WorkerKit

AI workers for E-commerce apps

Give an AI worker scoped access to the store — orders, inventory, customers and reports — and daily operations stop waiting on whoever has the admin open. Every permission is granted per worker, money moves only under hard caps, and every call is logged.

What is an AI worker for E-commerce?

An AI worker for an online store is a scoped connection to the store's admin: orders with their payment, fulfillment and return status, stock levels per location, customer profiles and spend history, and sales reports computed over complete data. A firewall between the worker and the store decides per worker what is readable, what is writable, and how much money any single operation may move — refunds and store credit carry per-operation caps, customer identities can be masked while order flow stays readable, and customer-facing emails are suppressed unless deliberately allowed.

What these workers do

The morning store brief

Yesterday's sales versus the week before, orders needing attention, refunds and returns, low stock on best sellers — one scheduled read-only run, delivered before the day starts.

Stuck-order triage

Unfulfilled orders older than a threshold, payment issues, high-risk flags and return requests get swept on a schedule and reported with the order numbers, so nothing sits unnoticed.

Low-stock sentinel

Stock is checked per location against thresholds, with days-of-cover computed from real sales velocity, and only new crossings get raised — no re-alerting every run.

Best sellers versus inventory

The top sellers ranked over a window with current stock joined in, flagging any winner whose cover has fallen below the restock lead time — the most expensive stockout caught early.

Customer intelligence

Top customers by spend, repeat rate, and previously-loyal buyers going quiet, computed from order history with identity masking and contact rules applied throughout.

How an AI worker runs in your E-commerce apps

The store admin is where a business actually lives, which is why handing an AI the whole login was never the answer. The jobs are real — the morning briefing, the stuck-order sweep, the low-stock alarm, the weekly numbers — and every one of them is cadence work that loses to whatever is urgent. A worker with exactly the permissions its job needs does that upkeep on schedule, and nothing else.

Start with reads and reports. Sales summaries, order triage lists, inventory health and customer insights are computed server-side over the complete window — never sampled — and a worker doing them holds no write at all. The reports carry comparisons built in, so "revenue versus last week" or "refund rate doubled" is a number the worker reads, not arithmetic it might get wrong.

Money is where this category differs from every other. Refunds, cancellations that refund, return processing and store credit are each a separate permission, and each is bounded by a per-operation cap you set in the store's own currency. A worker granted refunds up to a limit can settle small cases on its own and physically cannot exceed the cap; over-refunding beyond what was paid is refused outright, whatever the instruction says.

Customers notice sloppy automation, so shopper-facing noise is off by default. Every write that could email a customer — a cancellation, a refund, a fulfillment update — has its notification suppressed unless the account explicitly allows worker-triggered customer emails. The worker acts; the shopper hears about it when a person, or a deliberate setting, decides they should.

Know the platform's edges before promising work. Order history may be limited to a recent window until the connected app is granted full-history access, and the reports say so rather than letting a short window read as slow sales. Incoming stock reflects transfers between locations, not supplier purchase orders, which the platform does not expose. Alerts run on the worker's schedule — checked every few minutes or hours — not as real-time push.

E-commerce questions, answered.

Can a worker issue refunds?

Only if that exact permission is granted, and then only under a cap. Refunds are their own grant — separate from reading orders, separate even from editing them — and the account sets a per-operation ceiling in the store's currency that every refund path honors: direct refunds, the refund leg of a cancellation, and return processing alike. Refunding more than the customer paid is refused outright, whatever the worker was told. A worker without the grant can draft the recommendation and touch nothing.

Will customers get emails when a worker acts?

Not unless you deliberately allow it. The store platform likes to notify shoppers on cancellations, refunds and fulfillment updates, with defaults that vary call by call — so the firewall suppresses every one of those notifications on every worker write unless the account has explicitly enabled worker-triggered customer emails. Actions happen quietly; telling the shopper stays a human decision, or a setting you chose on purpose.

Which customer data can a worker see?

What its field grants allow, item by item. Customer names, emails, phone numbers and addresses are each maskable while order flow, totals and statuses stay readable — so a triage worker can work a queue without ever seeing who the buyer is. Account-wide contact rules go further: a blocked customer simply vanishes from lists and lookups, and the response says items were withheld rather than pretending they never existed.

Why can't the worker see older orders?

The store platform limits new app credentials to a recent window of orders — roughly the last two months — until the app is granted full-history access in its settings. Until then, reports and searches that reach further back say plainly that history was limited by the connection's scopes, so a short window is never mistaken for a quiet store. Granting the full-history scope on the connected app removes the limit for everything, including lifetime-value analysis.

Are the alerts real-time?

They are scheduled checks, honestly labeled. A worker watching stock or stuck orders runs on the cadence you set — every few minutes or hours — reads complete data, compares against its remembered baseline so it only raises new crossings, and stays silent otherwise. That suits operational alerting well; if a check ever needs to be faster, tighten the schedule. What the platform does not provide here is push notification, and the kit descriptions say so rather than implying it.

All apps