By 2030, AI agents could handle up to $5 trillion in global commerce. This reflects McKinsey’s conservative baseline estimate rather than an overly optimistic forecast. The entire customer journey is collapsing into a single moment driven by machines instead of humans.

To capture this momentum, businesses must first organize their internal processes. Currently, 88% of companies use AI, but only 39% report a clear impact on profitability. This article bridges that gap by exploring 11 specific AI use cases in e-commerce, complete with key metrics and deployment requirements. It also outlines the essential foundation needed to make autonomous commerce possible.

#1 Personalization as a Baseline Expectation Rather Than a Competitive Advantage

Many product teams assume they have already mastered personalization, which often leads them to underestimate its true impact. In e-commerce, personalization relies on three primary mechanisms: recommendation engines for related products, dynamic storefronts that adapt to individual users, and behavioral messaging triggered by customer actions rather than fixed schedules.

McKinsey data shows that these strategies go far beyond cosmetic changes. According to the Next in Personalization report, tailored experiences typically drive a 10% to 15% increase in revenue, with the overall impact ranging from 5% to 25%, depending on execution quality. Furthermore, fast-growing companies generate 40% more of their revenue from personalization than their slower-growing competitors. On a macroeconomic scale, bringing all U.S. businesses up to the level of top-quartile performers would generate more than $1 trillion in value.

The real pressure comes from the cost of inaction. Personalization has evolved from a premium feature into a baseline expectation that modern consumers simply take for granted.

What the McKinsey data shows (Next in Personalization):

The key takeaway is not the potential upside but the 76% dissatisfaction rate. Poor personalization is no longer neutral. It pushes customers toward competitors with more relevant shopping experiences. Personalization has become the cost of entry rather than a competitive advantage.

The main technical constraint is data quality. A recommendation engine cannot outperform the data it is trained on. When customer profiles are fragmented across CRM systems, email platforms, and backend databases, even the most advanced algorithm cannot deliver truly relevant recommendations. Success therefore depends far more on a unified customer data foundation than on the choice of AI model.

The figures above come from McKinsey’s 2021 research, published before the widespread adoption of generative AI. Since 2023, customer expectations have risen significantly.

The simplest level of AI delegation involves routine repeat purchases. According to McKinsey Quarterly 2026, around 23% of U.S. Amazon customers had an active Subscribe & Save subscription in 2024.

# 2 AI Search 

What e-commerce search is and why it is changing

On-site search is the primary way customers with a clear purchase intent find products without manually browsing the catalog. This is also where AI is reshaping the shopping experience most dramatically because traditional search and AI-powered search operate on fundamentally different principles.

Traditional keyword search matches the words in a customer’s query with text stored in product listings. It works only when shoppers know the exact product name or category, use the same terminology as the catalog, and make no spelling mistakes. Even a small variation can produce no results.

AI-powered search combines three distinct technologies that are often confused.

1. Semantic search understands the intent behind a query. Requests such as “a warm jacket for transitional weather under €80” or “a gift for a teenager who loves music” are automatically translated into product categories, attributes, and budget constraints, even when none of those words appear in the product title.
2. Visual search identifies products from an uploaded image. Customers simply provide a photo, and the system searches for visually similar items.
3. Automated categorization and tagging uses AI to assign attributes such as color, size, material, compatibility, and intended use. These metadata improve search, recommendations, and filtering across the entire catalog.

Search users represent the most valuable audience in an online store. They have already decided to buy and are looking for a specific product. As a result, failures at this stage have the highest business cost.

Case study: How Rozetka approached AI search

In March 2026, Rozetka launched Rozetka AI, an assistant that helps customers discover products using natural language instead of filters or keyword searches. Customers can describe what they need, specify a budget, or ask the assistant to compare products. The system analyzes the product catalog, pricing, and customer reviews before generating recommendations.

