.png)
A customer asks a retailer's AI agent whether a jacket can arrive tomorrow. The agent answers at once, politely and with confidence. Whether it's right depends on stock at the nearest warehouse, tonight's picking capacity and the carrier's cut-off, and the customer can see none of it.
Thirty years of e-commerce gave customers access to far more products and control over when and how they buy. Most retailers got there by adding to what they already had, putting new channels, systems, teams and fulfilment options on top of a business built for stores. Shopping became more convenient, and the business behind it became harder to run.
E-commerce also moved work onto the customer: finding the right product, comparing options, choosing a delivery slot, chasing a late parcel, arranging a return. AI agents now take on much of that work for the customer. Every answer they give draws on the order management, stock and pricing systems behind them.
Better CX has to come with better economics
Margins are thin and labour costs keep rising. Omnichannel promises rest on systems that were never designed to work together, and customers still expect lower prices and faster service. An AI project therefore has to improve the experience and make the business cheaper to run at the same time. One that delivers only one of those is hard to fund past the pilot.
Every retail experience has a line of visibility. Above it are the moments a customer notices, from getting help choosing to being recognised next time. Below it sit the data, decisions, processes, systems and people that produce each of those moments.
Anyone who has worked on a shop floor knows that the best customer-facing colleagues quietly fix what the back end failed to deliver. They make the customer believe there was never a problem. That skill is valuable, and it is also how a broken operation stays hidden for years.
An AI agent can play the same role at scale, answering the jacket question fluently whether or not the jacket is in stock. The real job for AI is to fix the cause. Hiding the symptom more effectively is the easy version, and it's what most retailers will get by default.
A common mistake is to automate the interface and leave the rest of the business as it was. A conversational assistant can work out what the customer wants. It creates value only when it can also reach reliable product information, see real stock, respect pricing rules, trigger fulfilment or service actions, and remember the customer from one interaction to the next.
Without those connections, the retailer has built a polished front door to the same fragmented operation. The customer gets a fluent reply and the same outcome as before. What they experience is whether the retailer understood the request and followed through.
The retailer also owns whatever the agent says. In 2024 a Canadian tribunal held Air Canada liable for a bereavement refund rule its website chatbot got wrong, and rejected the airline's argument that the chatbot was responsible for its own words. An agent that quotes a delivery date, a price or a returns policy is making a commitment on the retailer's behalf. The data behind it needs to be as reliable as a contract.
Here is what customers ask, and what a good response takes:
Take availability. A customer told that an item is out of stock has received information. What they wanted was a suitable alternative they can get hold of, with a delivery date the retailer will hit. In grocery the same decision happens thousands of times a night as pickers reach empty shelves: which substitute will this customer accept? Getting it right means combining what the retailer knows about the customer with stock, fulfilment, pricing and service rules in one decision, then carrying out the action that follows.
The value sits in that decision and the completed action. Customers get problems solved with less effort and can trust the answer. Retailers see it in conversion, cost-to-serve, sell-through and, over time, customer lifetime value.
The agents covered so far belong to the retailer. A growing share will belong to the customer. Assistants from OpenAI, Google and others can already search, compare and, in some cases, check out for a shopper. They never see the homepage, the merchandising or the brand campaign. They read product data, prices, stock and delivery promises, then choose.
A retailer with incomplete product attributes or an unreliable stock feed loses the sale to a competitor whose data the agent trusts, and never learns the customer was there. The work that makes your own agent useful, accurate data that other systems can reach directly, is the same work that gets you chosen by someone else's.
In customer care, "where is my order" contacts are often the largest single category. An agent that tells customers about a delay before they ask removes that contact and frees the team for cases that need a person. In stores, the stock count a colleague corrects at 7am becomes the delivery promise an agent makes at 9pm, so stock accuracy becomes part of the customer promise.
People keep the work that needs judgement, such as handling exceptions and deciding what the agent is allowed to do. Agents take on the chasing and re-keying between teams. AI will also show where the operating model is broken, because adding an agent to a poorly designed process makes the wrong process run faster.
Where should we start? Pick one customer question where the quality of the answer depends on operational data, such as availability or a delivery date. When we start this work with a retailer, we map that single journey on one sheet, from the customer's question down to the stock file, and mark every point where a handoff fails. The customer-facing piece usually turns out to be the smaller part of the job.
83% faster delivery. Under 10% rework. See exactly how Xceptor got there.
