AI Agents for Ecommerce: Support, Catalog, Carts and Returns

AI agents for ecommerce earn their keep where an online store bleeds margin every day: post-sale support drowning in "where is my order?", half-finished product pages, carts abandoned with nobody chasing them, and returns processed by hand. An AI agent connected to your platform — Shopify, WooCommerce, PrestaShop or a custom build — plus your carriers and payment gateway can operate most of that work end to end, leaving the human team focused on what actually differentiates the store.

Ecommerce is probably the sector where the gap between "adding a chatbot" and deploying agents shows most: volume is high, the data lives in connectable systems, and almost everything is measurable in euros. Here are the five fronts where we see the fastest return, with cost ranges and the limits worth respecting.

Why is WISMO the first process to automate?

Because "where is my order" (WISMO) inquiries typically account for 25-40% of all tickets in an online store, and they're the most mechanical work in support: look up the order, check the shipment status, type the same reply for the thousandth time. An agent connected to the order system and the carriers' APIs resolves the full inquiry in seconds, at any hour, in the customer's language.

The well-built version goes beyond answering: the agent spots the incident before the customer does. If tracking hasn't moved in 48 hours or the carrier logs a failed delivery, the agent writes first — "your parcel has been sitting at the X hub for two days; we've opened a claim" — files the claim with the carrier and follows it up. Turning a problematic delivery into impeccable communication is one of the most profitable review-and-repeat-purchase levers in the sector.

With WISMO and the FAQ handled by the agent (sizing, stock, delivery times, payments), 60-75% of a typical store's tickets get resolved without a human, and cost per ticket drops from the €3-6 of manual support to cents. The remaining team handles the cases with nuance: serious complaints, B2B orders, product advice.

Product pages: the catalog as a process, not a project

Writing and maintaining product pages is the eternally pending task of every store with more than a hundred SKUs. An agent turns it into a continuous process: from supplier data, specs and your tone rules, it generates title, description, attributes, SEO metadata and image alt texts for every new SKU, in every store language, and keeps existing pages current when data changes.

The nuances that separate this from "asking ChatGPT for descriptions": the agent works against the real catalog (it doesn't invent attributes — if the data isn't there, it requests it or flags it), applies your category rules, and routes sensitive pages — hero products, regulated items — to a human review queue. For a store listing 50-200 SKUs a month, that's dozens of hours freed monthly and, more importantly, an entire catalog optimized for search engines and for the AI shopping assistants that are sending a growing share of traffic.

Can you hand pricing over to an agent?

Partially — under rules you write. A pricing agent watches competitors and costs and proposes or applies adjustments within the limits you define: minimum margin per category, maximum daily variation, lists of untouchable products. The key phrase is "within limits": unconstrained repricing ends in penny wars or in margins destroyed by one bad data point.

The healthy split we recommend: the agent executes the tactics (follow the rule, catch the competitor's change, flag anomalies) and the human owns the strategy (positioning, promotions, how far to compete on price). In stores with large, comparable catalogs — electronics, spare parts, para-pharmacy — this piece alone can pay for the whole project; in own-brand stores with no direct comparables, it's the piece you need least.

Abandoned carts, recovered with context

65-75% of carts are abandoned, and the standard recovery — three generic automated emails — stopped pulling its weight years ago. An agent improves the two levers that matter: the reason and the channel. On the reason: it reads the session's behavior (where the customer hesitated, what they compared, whether they reached shipping costs) to pick the message — resolve the sizing doubt, restate the returns policy, flag genuinely low stock — instead of firing the same template. On the channel: it converses where the customer replies, including WhatsApp where regulation and consent allow, and answers the doubts that blocked the purchase inside that same conversation.

An honest design rule: recovery is not harassment. One or two useful touches that resolve the actual doubts beat five impacts with a fake countdown. And discounts, if used at all, run on the agent's judgment but your policy: cart-value threshold, new versus returning customer, product margin.

Returns: the process nobody wants and everybody measures

Returns are expensive (€7-15 of operational cost per return is a common range before reverse logistics) and their manual handling is pure paperwork: validate window and condition, generate the label, track the parcel, approve the refund, update stock. An agent runs the entire flow against your rules — windows per category, exceptions, auto-refund thresholds — and escalates the cases with friction: suspected fraud, disputed damage, high-value customers.

The side benefit is the data: the agent classifies the real reason behind every return ("runs small" on one product, "not as pictured" on another), and that monthly report is gold for purchasing and for the product pages themselves — closing the loop: better pages, fewer returns.

Costs and where to start

Ranges assume standard platforms with APIs (Shopify, WooCommerce, PrestaShop); custom builds add cost depending on the state of their internal APIs. Maintenance: 10-20% per year. Inference for a mid-sized store usually lands between €50 and €300/month depending on conversation volume. Starting with WISMO support is the default recommendation — fastest return, least disruption; the full priority order by store type is in our guide to AI agent use cases by department, and the general ranges in how much a custom AI agent costs.

What NOT to automate in an ecommerce operation

• Emotionally charged complaints and public crises. The agent detects and prepares context; a person replies.
• Assortment and brand decisions. What to sell, how to position, the brand's voice — that's strategy; the agent executes it, it doesn't define it.
• Borderline fraud. The agent flags patterns and blocks the flagrant cases by rule, but the grey zones — long-standing customers behaving oddly — get human review before anyone is accused.
• Refunds outside policy. Exceptions to the returns policy: always human-approved, even when the agent prepares the case.

If you want all of this grounded in your own store's numbers — tickets, catalog, abandonment rate, cost per return — that baseline diagnosis is how we start every project at our AI agency in Madrid.

Frequently asked questions

Does it work with Shopify, WooCommerce or PrestaShop?

Yes: all three expose APIs that agents work with as standard, as do the usual carriers and payment gateways. On custom platforms, feasibility depends on the state of the internal APIs — it's the first thing checked in the diagnosis.

Can the agent also handle WhatsApp and Instagram?

Yes, and in ecommerce that's where it pays most: WhatsApp via the Business API (with the templates and consents it requires), Instagram and Messenger through their messaging APIs. The key is one agent with one context across all channels — not three separate bots that don't know what the others said.

How long until the support agent is live?

A WISMO + FAQ agent connected to orders and carriers deploys in 4-6 weeks, including the run-in phase with human review. The next pieces (carts, returns, product pages) are added in 3-5 week phases, reusing the same foundation.

What happens during peaks like Black Friday?

That's where the agent makes the biggest difference: it absorbs 5x or 10x volume with no queue and no temp hires, and keeps its tone at 3 a.m. on the Friday. The preparation is human, though: updating policies, promotions and campaign answers before the peak is part of maintenance.

Don't customers prefer talking to people?

They prefer getting resolved: fast, first time, without repeating their order number. For the mechanical 70% of inquiries, the agent resolves better than a queue; for the nuanced 30%, the person serves better precisely because they're no longer buried in WISMO. Identifying the assistant as AI and offering a frictionless path to a human is mandatory — legally too, under the EU AI Act.