Several aspects of this launch stand out. It is the first large-scale implementation of this kind in Ukrainian e-commerce and illustrates the first-mover advantage discussed later in this article. According to Rozetka, the assistant has already processed more than one million queries, while the average product selection takes about 30 seconds. The company also positions the assistant as an enhancement to traditional search rather than a replacement.

One particularly valuable insight emerged during deployment. Customer questions vary depending on where they are in the shopping journey. On the checkout page, users ask about payment and delivery. On product pages, they ask about specifications and compatibility. This demonstrates the advantage of semantic understanding. Instead of matching keywords, the system responds to user intent within its current context.

The state of search in e-commerce

How search fails by query type

To see exactly where classic search breaks down, Baymard sorts every shopper query into eight types. For each type, researchers measured the share of sites that handle it poorly. The pattern is clear: the further a query moves from an exact product name toward natural language, the more often search fails.

An exact name like “Keurig K45 Elite” trips up only 12% of sites. But the moment a shopper describes a problem, a use case, or a feature — the way a real person actually talks — the share of failing sites doubles or triples. These are precisely the query types that semantic AI search is built to handle.

How poorly search handles different query types (share of sites with issues, Baymard Institute):

#3 AI-Powered Pricing (Dynamic Pricing)

What it is

Dynamic pricing means automatically adjusting a product’s price in real time based on demand, competitor prices, stock levels, season, and shopper behavior. Instead of a static price a manager updates by hand, the algorithm recalculates continuously.

Two main approaches:

• Rule-based — a human sets the conditions: “if a competitor drops below our price, match within 2%, but never below cost +10%.” Transparent and easy to audit, but it can’t cover every situation.
• ML-based pricing — a model trains on sales history and predicts the price most likely to hit a goal (maximize revenue or protect margin), weighing dozens of variables at once.

This also covers discount and promo optimization: when, on which product, and by how much to mark down.

What the data shows

The key point in these numbers: margins grow twice as fast as sales (5-10% vs. 2-5%). The algorithm doesn’t just cut prices to drive volume — it finds where a price can hold higher without losing demand. The cases confirm it: +10% gross margin for one player, +4.7% EBIT for another in three months.

Amazon’s number shows the other side — speed. Back in 2013, by Profitero’s estimate, Amazon changed prices ~2.5M times a day versus ~50,000 times a month at Best Buy and Walmart. The figure is old and no longer current, but it captures the gap between continuous and manual pricing.

# 4 Customer Service: Chatbots, Request Handling, Voice Commerce

What it is

AI in customer service works on three levels. First, RAG-based chatbots (retrieval-augmented generation): instead of inventing an answer, the assistant pulls it from real sources — the catalog, order history, knowledge base — so it responds on substance, not with a script. Second, automated handling of returns and complaints: checking status, processing a return, escalating complex cases to a human. Third, voice commerce: voice-based interaction and support.

What the data shows

By McKinsey’s estimate, generative AI can cut the volume of requests reaching live agents by up to 50%, depending on how automated the process already is. Gartner looks further out: by 2029, AI agents will autonomously resolve 80% of common requests, delivering roughly a 30% cut in operating costs. 

Case study: Klarna — the scale and its limit

Klarna is the most instructive public case because it shows both sides. In February 2024, Klarna’s OpenAI-powered AI assistant, in its first month:

• handled 2.3 million conversations — two-thirds of all support requests;
• did the work equivalent to ~700 full-time agents;
• cut resolution time from 11 minutes to under 2;
• reduced repeat inquiries by 25%;
• was projected to improve profit by $40 million in 2024.

(Source: Klarna press release, OpenAI.)

What happened next matters just as much. In May 2025, Klarna’s CEO told Bloomberg that the drive to cut costs through AI had “gone too far”: quality slipped on complex requests, and the company began hiring human agents so customers would always have the option to reach a person. Klarna later clarified this was more a recalibration than a reversal — by then the AI assistant was doing the work of around 800 roles, and the move was a pilot with a few experienced agents alongside the AI, not a mass return of staff.

