Ecommerce AI agents split into four jobs: catalog discovery, shopping assistance, order support, and back-office work. Map each agent's authority, cart and checkout boundaries, identity requirements, and write permissions.
An ecommerce AI agent is not one product. The label covers four different jobs: helping shoppers discover products, assisting with cart actions, resolving order issues, and doing back-office work. Each job has its own data source, permission level, and failure mode. Treating all four as a single project is the most common reason agent deployments stall.
This guide maps the four jobs, shows where each agent's authority begins and ends, and gives you concrete tests to run before you widen any agent's powers.
TL;DR. Ecommerce AI agents fall into four distinct jobs: catalog discovery (read-only product answers), shopping assistance (product selection plus cart actions up to but not including payment), order support (customer-specific lookups, then carefully scoped writes like returns and refunds), and back-office operations (inventory, fulfillment data, and reporting). Buy or build agents job by job, test each authority boundary, and keep order writes capped and logged. Start with read-only work before granting change powers.
A storefront chatbot can answer product questions, look up an order, process a return, and flag low inventory. Those all sound similar in a demo, but they are different engineering problems. Write down which of these four jobs you want before comparing platforms.
Catalog discovery answers “Do you have this? In this size? What does it cost?” It reads product data only. Shopping assistance goes one step further: it can place items in a shopper's real cart. Order support works on a customer's records, meaning order status, tracking, returns, and refunds. Back-office work touches inventory, price lists, fulfillment data, and reporting, often without a customer in the conversation at all.
Authority escalates from read-only catalog access to money-moving writes. That escalation is where you will find most of the risk and most of the value.
Catalog agents are trained on product data: titles, descriptions, variants, prices, availability. Their job is to answer factual questions accurately and steer shoppers toward what they need. Because this agent never touches customer records, its permission surface is small.
The main quality problem is freshness. If the agent's knowledge base reflects last quarter's catalog, it will confidently recommend a discontinued colorway or quote a wrong price. Sync matters more here than clever prompting. Some platforms index your website and documents; others connect directly to store product APIs. A store-connected sync narrows the gap between what the agent knows and what is actually for sale. Test the agent against known changes: retire a product, update a price, then ask about those exact items.
Watch for the inventory hallucination failure: the agent says something is in stock when it is not. In a catalog agent that is a knowledge error, not a permissions error. You fix it with fresher data or by marking out-of-stock states clearly in the source.
Shopping assistance builds on the catalog and adds the cart. Its defining feature is native store actions. With YourGPT's Shopify AI Copilot, connecting a Shopify store enables actions that search the catalog, render product cards, show details, add items to the real cart, and display the cart with a checkout button. The shopper's cart is the same one they would see at checkout, not a separate chat-only cart. These actions activate only when the widget runs on that store and when Agent Mode is enabled.
A hypothetical example makes the boundary clear. Assume a shopper types: “Do you have the merino crewneck in forest green, medium?”
This scenario demonstrates documented platform actions, not a tested result. Notice where the agent stops: it moves the shopper to checkout but does not process payment. If an item is out of stock, the card marks it clearly and blocks the add. That is the cart/purchase boundary. The agent can assemble a cart; completing a purchase remains the shopper's action on the store's checkout.
One practical caution: YourGPT toggles all five store actions on together rather than letting you disable individual ones. Verify whether a platform lets you turn off a specific action, like adding to cart, if you only want discovery for now. Native store actions cover standard retail flows. When you need behavior beyond those, such as loyalty-point redemption or custom shipping rules, you need custom workflows via an API, webhooks, or a workflow builder rather than a canned integration. If that custom work calls Shopify's Admin API, build and install the app through Shopify's Dev Dashboard or CLI: since January 1, 2026, new custom apps cannot be created directly in Shopify Admin, and any admin-generated token instructions you find online are legacy-only. Follow Shopify's current authentication docs and scope tokens to exactly the resources the workflow needs.
Order support is a different authority class because it involves one customer and their money. An agent that looks up order status is reading a record. An agent that processes a return or issues a refund is changing it. Read and write powers should be designed separately.
Start with reads: order status, carrier tracking, return-window eligibility, and policy explanations. These reduce your highest-volume tickets quickly. Give the agent order lookup with strict ownership checks, meaning the order must belong to the authenticated customer. Only after reads run cleanly for weeks should you consider writes: return labels, replacements under a threshold you set, then refunds and cancellations with a per-order cap, a log humans can open, and a same-day reversal path.
A typed email, order number, or account ID is a lookup key, not proof that the caller owns that record. Customer-specific actions require an authenticated customer session or a server-validated identity flow, then object-level authorization on every request. Do not collect passwords, admin tokens, or one-time codes in chat; a secure login or verification screen should handle those. For guest lookups, send a secure link to the stored contact address rather than exposing the record in the conversation.
The three-order test remains essential. Run a real order, an order ID that does not exist, and an attempt to view an order from a different customer's session or with a mismatched contact address. The third is the tell: if the agent reveals another customer's data, it must fail closed. Do not widen the agent from read to change until that test passes consistently.
When the agent hands off to a human, billing models matter. Under Gorgias's current terms, a helpdesk ticket fee applies to any ticket replied to from the helpdesk, and an additional automation fee applies only when AI fully resolves the conversation without handover. Partially automated conversations therefore cost a ticket fee; fully automated ones cost both. Legacy accounts created before 28 May 2025 may be billed under different terms, so confirm which model applies to your account.
