WISMO Automation with AI Agents: Order Lookup, Tracking, and Escalation (2026)

“Where is my order?” is not a creative question. It is a state lookup dressed as conversation. Stores still burn associate time because the answer lives in two systems — OMS and carrier — and the customer typed a name, a phone, or a screenshot instead of ORD-1001.

Gorgias has put WISMO at about 18% of incoming requests and called it the most common ticket type, as reported by Redo. Pull your own tag mix. If it is not in your top three, start somewhere else on the 15-automations map.

On 17 June 2026, AgentCore Harness reached GA (What’s New). Agents Classic is in maintenance for new customers after 30 July 2026. A WISMO agent is a single-domain, read-heavy loop: Harness is the default. Do not start on Classic.

Baymard 70.22% cart abandonment (50 studies, updated 22 Sep 2025) is a different leak. Those shoppers never paid. WISMO is people who did. Do not pitch a WISMO bot as a conversion program.

This is post 3 in the series. The control plane (tools, HITL, Policy) is in the support-agent post. Warehouse holds and payment exceptions are order-exception management — keep those tools off this harness.

Outcomes here are directional: less associate copy-paste, faster truthful status, fewer “you said Tuesday” arguments. We are not publishing a ticket-cut or repeat-purchase number.

First-party signals we reuse (not eCommerce client outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Model WISMO volume on the AgentCore pricing calculator.

Reproduce this — States, notify rules, and ten golden evals: wismo-state-machine.md. Series folder: ecommerce-ai-agents-series/. Week plan: monday-checklist.md.

Opinionated take: quote carrier tool fields only. Never interpolate a clock time from “out for delivery.” Trade-off: some chats stay unsatisfying (“we don’t have a window”) instead of confidently wrong.

What WISMO automation actually is

Three jobs:

1. Identify the order (id, email + last-4, guest token).
2. Read fulfillment + carrier state without opening five tabs.
3. Notify or escalate — including before the ticket, when an exception lands.

That is it. A model that “sounds empathetic” while inventing a ship date is a CX defect.

There is no native Shopify AgentCore connector. getOrder and getShipment are your OMS and carrier (or aggregator) APIs behind Gateway. Shopify, custom OMS, Magento — same contract.

flowchart TD
  WISMOTicket --> OrderLookup
  OrderLookup --> FulfillmentState
  FulfillmentState --> CarrierTracking
  CarrierTracking --> DelayOrETA
  DelayOrETA --> NotifyOrEscalate

If OrderLookup returns two open orders, stop and ask which. If it returns none, request the order number. Do not search by first name across the whole tenant.

Order lookup

Allowed identifiers: order id, email plus last-4 of phone or postcode if that is how you already authenticate in the helpdesk, plus a signed guest link.

Not allowed: “I’m Jane in Austin, it was a blue sweater.” That is a fishing prompt. Escalate to a human who can run your existing PII-safe search.

Guest checkout with no email match → request order number. Preorder / future ship → policy snippet, not a carrier ETA. Cancelled order → do not quote tracking from a stale shipment object.

Split shipments: return both tracking ids. The most common “your bot is wrong” screenshot is one of two boxes.

Fulfillment vs carrier (do not mash the states)

The state machine keeps these separate on purpose.

fulfilled_no_scan is where stores lie with “it’s on the way.” It is not on the way until a scan exists. A wait window, then an ops ticket, is the honest path. That ops ticket is adjacent to order exceptions — WISMO agent opens it; it does not cancel or split.

Set the wait window per carrier, not as a store-wide “48 hours.” Some labels scan at pickup; some scan at the first hub the next morning. Encode the window in config, not in the system prompt. When the window expires, the agent files ops with order id, label time, and last (empty) scan — it does not tell the customer the warehouse “probably forgot.”

Delays, ETA, notifications

ETA. One field, from getShipment. Missing field → say missing. Do not average historical transit days in the prompt.

Delays. When the carrier exception lands, send a proactive delay notice if you can. That is the workload win: the 18%-shaped pile never becomes a ticket. We are not claiming your ticket rate will fall 18%. Gorgias’s mix is a prioritization hint.

Notifications. Never send a second “your order shipped” if getShipment already triggered that email. Duplicate shipped mail is how customers decide the brand is noisy and then still open a WISMO chat.