The model that McKinsey’s data, Klarna’s case, and Gartner’s forecast all converge on is hybrid. AI takes on the routine volume (hence the −50% of requests and 80% automation figures), while humans focus on the complex. RAG matters precisely because it ties answers to the store’s real data and lowers the risk of hallucination on sensitive questions like returns and payments.

#5 Advertising and Marketing: Prediction, Lookalikes, Content Generation

What it is

AI covers three jobs in marketing. Predictive analytics forecasts who will buy, when, and who’s about to churn. Lookalike modeling finds new users who resemble existing buyers, based on behavioral data. Content generation produces product descriptions, banners, emails, videos, and ad creative.

What the data shows

Content generation is AI’s leading job in retail

NVIDIA’s State of AI in Retail and CPG (2025) reports that more than 80% of companies in the sector already use or pilot generative AI, and its most common application is creating marketing and advertising content — text, images, video. Secondary summaries of the same report put this use case at roughly 60-67% of genAI deployments in retail, though the exact figure varies by source.

#6 Logistics and Warehousing: Demand Forecasting, Routing, Fulfillment

What it is

AI works across three areas in logistics. Demand forecasting predicts how much to order and of what, to avoid both shortages and overstock. Route optimization builds efficient delivery routes around load, time windows, and traffic. Fulfillment automation speeds up picking, packing, and shipping orders in the warehouse.

What the data shows

Demand forecasting is where AI delivers its cleanest measurable impact. McKinsey (AI-driven operations forecasting):

The logic runs in sequence: a sharper forecast means less shortage and less overstock, which means lower losses. When AI cuts forecast error by 20-50%, that converts directly into a reduction in lost sales and product unavailability of up to 65% — and drags warehousing (−5-10%) and administrative costs (−25-40%) down alongside it.

Demand forecasting is the leading supply-chain AI use case

Per NVIDIA’s State of AI in Retail and CPG, demand forecasting is the leading application of AI in the supply chain, and the supply chain itself has become the number-one priority: 64% of companies reported year-over-year growth in supply-chain complexity, and the industry is responding by deploying AI for operational efficiency (51%).

#7 Fraud Prevention and Payment Protection

Why this is its own area

Personalization or search brings in extra revenue. Fraud prevention does something different: it protects money the business has already earned. The pain is direct and measurable. Every successful fraudulent transaction costs a merchant more than the price of the goods, because fees, penalties, manual review, and chargebacks pile on top. AI flags payment fraud, chargebacks, account takeover, and synthetic identities in real time, at a speed manual review can no longer match.

What the data shows

The figure that matters most here is $4.61. It shows that the stolen amount is only the surface: the real cost of fraud to a US merchant runs 4.6 times the face value of the stolen transaction, once fees, penalties, replacement goods, and manual handling are counted. Juniper puts the scale of the problem at more than $362 billion in cumulative losses across 2023–2028.

Meanwhile, 41% of North American merchants still handle this manually, and only 6% of US e-commerce runs fully automated fraud prevention. That distance between the size of the threat and the methods used to defend against it is where AI has room to work.

Why AI specifically

The primary sources are direct on this point. LexisNexis says that staying ahead of fraudsters calls for AI-powered detection and a multi-layered approach that catches fraud in real time. Juniper makes a similar recommendation for ML platforms, on the reasoning that fraudsters attack at scale and now use AI themselves, while manual review and static rules fall behind.

The mechanism is straightforward. Fraud has to be caught in the window between order and fulfillment, and that window keeps shrinking, especially for digital goods delivered instantly. An ML model scores hundreds of signals in milliseconds — behavior, device, history, context — and catches the anomaly at a speed a manual team cannot reach.

#8 Virtual Try-On (AR) and Reducing Returns

Why this is its own area

Returns are one of the biggest holes in e-commerce margin. An item comes back, but the money spent shipping, processing, inspecting, and re-listing it doesn’t come back with it. Online this hits harder than in physical stores. A shopper can’t hold the product before buying, so orders get placed just to take a look. AI targets exactly this hole through visualization (AR/3D) that lets a shopper see the product in a real setting before ordering.

Per the NRF 2025 Retail Returns Landscape report:

The gap between online and retail overall (19.3% versus 15.8%) shows where the issue sits: the online channel structurally produces more returns. For a sense of scale, 2024 saw total returns of $890 billion at a rate of 16.9%. This is a standing cost line rather than an edge case.

What AR/3D visualization delivers

The cleanest data here comes from Shopify’s own cases, where the effect was measured on real brands.

After adopting 3D and AR, the brand Rebecca Minkoff saw these shifts in shopper behavior (Shopify):

The second case is Gunner Kennels, a maker of travel crates for dogs. Here AR solves a specific sizing question: a shopper places the 3D crate model next to a photo of their dog and checks whether it fits. The result was a 40% lift in order conversion and a 5% drop in the return rate (Shopify).

Both cases follow the same logic. AR closes the gap between expectation and reality ahead of purchase. A shopper who has already seen the size and appearance is less likely to be let down on delivery, and less likely to send the item back.

AR and 3D affect margin in two ways at once: they raise conversion because the shopper feels more confident, and they lower returns because there’s less mismatch between expectation and product. The biggest payoff lands on items where size and spatial appearance are what matter most, such as furniture, appliances, bulky goods, footwear, and accessories. A good starting point is the categories with the highest return rates rather than the full catalog.

#9 Conversion Optimization and Abandoned-Cart Recovery

Why this is its own area

The traffic is already paid for. A shopper who filled a cart and left is the cheapest revenue on the site to chase. AI works here narrowly: predicting who’s about to leave, timing the follow-up, personalizing the return.

Scale of the problem

Per Baymard Institute (meta-analysis of 50 studies, updated Sept 2025):

Why carts get abandoned

Not all of it is recoverable. Baymard’s survey (n=4,384 US shoppers, multi-select) splits it:

What recovery actually delivers

Cleanest public data is Klaviyo’s Benchmark Report (143,000+ abandoned-cart flows sent in 2023):

The eightfold gap between average and top-decile RPR says the lever is execution — timing, segmentation, sequence — not the mere existence of a flow. That’s where AI operates: predicting abandonment risk, timing sends per user, generating personalized content at scale.

#10 Agentic Commerce — Where We See the Most Potential

Of the ten directions in this article, this is the one we’re most convinced about. Not because it’s the flashiest, but because it’s the layer where all the others converge: search, pricing, service, recovery, and returns each stop being a standalone tactic and become inputs an agent reads at the moment it mediates a purchase. Everything upstream is optimization; this is the shift in the interface itself. If you invest ahead of one curve, we’d argue it’s this one.

McKinsey frames the same shift at the level of the whole organization. Its Q2 2026 Quarterly argues that applying AI to isolated tasks is precisely why most deployments underperform — the gains come from redesigning entire workflows so people, agents, and robots work together. The scale of the current miss is the starting point:

The through-line to the other blocks: the gap is a workflow-design problem, not a technology one. A chatbot bolted onto a legacy process delivers little; agents embedded across a redesigned process is where value shows up.

The size of the prize

The forecast for the agentic layer specifically (McKinsey QuantumBlack, reaffirmed in the Q2 2026 Quarterly):

One caveat: “mediate” is broad — it counts goods where an agent influenced or orchestrated the sale, not only checkouts completed inside an AI. Firms drawing the line more narrowly land far lower (eMarketer ~$144B, Morgan Stanley $190–385B). Treat $3–5T as the widest frame, not a consensus.

The behavior is already measurable

Adobe Analytics (Q1 2026 report, March 2026 data, 1T+ retail visits):

The sign flip is the story: AI traffic went from worst-converting to best-converting in twelve months, because a shopper arriving from an assistant has already researched and narrowed the choice. Two caveats: the data is Adobe’s own (vendor-stated, not independently audited), released alongside a product Adobe sells to fix the problem it describes. Later months run higher (+54% in May 2026), so trust the direction and treat any single decimal as indicative.

The automation curve: where a shopper’s trust actually goes

McKinsey’s Q2 2026 Quarterly lays out a six-level curve of how much of the journey a consumer will delegate. It’s a curve, not a ladder — the goal is optimal delegation, not maximum autonomy:

Where the curve bends — the part that tells you where to start

This is the most practically useful piece McKinsey offers, and it maps cleanly onto where a merchant should move first. Delegation doesn’t advance evenly; it splits by category:

The strategic read: value pools migrate. As the funnel compresses, advantage shifts to merchants that reliably execute against agent constraints rather than those that capture human attention — margins get shaped by service guarantees, fulfillment reliability, and policy clarity. Discovery-dependent players face real disintermediation risk. This is also why we’d tell a merchant to start where their return-rate and repeat-purchase categories are utility-shaped: that’s where autonomy arrives first and operational readiness pays off fastest.

The rails are live — and still shaking out

Current as of mid-2026, not settled:

A correction to the common narrative: ChatGPT’s original Instant Checkout underperformed and was pulled — Walmart’s pilot showed in-chat conversion at roughly a third of clicking through. What made agent payments viable was Visa carrying settlement on rails merchants already trust, not the first in-chat button.

The critical thesis: readable or invisible

McKinsey states it plainly: if a retailer’s catalog, policies, and value proposition aren’t machine-readable, agents — and by extension shoppers — simply won’t find them, no matter how beloved the brand. Every level of the curve raises the bar: level 2 needs API-first merchandising; levels 3–4 need the rules and policies exposed, not just the catalog, so an agent can reason about eligibility, substitutions, and guarantees.

Adobe’s audit puts a number on the gap: retail product pages scored just 66% on AI readability, the lowest of any page type, while text-heavy pages (FAQs, returns) scored above 80%. One honest limit: the causal strength of structured data alone is contested — Ahrefs’ test on 1,885 pages found schema barely moved citations, because pages with schema already do the rest of the SEO work well. So the defensible framing: machine-readable catalog data is table stakes for agentic visibility, not a standalone lever.

Case Study

Everything in the previous ten blocks assumes a foundation most retailers don’t have: a modular, integration-ready, machine-readable platform. Without it, none of the AI directions actually run. Visdeal is what building that foundation looks like in practice.

Visdeal is a fishing-gear e-commerce business selling 80,000+ SKUs across 8+ European countries. Anadea has partnered with it for nearly 14 years — long enough to have carried it through a full platform generation. The starting point was a monolith built on the GetSocio website builder: pages rendered in up to 10 seconds, there were no APIs to connect payments, analytics, or warehouse tools, and eight separate country sites ran independently, each needing manual content duplication and its own tax, shipping, and pricing setup.

The rebuild replaced that monolith with a modular architecture on Ruby on Rails and Elixir/Phoenix. Critical components — shopping cart, product catalog — were rewritten from scratch while staying backward-compatible to avoid downtime. A Self-Contained Systems approach let one market update without touching the others, and Micro Frontends adapted each country’s design to local preferences. A central admin layer handled AI-powered catalog translation via DeepL, and ChannelEngine automated listings on Amazon and Bol.com with centralized inventory, settled through the Adyen payment gateway.

Conclusion

The pattern across all eleven directions is the same: the companies getting real results stopped adding AI to single tasks and fixed the workflow and data underneath. The direction with the most upside is agentic commerce, because it’s where everything else comes together: once an agent handles the purchase, your search, pricing, catalog, and returns all become one thing the agent can either read or skip. McKinsey puts it plainly — if your catalog and policies aren’t machine-readable, agents and their shoppers won’t find you. It’s worth getting ready now.