In high-volume retail, every conversation with a customer contains an operation waiting to be completed. An order status to check, a return to open, a warranty to activate, a lead to call back before they buy elsewhere.
To carry them out, artificial intelligence in enterprise retail has taken the shape of a team of specialised AI Agents. They cover the first line of contact on chat, WhatsApp, and voice, carry out these operations directly on the company's systems, and hand over to people the cases that call for a decision, delivering them with the context already gathered.
Their value is measured on response time, conversion, and cost per contact, three indicators that pull against each other at peak times. An enterprise retailer adopts AI Agents to keep the three in balance all year round. Peaks are simply the moment that balance is put to the test, because volumes make it impossible to hold with headcount alone.
The pressure has grown on both volumes and customer expectations. For the days between Black Friday and Cyber Monday 2025, the Netcomm–Politecnico di Milano's B2c eCommerce Observatory estimated online spending of €2.2 billion in Italy, 7% more than the year before, with the most competitive retailers able to take up to three times the revenue of an average day in a single day.
Over the same season, according to Adobe, traffic reaching US retail sites from generative AI tools grew by 693% year on year. There are more customers, and they are better informed and less patient than a year ago. Answering is no longer enough. The job is to execute.
The peak does not end with Black Friday
The commercial peak lasts a few days. The service peak lasts for weeks. For the ten days around Black Friday 2025, Netcomm's Delivery Index estimated 41.5 million parcels moving across Italy, 4.8% of a full year's volume. Parcels generate questions about delivery, and a predictable share of them comes back.
The fourteen-day right of withdrawal pushes the wave of returns into December and, for Christmas purchases, into January.
A retailer's customer care, therefore, faces two peaks in sequence. The first, short and intense, is made of pre-purchase questions and confirmations. The second, long and costly, is made of order status queries, delays, returns, and refunds.
An organisation whose capacity depends on headcount meets this asymmetry with overtime, temporary reinforcements, and waiting times that keep growing. AI Agents absorb the asymmetry, because they handle a thousand conversations with the same quality as a hundred, without sizing headcount for the busiest day. In high-volume retail, the return on investment is concentrated in precisely the weeks when the cost per contact would otherwise be highest.
From answering to executing. What AI Agents actually do in retail
For a long time the central question was the quality of the answer. Today, understanding language is a baseline requirement, and what separates one solution from another is what the AI Agents execute on the company's systems.
Orders, returns, and payments online
In e-commerce, AI Agents read the real status of an order from the order management system and the carrier, not from an FAQ page, and reply with the updated delivery date. They open a return within the permitted window, generate the label, book the collection, and confirm when the refund has been issued. They cancel an order for as long as logistics allow, correct a shipping address, and check the status of a payment.
Every operation takes place with permissions defined for that single action, and every action is logged.
Warranties and after-sales appointments
In physical retail and technical after-sales, the work is different, but the logic is the same. AI Agents activate a warranty from the receipt or serial number, book an installation or service appointment by reading the technicians' schedules, guide the customer through using the product based on the manuals, check availability in store, and reserve a click-and-collect pickup. When a need arises that calls for specialist expertise, they pass the conversation to a consultant along with the information already gathered.
Integration as a precondition
All of this depends on integration with the systems that matter: the order management system, the CRM, the ticketing platform, the carriers, the payment systems, and the leading e-commerce platforms. Which operations the Agents can carry out depends on which of these systems are connected. Without these connections, the Agents remain an information channel. With them, they become operational capacity.
Don't lose the lead when it is worth the most
On the commercial side, the decisive parameter remains Lead Response Time, the interval between a prospect's inquiry and the company's first reply. During promotional peaks, this interval stretches at exactly the moment purchase intent is highest, and a lead who hears nothing within a few minutes is likely to buy elsewhere.
Follow-up, qualification, and appointments
AI Agents intervene at three points.
- Follow-up. With the customer's consent, they follow up within minutes with anyone who has abandoned a basket or filled in a form, on WhatsApp or by phone, including when the sales team is off shift, and always within the time windows the company has set.
- Qualification. They qualify the request and, where the sale is self-service, guide the customer through to purchase.
- Appointment. Where the product calls for advice, as it does in furniture, home appliances, or configurable services, they book an appointment with a consultant and pass on the context gathered so far, so the conversation starts from the right place.
Customers arriving from AI assistants
There is a further reason to guard this moment. A growing share of customers reaches the site after comparing products and prices with a general-purpose AI assistant. These assistants are also trying to let shoppers complete the purchase inside the conversation, without going through the seller's site, but the first of those experiments was withdrawn after around six months, with a product range that remained limited and product information that was not always up to date.
For now, the direction is this. Product discovery increasingly happens in AI, while purchase and service stay with the brand, on its own site and channels. A customer arriving from an AI assistant has already compared, already read, and expects to find an equally capable counterpart on the other side. A brand without its own AI Agents receives qualified traffic and serves it worse than the channel that delivered it.
A team of AI Agents, not a jack-of-all-trades assistant
In the most advanced enterprise architectures, the single general-purpose bot is giving way to a team of specialised AI Agents, each responsible for one stage of the customer journey. One covers orders, tracking, and returns, another handles technical support on complex products, another manages loyalty cards, promotions, and store availability, and another takes care of commercial follow-up. An orchestration layer decides which Agent steps in, when, and with which permissions, and the Agents pass the conversation between them without the customer having to repeat anything.