International customs hold: last event only. Do not quote a clearance date the tool does not have.

Platform latency: Gateway ~180 → ~95 ms on the CRM canary is not your carrier API. Expect the tracking provider to dominate. Cache last-scan within a small TTL if the vendor rate-limits; do not cache a fake ETA.

Escalation (especially delivered-not-received)

Human path, same queue as the support agent:

• Delivered + “I don’t have it” — no refund tool on this harness
• Chargeback language — stop tools, escalate
• Stalled scan past your SLA — exception copy and ops ticket
• RTS — HITL for reship vs refund
• Address change after label — usually too late; escalate, do not call updateAddress from WISMO

Session id + tool trace on every handoff. Associates re-doing carrier lookup from a pasted chat is how you fail the speed thesis.

Cost: if WISMO is most of chat volume, the published 50K sessions/mo ~$791/mo support-style silhouette is the right shape to model — platform + tokens, not a store invoice. Browser off. You do not need a headless carrier site if you have a tracking API.

What broke — An early WISMO harness answered “out for delivery” with “by 2pm” because the prompt said “be specific” and the model interpolated a local courier pattern. The carrier tool had no appointment window. Detection: golden eval #4-equivalent failed in staging once we added an assertion: reply must not contain a clock time unless getShipment.etaClock is set. Production would have shown it as angry “you promised 2pm” tickets. Recovery: strip clock-time from the instruction; add the eval; forbid free-text dates unless copied from the tool JSON. Lesson: specificity without a field is fabrication.

What to Do This Week

1. Tag last month’s helpdesk for WISMO vs everything else. If it is not a top slice, stop.
2. Implement getOrder + getShipment only. No refund, cancel, or address write.
3. Encode the seven states from wismo-state-machine.md.
4. Run the ten goldens in that file (two orders, guest, preorder, split, customs, cancelled, chargeback, delivered-not-received).
5. Wire proactive delay notify on carrier exception — dedupe against existing shipped emails.
6. HITL for delivered-not-received and RTS. Attach traces.
7. Price expected session volume on the AgentCore pricing calculator. Follow monday-checklist.md.

If the blocker is Gateway Policy on the adjacent refund path, contact us. For retail-shaped AWS work see AWS for retail / eCommerce and Generative AI on AWS. FactualMinds keeps the state machine in git so the argument is “which state is this order in,” not “what should the bot make up.”

What This Post Doesn’t Cover

• General support policy / product Q&A — AI customer support agent.
• Holds, splits, payment capture failures — order exceptions.
• Returns eligibility and RMA writes.
• Multi-carrier rate shopping or label purchase.
• Native marketplace (Amazon, eBay) A-to-Z claim flows.
• Percent reductions in WISMO ticket count or repeat-purchase rate presented as a measured client outcome.
• A Shopify-native AgentCore app (there isn’t one).

FAQ

When should we NOT automate WISMO with an AI agent?

Skip the agent when OMS and carrier webhooks already send accurate status and customers can self-serve tracking without a ticket. Also skip if you cannot call getOrder and getShipment as authenticated tools — a model guessing from a forwarded email is worse than the queue. Do not start on Agents Classic after 30 July 2026.

What could go wrong if the agent invents an ETA?

Customers plan around a clock time the carrier never published. You then eat reships, refunds, and trust. Quote last scan and carrier ETA fields only. Out for delivery is not 2pm unless the tool said 2pm.

What could go wrong if delivered-not-received auto-refunds?

You train porch-theft and friendly fraud, and you skip the investigation the warehouse and carrier need. Escalate. Human approval for reship or refund after delivered. The support-agent tool policy forbids createRefund in week one for this reason.

Is WISMO really about 18% of tickets?

Gorgias has described WISMO as about 18% of incoming requests and the most common ticket type, as reported by Redo. That is not our measurement and not your store. Pull your own helpdesk tag mix before you staff the project.

Should we wait for the customer to ticket a delay?

No, if you have a carrier exception event. Proactive delay notice is the workload win: the ticket never opens. Do not send a second shipped email if getShipment already triggered that notification.

How does WISMO fit the rest of the 15 automations?

WISMO is the highest-volume support slice. It is not order-exception management (holds, splits, payment) and not returns eligibility. Keep those tools off the WISMO harness. See the pillar map and the support-agent post for the shared read tools and HITL queue.