AI agents for e-commerce: from order questions to catalog upkeep

AI agents for e-commerce pay back fastest in two places. The first is the small set of requests that make up most support volume: where is my order, can I change it, and how do I return it. The second is the questions shoppers ask before they buy, where a fast, correct answer removes a reason not to order. Behind the scenes, agents also keep the catalog complete and turn reviews into a list of things to fix.

This guide is for store owners and e-commerce managers. It covers the jobs worth handing over, the guardrails, planning for peak season, and how to choose between the agent in your platform and a custom build.

#Customer-facing jobs

• Order status. The agent reads the real tracking events and explains them in plain words, including what happens next when a parcel is delayed.
• Order changes. Address corrections, adding a note, or cancelling, up to the moment the order is released to the warehouse. After that it explains the options.
• Returns and exchanges. It checks the order against your policy, starts the return and sends the label. Where an exchange makes sense, it offers one first, which keeps the sale.
• Pre-purchase questions. Sizing, materials, compatibility, stock, delivery dates to a given ZIP code. These are asked at the moment of decision, often outside working hours.
• Subscriptions. Skip, pause, swap, change the date. Customers who can do this easily are less likely to cancel.
• Back in stock and delivery problems. Proactive messages, where the customer has opted in.

#Behind-the-scenes jobs

These get less attention and often pay back sooner, because no customer sees the first draft.

• Catalog upkeep. Writing and improving product descriptions, filling in missing attributes, image alt text, translations. The agent also finds the gaps: products with no dimensions, no materials, no care instructions.
• Review mining. Reading every review and support ticket, and reporting the recurring themes by product: "runs small", "arrived damaged", "instructions unclear". This is a product improvement list that usually sits unread.
• Review replies. Drafts for a person to approve.
• Search tuning. The searches on your site that return nothing, and the synonyms that would fix them.
• Chargeback evidence. Collecting the order, delivery proof and correspondence into a response.
• Supplier follow-up. Chasing purchase order confirmations and delivery dates.

#What should stay with people

• Damaged, unsafe or faulty products.
• Refunds above a set amount, and anything outside policy.
• Decisions about suspected fraud.
• Wholesale and high-value customers.
• Pricing and promotions.

#Guardrails for a store

• Stock and delivery promises come from your systems. The agent never guesses when something will arrive or come back into stock.
• Limits on money. Refunds, credits and free shipping up to an amount you set. Above it, approval.
• No discount codes on request, unless you have written a rule for when one is offered.
• No card details in chat. Payments stay inside your checkout.
• Product claims come from your product data. This matters most for health, safety and children's products.
• Consent for marketing messages. Text and messaging-app marketing need opt-in in most countries. Service messages about an order are treated differently, so keep the two apart.
• Say it is an AI, and offer a person.
• Watch for returns abuse. The agent should flag patterns, not decide on them.

#Planning for peak season

Peak is when an agent is most valuable and most risky.

• Be live early. At least six weeks before your peak, so the problems appear while volume is low.
• Test against last year. Run last peak's real questions through it, especially delays, stock-outs and gift returns.
• Agree the rules in advance. What do you offer when a parcel will miss the date? Decide before the rush.
• Check the bill. Per-conversation pricing at five times normal volume can surprise you. Work it out in October, not in January.
• Freeze changes during the peak, exactly as you would for the store itself.

#Platform agent or custom build?

Products with rules are the most common reason for custom work. "Will this part fit my 2019 model?" cannot be answered from a help center. It needs your compatibility data, and an agent that knows to ask for the model before answering.

#Being found by other people's agents

More shopping now starts in an AI assistant instead of a search box, and some assistants can complete a purchase. Whether an assistant recommends your product depends largely on whether it can read your product data. Complete attributes, accurate prices and stock in your product feeds, and structured data on product pages are the basics. The catalog work above helps twice: once for your own customers and once for everyone else's agents. If this matters to you, our AI Search Visibility work covers it.

#What to measure

#What it costs

• Platform and help desk agents are usually charged per seat or per resolved conversation, often around $1 to $2 per resolution at list price. A store with 3,000 resolved conversations in a normal month and 12,000 in its peak month should budget on the peak.
• A custom e-commerce agent typically costs $8,000 to $40,000 to build, depending on how many systems it touches. Running costs are often a few cents per conversation in model usage, plus hosting and a few hours of upkeep a month.
• Catalog and review automations are smaller projects, often a few thousand dollars.

The full breakdown is in how much does an AI agent cost. For reference, our own automation projects start from $4,800 and custom AI agents from $9,800, each quoted as one fixed price before work begins, with a 30-day fix window.

#Where to start

Export three months of support tickets and count them by type. If order status, changes and returns are most of the volume, and your platform's agent can see your orders, switch that on first. If the answers your customers need live in systems your platform cannot reach, book a free 20-minute call and we will tell you what a build would involve. For the support side in more depth, see AI agents for customer service.

This guide is part of our series on AI agents for business.

Common questions.

• What can AI agents do for an online store?

  For customers: answer product, sizing, stock and shipping questions, give real order status, change an address or cancel before dispatch, handle returns and exchanges inside your policy, and manage subscriptions. Behind the scenes: write and complete product data, translate listings, reply to reviews, spot recurring product problems in reviews and tickets, and prepare evidence for chargebacks.

• Which AI agent is best for e-commerce?

  For most stores the best first step is the agent built into the store platform or help desk, because it already connects to orders, customers and products. Shopify, BigCommerce and the main e-commerce help desks all offer one. A custom agent is worth it when stock, pricing, compatibility or returns rules live in systems those products cannot reach.

• Will an AI agent increase sales or only cut support costs?

  Both are possible, and they come from different jobs. Order status, changes and returns reduce support workload. Answering pre-purchase questions quickly, such as sizing, compatibility and delivery dates, removes reasons not to buy, and offering an exchange in place of a refund keeps revenue. Measure the two separately, because they pay back in different ways.

• How should an online store prepare an AI agent for peak season?

  Have it live and stable at least six weeks before your peak. Test it against last year's peak questions, including delays and out-of-stock items. Agree the rules for late deliveries in advance. Check what per-conversation fees will cost at peak volume. Then freeze changes during the peak itself, as you would for the store.

• How much does an e-commerce AI agent cost?

  Platform and help desk agents are usually charged per seat or per resolved conversation, often around $1 to $2 per resolution at list price, which adds up during peak season. A custom e-commerce agent typically costs $8,000 to $40,000 to build, with running costs of a few cents per conversation. Catalog and review automations cost less.