This walkthrough of Gorgias' AI Agent on Shopify tickets shows how an order-support agent moves from a customer question to a resolved ticket:
Watch the video above to see the agent retrieve an order, verify the customer, and close a shipping inquiry without handing off to a human. If the video does not load, review Gorgias's AI Agent documentation for the same workflow.
Internal agents run on a different clock. They monitor inventory levels, flag price discrepancies, reconcile fulfillment data, and prepare reports. No customer sees them, but their errors ripple outward: a stale price feed becomes a cart mistake; a missed low-stock alert becomes an oversell. Back-office agents usually need deeper API access to your store's admin endpoints, so scope their permissions tightly. A reporting agent does not need to edit products; an inventory-sync agent might.
A concrete, hypothetical example: an inventory-sync agent polls the store's product API hourly, writes stock levels to a warehouse system, and alerts when a SKU crosses a reorder threshold. Its failure mode is silent corruption, meaning a successful-but-wrong write that no one notices until an oversell happens. Test it by seeding a known stock level, forcing a sync, and checking both systems match. Give it read access to products, write access only to the inventory field, and an audit log of every change. Remember that an API 200/202 response or a timeout is not proof the write completed: check the operation state or the authoritative record, and use idempotency plus duplicate prevention before retrying after a timeout. A price-discrepancy agent should compare list prices against the catalog and flag differences for a human, not edit prices itself. The rule for back-office is the same as for order support: the larger the write surface, the stricter the approval loop.
Here is the decision rule that recurs across all four jobs: if an agent can change a record, it needs a reason, a cap, a log, and a reversal path. Read-only work is cheap to delegate; writes are where you earn or lose your team's trust.
| Job | Data source | Typical authority | Failure mode |
|---|---|---|---|
Catalog discovery | Product catalog | Read-only | Stale or missing product info |
Shopping assistance | Catalog + cart API | Add/view cart, not payment | Wrong variant, out-of-stock add |
Order support | Orders, returns, payments | Read first, then writes | Leaked order data, unauthorized refund |
Back-office | Inventory, prices, channels | Depends on workflow | Silent data corruption |
Start with the job that has the clearest ROI and the smallest permission risk. For most stores, that is catalog discovery plus order-status reads. Both are read-heavy, high-volume, and easy to test. Add shopping assistance once your catalog sync is fresh. Reserve writes, meaning returns, refunds, and cancellations, for after you have watched read behavior for a few weeks and built the logging and approval loop.
When comparing platforms, ask how each handles the specific job you are starting with. Chatbase's pricing includes a Shopify integration even on its entry-level Hobby plan, with message-credit pools rather than per-seat pricing. YourGPT advertises built-in Shopify actions and order and return handling in its public materials; verify which actions are enabled on your plan and whether its Studio builder is needed for your workflows. Gorgias ties AI pricing to fully resolved interactions, which aligns cost with automation rate but means you pay a helpdesk ticket fee even when AI hands over. Zendesk prices per seat with AI usage billed per automated resolution; budget for both components. These are different payment shapes, not equivalent plans. Compare them on the job you are actually automating, not on monthly sticker price alone.
For implementation steps and platform setup, see our ecommerce AI agents implementation guide. For choosing support software specifically, start with our ecommerce AI support comparison, and cross-reference the broader list of AI customer service software for ecommerce.
Prices and capabilities cited here come from vendor pricing pages and public documentation reviewed on 26 October 2026. Vendor-documented claims are not independent benchmarks. Confirm current rates and allowances with a quote before committing.
Can one AI agent handle product questions and order support at the same time? Yes, but give it two distinct tool sets. Catalog questions should never need customer data, and order lookups require an authenticated session tied to the order owner. If both live in one agent, keep the permission boundaries separate and test the ownership check by attempting to view an order from a different customer's session.
Should my agent be able to add items to a shopper's cart? Only if the store integration exposes a real cart and you are comfortable with the action. Native store actions typically add to the actual cart, not a demo one, and out-of-stock items are blocked. The checkout step stays with the shopper; the agent should not process payment.
How do I prevent an agent from issuing refunds it should not? Scope refunds as a distinct capability with a per-order cap, an approval rule for larger amounts, and a log your team reviews. Delay enabling writes until order-status reads have run cleanly for weeks, and keep a same-day reversal path. Make sure the refund action is authorized against the authenticated customer's session on each request, not just a typed email or order number.
What is the difference between message credits, seats, and per-resolution pricing? Platforms bill differently. Chatbase uses monthly message-credit pools with auto-recharge at $40 per 1,000 credits. Gorgias charges per helpdesk ticket plus an automation fee only for fully AI-resolved conversations. Zendesk and Freshdesk price per seat with separate AI usage units. Compare platforms on how their unit maps to the work you are automating.
Does a store-integrated agent need custom code? Not for standard retail flows. YourGPT documents native Shopify actions for product search, product cards, add-to-cart, and cart view without custom integrations. Verify what each platform's store connection actually exposes, because cart and order-data authority varies by platform and plan.
How do I know my agent's inventory answers are current? Use an integration that syncs product data from your store rather than relying only on static uploaded files. Retest after price or stock changes, and confirm out-of-stock items are visibly blocked from being added to cart.
Separate the jobs, match authority to the job, and prove every boundary. Start read-only, add cart assistance when the catalog is fresh, and reserve order writes for later, with logging and caps. A catalog bot that recommends well, a shopping helper that fills carts correctly, and an order agent that verifies identity and authorizes before it acts will outperform a general-purpose chatbot that tries to do everything at once.