The role of each channel
Each channel has a precise role. Web chat catches the visitor at the point closest to purchase intent. WhatsApp is the asynchronous, personal channel for follow-ups, delivery updates, and reminders. Voice remains the channel for complex cases and advisory sales, and it is also where peaks are felt most, because a telephone queue cannot be hidden.
AI Agents take the recurring calls, such as the status of a delivery, a store's opening hours, or the start of a return, and leave to operators the ones where a person is needed, such as a dispute, an advisory sale, or a customer who has already called twice.
Priority rules at peak
At peak times, the orchestration applies priority rules that the company has defined in advance. Loyal customers and high-value orders move to the front, repetitive requests are pre-empted with proactive messages on delivery status, and cases that meet the escalation criteria set by the company are transferred to an operator before the customer walks away.
The knowledge changes with every campaign
In retail, the information that feeds AI Agents changes with every campaign, at a pace few other sectors know. Discount codes expire, catalogues are refreshed with every drop, return policies are extended for the Christmas period and then reverted, and delivery times shift with carrier load. A knowledge base prepared in October and then abandoned is already wrong by December.
Owners, versions, and validity dates
The most mature solutions manage the knowledge base through an editorial process. Every piece of content has an owner, a version, and a validity date, so that an expired promotion stops being offered on the day it expires, not when someone happens to notice.
Data that changes constantly, such as prices, availability, and order status, is never copied into the knowledge base but read from the systems through integrations.
And the conversations themselves become the diagnostics. Questions left unanswered during the peak map out the content that needs adding, and Self-improving Agents flag them to the team, which decides which updates to approve before they go into production. The next peak starts from a better knowledge base than the one the last peak ended with.
Retail rules and the AI Act. The person decides, the Agents execute
Retail operates within precise rules, which for AI Agents are a hard limit. The fourteen-day right of withdrawal for distance purchases and the two-year legal guarantee of conformity set out in the Italian Consumer Code, the GDPR on purchase and payment data, and, since 2 August 2026, the transparency obligations of the AI Act, starting with the most visible of them. Customers must know they are talking to an artificial intelligence system.
The way the most advanced companies meet this framework has a name, human in the lead.
AI Agents execute what the policies provide for, such as the return within the deadline, the refund by the book, and the warranty activation once the requirements are met. People decide what the policies do not provide for, such as an exception in favour of a long-standing customer asking for a late return, a disputed refund, or a complaint about a product that arrived damaged.
Every action taken by the Agents is logged and traceable, and every decision belongs to a person who received the case fully prepared. It is an arrangement that withstands an audit and keeps the customer relationship in the hands of those accountable for it.
Evidence from the field
Two cases show the conditions on which resilience at peak depends: a high automation rate on large volumes, and complex product knowledge, made of catalogues, price lists, and promotions that change over time.
Würth Italia has entrusted AI Agents with support for professional customers browsing its website, guiding them to the right point of contact among the website, the stores, and sales representatives. The Agents handle 96% of chats fully autonomously, and more than 30,000 unique users interacted with them in the first six months.
Flou, which serves thousands of specialist retailers in 58 countries, has brought to WhatsApp AI Agents that answer questions on product sheets, quotes, and assembly instructions and keep the network updated on launches and current promotions, reaching an 88% automation rate.
A scope that grows with the data
There is no single right place to start. Some companies begin with order status and returns, because that is where the volume is. Others with lead and abandoned-basket follow-up, because that is where the value is. Others again with warranties and after-sales appointments, or with voice for the recurring calls to the helpline.
The criterion is the same in every case. Choose the flow where the data shows the greatest impact, go live a few weeks ahead of the busiest season, and use that time to measure and correct. The scope then widens as the results confirm it.
What to measure
Metrics are set before go-live and monitored continuously. For customer service, the automation rate, first response time, and satisfaction at the end of the conversation. For sales support, Lead Response Time, the conversion of interactions handled by the Agents, and the average value of the orders generated. For both, the cost per contact and the share of conversations transferred to an operator, which measures how much load the Agents genuinely absorb.
Operational capacity all year round
Peaks remain the harshest stress test for a retail organisation. But the companies that have adopted AI Agents did not do so to get through a promotional weekend. They did so to build an operational capability that works all year round, at the same cost per contact in January as in November.
FAQ
What AI solutions does a high-volume retailer need?
A team of specialised AI Agents, integrated with the order management system, the CRM, the ticketing platform, the carriers, and the payment systems, and active on web chat, WhatsApp, and voice. The Agents execute recurring operations, such as order status, returns, refunds, warranties, appointments, and lead follow-up, and transfer to people the cases that call for a decision, with all the relevant information. The solution should be judged on response time, conversion, and cost per contact, and on its ability to keep them in balance at peak.
How do AI Agents help retail through the Black Friday and Christmas peaks?
Their capacity grows with volumes and shrinks when the peak subsides, without overtime or temporary reinforcements. During the commercial peak, they follow up with leads within minutes and answer pre-purchase questions. In the weeks that follow, they absorb the wave of order status queries, delays, returns, and refunds by reading live data from the systems. Priority rules defined in advance move loyal customers and high-value orders to the front and transfer cases that meet the escalation criteria to an operator.
What does it take to integrate AI Agents into an enterprise e-commerce stack?
Connectors to the retailer's operational systems, namely the e-commerce platform, the order management system, the CRM, carriers, payments, and the help desk, with permissions defined for every operation. On the security side, SSO, access control, data retention rules, and separate environments for development, testing, and production. On the content side, a versioned, approved knowledge base with validity dates for promotions and seasonal policies. On these foundations, a company goes live on one high-impact queue and widens the scope based on the results.
