shopify or your fraud app already provides and puts them in front of the right person with context. It doesn't score fraud on its own and shouldn't be treated as a fraud tool. If you sell high-value or high-risk goods, keep a dedicated fraud service and use this triage step to organize the review work.

What does it cost to run?

A low-cost setup uses a Google Sheet as the queue, a Zapier or Make plan sized to your order volume, and pay-per-use LLM calls on short order records. Free tiers can cover testing with past orders. Costs grow mostly with order count and with how often the workflow runs, so filtering out test orders and duplicates keeps usage down.

Want this set up for you?

Build it yourself with the steps above, or tell us what you need. We scope the work with you and send a written quote before anything starts.

Request a quote

Sources

Written by Agents Lookup. Official pages for the tools this recipe uses: