Chapter 8 — Retail and e-commerce

Retail and e-commerce are the domains where machine learning has been longest-deployed at consumer scale. Amazon’s collaborative-filtering recommendation system, the canonical reference for the field, was running in production by 1998 (Linden et al., 2003). Netflix’s recommendation system has been a research-and-deployment focus since the 2006–2009 Netflix Prize. YouTube’s two-tower neural recommendation architecture (Covington et al., 2016) and TikTok’s For You algorithm have shaped the contemporary default. By 2024, recommendation-system techniques touched essentially every major e-commerce platform, every content platform, and many adjacent products. The deployment maturity makes retail a useful contrast case to healthcare (Chapter 7), where deployment is structurally constrained, and to finance (Chapter 6), where deployment is high-stakes and heavily regulated. Retail is where AI deployment is normal.

The 2022–2026 generative-AI wave changed the deployment landscape in two distinct directions. The personalisation track — using foundation models to produce individually-tailored content (descriptions, recommendations, search results, ads) — extended the collaborative-filtering tradition into a regime where the personalised content itself is generated rather than retrieved. The agentic track — using foundation models to act on behalf of merchants or customers (drafting listings, handling support inquiries, completing purchases) — opens a new operational frontier whose maturity is uneven. Both tracks have produced commercial successes and prominent failures; the field’s most-cited cautionary case (Klarna’s 2024–2025 AI customer-service reversal) is from this period.

This chapter develops the retail-and-e-commerce AI landscape across thirteen sections. Section 8.1 traces the recommendation-systems arc that defines the field. Section 8.2 develops the Stitch Fix case in detail, as the canonical data-flywheel deployment. Section 8.3 covers Amazon as the AI-native incumbent. Section 8.4 develops the Klarna AI customer-service post-mortem, the field’s most-detailed contemporary cautionary case. Section 8.5 covers the Southeast Asian platforms (Shopee, Lazada, Tokopedia). Section 8.6 covers generative AI in retail operations. Section 8.7 covers visual and image-based commerce. Section 8.8 covers conversational and agentic commerce, the 2024–2026 frontier. Section 8.9 covers demand forecasting and inventory. Section 8.10 covers fraud and trust-and-safety. Section 8.11 covers operations — warehousing, last-mile, returns. Section 8.12 covers Australian and Malaysian regional context. Section 8.13 sketches the 2026 frontier and what comes next.

8.1 The recommendation-systems arc

The field’s foundational reference is Amazon’s collaborative-filtering paper (Linden, Smith, and York, 2003), which described the item-to-item collaborative filtering algorithm Amazon had been running in production since 1998. The paper’s contribution was less the algorithm than the demonstration that recommendation systems could operate at consumer-internet scale (millions of users, millions of items) on commodity infrastructure. The Amazon paper became the benchmark against which subsequent recommendation-system work was measured.

The Netflix Prize, announced October 2006, offered USD 1 million for an algorithm that improved on Netflix’s existing Cinematch recommender by 10% on RMSE for movie ratings. The competition produced three years of methodological development before the BellKor’s Pragmatic Chaos team won in September 2009. The technical contribution was matrix factorisation — representing users and items as vectors in a shared latent space, with predicted ratings as the inner product (Koren, Bell, and Volinsky, 2009). The methodological influence outlasted the prize itself; matrix factorisation became the standard recommendation-system approach through the early 2010s.

The deep-learning era reshaped the field. YouTube’s two-tower neural architecture (Covington, Adams, and Sargin, 2016) replaced matrix factorisation with deep networks that could ingest much richer feature sets — watch history, search queries, user demographics, item metadata, contextual signals. The two-tower pattern (one tower encoding the user; one tower encoding the candidate item; their dot product producing the score) became the default for production recommendation at internet-scale firms. Pinterest, Instagram, LinkedIn, and most major recommendation systems now use variants of this architecture.

TikTok’s For You algorithm (operationalised by ByteDance from 2017) extended the neural-recommendation paradigm with two distinguishing features: extreme reliance on implicit signal (watch time, completion rate, scroll behaviour) over explicit signal (likes, follows); and a recommendation-as-the-product positioning, in which discovery rather than search or follow-graph drives essentially all engagement. The 2024–2026 dominance of TikTok in attention markets — and the regulatory responses, including the 2024 US ban-or-divest legislation — reflects the strategic value of the recommendation system as the platform’s primary moat.

The 2022–2026 generative-AI wave overlays foundation-model capabilities on this architecture. Large language models can be used to embed text richly (semantic search across product descriptions; cross-language recommendations; better understanding of user queries); to generate personalised content (product descriptions tailored to user segments; personalised marketing copy); to produce conversational interfaces over the recommendation infrastructure (the “shop with AI” pattern emerging from Shopify, Amazon, and various platforms in 2024–2025); and to support agentic shopping (where the AI completes purchases on the user’s behalf). The architectural pattern is hybrid: foundation models supplementing rather than replacing the underlying recommendation infrastructure.

The competitive dynamics in 2026 reflect this architectural layering. The platforms with the strongest underlying recommendation systems (Amazon, TikTok, Meta) extend their moats by adding foundation-model capabilities; the foundation-model-only entrants (e.g., Perplexity Shopping, OpenAI’s commerce features) compete on capability but lack the data flywheel that the platforms have accumulated. As Chapter 3’s framework predicts, the long-run advantage accrues to whoever turns the data flywheel fastest; foundation-model capability is increasingly commoditised, while data-flywheel position is structural.

8.2 Stitch Fix — the canonical data-flywheel case

Stitch Fix, founded in 2011 by Katrina Lake while she was a Harvard Business School student, is the most-detailed public case of an AI-native consumer business that built a durable data flywheel. The company’s positioning was personal styling delivered via a subscription “Fix” — a box of clothing items selected for the customer, who keeps what they want and returns the rest. The business model was structurally amenable to AI from the start: every Fix produced an outcome (kept vs returned, with reasons) that could train the recommendation system; every customer interaction produced data; every shipment was an experiment. The business was, in Iansiti and Lakhani’s (2020) characterisation, an AI factory by design.

The Concierge MVP. Lake’s initial operation was a Concierge MVP in the Chapter 21 sense: she hand-picked clothes for friends and shipped them from her Cambridge apartment, learning what selection criteria mattered before any algorithmic recommendation system existed. The Concierge stage ran for approximately 12 months (2011–2012). The lessons accumulated during this period — what styling questions to ask, what silhouettes worked for what body types, what price points converted — became the design specification for the data infrastructure that followed. The methodological discipline established the pattern: build the data infrastructure to capture the structured data that the human stylists were already learning to use.

The styling data infrastructure. From 2013 onward, Stitch Fix built what the company called the Algorithms team (later renamed Algorithmic Merchandising and Data Platform), eventually one of the largest in-house data-science operations in retail. The infrastructure captured 60+ structured data points per Fix: the customer’s stated style preferences, body measurements, lifestyle context, prior keep/return decisions, free-text feedback per item, frequency preferences, budget ranges, and many more. Each shipment generated an outcome label per item — kept (with the implicit endorsement) or returned (with a reason from a structured menu). The data-collection design meant that every customer was simultaneously a paying customer and a labelled-data-point contributor. Brad Klingenberg, Stitch Fix’s chief algorithms officer (2013–2024), built and led the team at peak scale of 80+ data scientists.

The algorithmic merchandising approach. The Stitch Fix system used a hybrid of machine-learning recommendations and human stylists. The ML system narrowed the inventory of 100,000+ SKUs to a candidate set of perhaps 30 items per customer per Fix; a human stylist made the final selection of 5 items from the candidate set, with the ability to override or add items. The hybrid was not a bug or a transitional arrangement; it was the architecture. Pure ML recommendation would have missed the contextual nuance that human stylists could add (a customer’s recent comment about an upcoming event; a specific outfit goal); pure human curation would not have scaled. The hybrid maximised what each could contribute. Lake, in her 2018 Harvard Business Review article, characterised the approach as “human plus machine.”

The IPO and scaling. Stitch Fix went public on NASDAQ (SFIX) in November 2017 at USD 15 per share, a USD 1.6 billion valuation. Revenue grew from USD 730 million in fiscal 2017 to USD 2.1 billion in fiscal 2021. The COVID-19 pandemic produced a scaling shock — fashion subscription was somewhat counter-cyclical in 2020, and the company’s algorithms struggled to adapt to abrupt category-mix shifts (loungewear up, professional wear down) — but the company recovered through 2021.

The 2022–2024 restructuring. Beginning 2022 Stitch Fix faced a structural challenge: the broader retail apparel environment, the company’s own Direct Buy expansion (which ate into the curated-Fix value proposition), and post-pandemic demand normalisation produced revenue declines. The company laid off 20% of staff in January 2023 and another 14% in June 2023, including substantial cuts to the data-science team. Founder Katrina Lake stepped back from the CEO role in 2021 and returned briefly in 2024. By 2024 fiscal results, revenue had declined to ~USD 1.4 billion, and the company’s market capitalisation had fallen to under USD 400 million. The Stitch Fix story is not yet finished; the data-flywheel asset remains, but the operating-business challenges of late-stage 2024–2026 have substantially reshaped the company.

The structural lessons. Despite the company’s late-stage challenges, the Stitch Fix data-flywheel architecture remains the canonical contemporary example of building a durable AI-native business. Five lessons recur:

  1. Data collection is the product, not separate from it. Stitch Fix did not run separate “data collection” projects; every customer interaction produced data automatically. The architectural decision in 2012–2013 to build the Fix experience around structured data capture is what produced the cumulative-data moat.

  2. The Concierge MVP precedes the data infrastructure. Lake learned what data to collect by hand-picking Fixes herself for 12 months. Without that learning, the data infrastructure would have captured the wrong variables.

  3. Hybrid human-and-algorithm beats pure-algorithm at most realistic scales. The hybrid stylist-and-algorithm approach was not a stepping stone to full automation; it was the architecture. Customer-facing AI in 2026 increasingly looks like Stitch Fix’s hybrid pattern, not like the autonomous-AI-replaces-human pattern that early-2020s narratives suggested.

  4. Data-flywheel advantages compound over years, not months. Stitch Fix’s competitive moat from accumulated style preference data only became material after 4–6 years of data accumulation. Competitors that wanted to enter the personalised styling category in 2018 faced a data deficit measured in millions of Fixes.

  5. Data-flywheel advantages are operational, not absolute. The Stitch Fix case also shows that data flywheels do not protect against operating-business failures. A company can have an extraordinary data asset and still face revenue decline if the broader operating environment shifts unfavourably. The flywheel is necessary, not sufficient.

These lessons are why Chapter 23 (§23.3.7), Chapter 22 (§22.3.5), and Chapter 25 (§25.3.1) all reference Stitch Fix as the data-flywheel exemplar. The contemporary application — for student teams building AI MVPs in 2026 — is to instrument data collection from the first paying customer, to start with a Concierge MVP, and to plan the data-flywheel build as the structural advantage.

8.3 Amazon — the AI-native incumbent

Amazon’s relationship with AI has three distinct dimensions. As an e-commerce platform, Amazon’s recommendation, search, and personalisation infrastructure is the field’s longest-running and largest-scale deployment. As a cloud infrastructure provider, AWS provides the picks-and-shovels for much of the rest of the AI industry. As a foundation-model investor and operator, Amazon’s USD 8 billion total investment in Anthropic (announced in tranches through 2023–2024) and its in-house Bedrock and Titan model development position Amazon at the foundation-model layer of the stack.

The recommendation infrastructure. Amazon’s item-to-item collaborative filtering described in Linden et al. (2003) was a research achievement two decades ago; the contemporary system is much more sophisticated, incorporating session-aware models, real-time signal processing, multi-task learning, and personalisation across the company’s many product surfaces (the front page, search results, product pages, cart, email). The system is opaque from the outside, but internal accounts and academic papers from Amazon researchers (e.g., Smith and Linden, 2017) suggest a deep-learning-based architecture similar to YouTube’s two-tower pattern. The system processes hundreds of millions of recommendations per day across Amazon’s customer base.

The data-flywheel structure. Amazon’s data-flywheel position is among the strongest of any commercial firm. The company captures essentially every interaction across its e-commerce platform: searches, browsing, cart adds, purchases, returns, reviews, ratings. Each interaction is a labelled data point for some recommendation or personalisation task. The cumulative dataset has been growing for ~30 years. The result is a moat that competitors with shorter data histories cannot match by buying or hiring.

AWS as picks-and-shovels. The 2006 launch of AWS established the cloud-computing pattern that has shaped every subsequent technology business. By 2024, AWS revenue exceeded USD 100 billion annually, with operating margin substantially above the e-commerce business. The AI-specific cloud services — SageMaker for ML training and deployment, Bedrock for foundation-model access, Titan for in-house models, Rekognition for vision, Comprehend for NLP — capture much of the infrastructure spend that AI startups and enterprises generate. Amazon’s economic exposure to the AI wave is therefore double: it benefits from AI directly (in its own platform) and it benefits as the infrastructure provider for everyone else’s AI.

The Anthropic investment. Amazon’s investment in Anthropic, announced as USD 4 billion in September 2023 and increased to USD 8 billion total by November 2024, is the company’s largest single foundation-model bet. The investment is paired with a commercial agreement under which Anthropic uses AWS as its primary cloud provider and Amazon’s Trainium chips for training. The strategic logic combines (a) hedge against OpenAI/Microsoft and Google as foundation-model incumbents; (b) preferred-customer access to Anthropic’s frontier models for AWS customers; (c) protection of AWS’s enterprise-customer base against churn to Microsoft Azure or Google Cloud; (d) acceleration of the in-house Trainium chip roadmap.

Alexa and the conversational-commerce ambition. Amazon’s bet on Alexa as a conversational-commerce platform has been more mixed. Alexa devices sold at scale (over 100 million units by 2019), but Alexa-mediated commerce has remained a small fraction of Amazon’s overall sales. The 2022 Reuters report (Lerman, 2022) on Alexa’s USD 10 billion annual operating loss — and the subsequent restructuring, with thousands of Alexa-team layoffs in 2022–2023 — suggests that the original conversational-commerce thesis has not played out as the company hoped. The 2024 reset with the “new Alexa” announcement (with foundation-model integration, paid subscription tier) repositions the product but does not yet establish a different commercial trajectory.

Robotic warehousing and operations. Amazon’s 2012 acquisition of Kiva Systems for USD 775 million was an early-mover bet on robotic warehouse automation. By 2024 the company operated over 750,000 robots across its fulfilment centres. The 2024 introduction of “Sequoia” — a new generation of automation that handles the long tail of items previously requiring human picking — extends the trajectory. The economic case for robotic warehousing is well-established; Amazon’s 2024 financial reports cite warehousing-automation cost savings as a material contributor to fulfilment-cost reduction. The case generalises: similar automation is being deployed at Walmart (with the Symbotic partnership), at Carrefour, at Tesco, and at many other major retailers.

The Amazon case is structurally important because it shows what AI deployment at incumbent scale looks like. The combination of recommendation infrastructure (long history, deep data), cloud platform (broad reach), foundation-model investment (frontier capability), and operational AI (warehousing, forecasting) produces a position that is hard to disrupt at the corporate level. Specific competitor entry into specific Amazon-occupied surfaces is possible (Shopify in merchant tooling; TikTok Shop in social commerce); replacing Amazon as the integrated default is much harder.

8.4 Klarna — the AI customer-service post-mortem

The Klarna case is the field’s most-detailed contemporary cautionary tale of AI deployment, and the case the playbook chapters of this textbook reference repeatedly. The detail merits a deep treatment.

Klarna context. Klarna AB is a Swedish buy-now-pay-later (BNPL) firm founded in 2005, with major operations in Europe, the United Kingdom, the United States, and Australia. The company served approximately 150 million consumers and 600,000 merchants by 2024. After a substantial valuation peak in 2021 (USD 45.6 billion), Klarna faced a significant valuation correction in 2022 (down to USD 6.7 billion) reflecting the broader fintech reset. The company recovered partially through 2023–2024, returning to profitability and pursuing a public listing. Klarna listed on the NYSE in November 2024 at a USD 14 billion valuation. CEO Sebastian Siemiatkowski has been a prominent and outspoken figure throughout, with a public profile that contributes substantially to the case’s visibility.

The February 2024 announcement. On 27 February 2024 Klarna announced that its OpenAI-powered AI assistant had handled approximately 2.3 million customer-service conversations in its first month of full deployment. The announcement included specific metrics: the AI handled work equivalent to ~700 full-time customer-service agents; resolution times had fallen from 11 minutes to under 2 minutes; customer-satisfaction scores were “on par with human agents”; the company expected USD 40 million in profit improvement annually. The announcement was widely covered in business and technology press; Klarna’s Twitter account and Siemiatkowski personally amplified the message. Klarna positioned the deployment as evidence that generative AI was ready for customer-facing operations at scale.

The deployment design. Klarna’s deployment was unusual in scale and speed. The AI assistant was built on OpenAI’s GPT-4 (later GPT-4 Turbo) with retrieval over Klarna’s internal knowledge base, and was rolled out across Klarna’s customer-service infrastructure within weeks. The deployment did not use a phased rollout (pilot with a subset of customers; gradual scaling based on validated outcomes); the AI essentially replaced the front-line customer-service tier across markets simultaneously. The agent attrition that followed was substantial — Klarna had reduced its workforce by ~20% (~700 positions) over the prior year, with attribution to AI capability cited prominently. The decision to publicly announce both the workforce reduction and the AI deployment in close temporal proximity was a strategic communication choice; it produced the maximal visibility for the AI deployment but coupled the company’s reputation closely to the deployment’s success.

The 2024 metrics and what they did and did not measure. Klarna’s published metrics — resolution time, conversation volume, NPS — were genuine, internally measured, and not (as far as public reporting indicates) misrepresented. What the metrics did not capture was the quality of resolution: whether the AI’s responses were accurate, whether the customer’s underlying problem was solved, whether the customer would ultimately remain a Klarna customer. Resolution time can fall not because problems are resolved faster but because the AI closes conversations the customer has not had resolved; NPS measured at the moment of conversation closure does not capture the customer’s eventual satisfaction. The metrics design is itself the failure mode that Chapter 23’s evaluation discussion identifies — measuring the wrong thing efficiently produces misleading confidence.

The 2024–2025 reversal. Through late 2024 and into 2025, customer-experience metrics that had not been visible in the initial reporting began to surface. Independent customer-satisfaction trackers (e.g., Trustpilot for Klarna’s UK and US operations) showed declining ratings through 2024. Customer-complaint volumes via regulatory channels (UK Financial Ombudsman, US CFPB) increased. Internal Klarna data — referenced in subsequent press but not formally released — reportedly showed the resolution-quality gap that the 2024 metrics had not captured.

In May 2025 Siemiatkowski publicly acknowledged the reversal. In a Bloomberg interview (May 2025), he said: “From a brand perspective, the cost of using AI was that quality became lower. Investing in the quality of the human support is the way of the future for us.” The company began rehiring customer-service agents — initially through a “Klarna gig-worker” model where former employees and others could provide on-demand customer-service work, and subsequently through more conventional employment. By Q3 2025 Klarna’s customer-service workforce had recovered substantially. The company did not abandon AI in customer service entirely; rather, it positioned AI as augmentation for human agents (handling routine inquiries; surfacing relevant information; drafting responses for human review) rather than as autonomous resolution. The May 2025 acknowledgement was widely covered as the most-prominent AI-deployment reversal of 2025; it has since become a standard reference in customer-service-AI deployment literature.

The structural lessons from the Klarna case mirror Watson Health’s (Section 7.3) but in a contemporary context.

Lesson 1 — alpha-skipping has high tail risk. The Klarna deployment moved from internal testing directly to ~67% of customer-service traffic without a meaningful staged rollout. A 5–10% rollout to a defined cohort, with measurement against the existing-agent baseline, would have surfaced the resolution-quality problems before brand impact accumulated. The cost of the staged rollout would have been weeks of additional deployment time; the cost of skipping the staged rollout was the brand-trust unwind that the May 2025 reversal acknowledges. This lesson is what Chapter 24 (§24.3.5) develops as the alpha-discipline argument.

Lesson 2 — the wrong metrics produce confident-but-wrong conclusions. The metrics Klarna emphasised in February 2024 — resolution time, conversation volume — measured efficiency, not effectiveness. The right metrics for customer-service AI are downstream: customer-satisfaction trajectory over weeks; problem-recurrence rates; customer churn rates among AI-handled vs human-handled customers; customer-effort-score on multi-step issues. None of these were prominent in Klarna’s 2024 reporting. The lesson generalises: AI-deployment evaluation must measure the outcomes that matter to the business, not the proxies that are easy to measure. This is what Chapter 23’s evaluation discipline addresses.

Lesson 3 — public commitment ahead of validation creates reputational sunk costs. Klarna’s February 2024 announcement was high-profile, including direct comments from Siemiatkowski and prominent media coverage. The public commitment made the eventual reversal more reputationally costly; quietly walking back the deployment was no longer an option once the deployment had been positioned as a strategic win. The lesson: align public commitment with validated evidence. Pre-validation announcements convert weak evidence into hard-to-reverse positions.

Lesson 4 — substitution willingness ≠ augmentation willingness. Klarna’s pre-deployment customer research (referenced in subsequent reporting) reportedly indicated that customers were “comfortable using AI for service queries.” This was true but incomplete. Comfortable using AI alongside human agents is different from willing to remain a customer when AI is the only available channel. The framing matters; Klarna’s research framing did not distinguish the two. Chapter 20 (§20.3.5) develops this lesson as the customer-discovery framing principle.

Lesson 5 — the brand damage from premature deployment is durable. The reputational cost of the Klarna reversal extended beyond customer-service: the company’s broader AI initiatives faced increased scepticism from regulators, business press, and Wall Street. Recovering from that scepticism takes years; the brand cost is not paid only at the moment of reversal but persists. For an early-stage startup, this lesson is sharper: a reversal at scale could destroy the company’s customer relationships entirely. For Klarna, the company has the financial position and customer base to recover; for a startup, the same pattern would be terminal.

Klarna’s case is recent enough that the full implications are still working through. The May 2025 reversal is the most-detailed public-record AI-deployment reversal of the contemporary era. Subsequent customer-service-AI deployments (at Wendy’s, at Air Canada, at various retail chains) explicitly cite Klarna lessons in their deployment design. The case has become an industry reference; its incorporation into this textbook’s playbook chapters reflects the durability of the lessons rather than the specifics of any one company.

8.5 Shopee, Lazada, Tokopedia — the Southeast Asian platforms

Southeast Asian e-commerce is dominated by three platforms — Shopee, Lazada, and Tokopedia — whose competitive dynamics, technical architecture, and AI deployment patterns differ substantially from the Amazon-centric pattern that dominates the US literature. The dynamics matter for KL-based students directly: most Malaysian e-commerce, including the projects students may build, runs through these platforms or interacts with them.

Shopee (Sea Group). Sea Limited, founded 2009 in Singapore by Forrest Li, operates Garena (gaming), Shopee (e-commerce), and SeaMoney (digital financial services). Shopee launched in 2015 and has become Southeast Asia’s largest e-commerce platform by volume. The company listed on NYSE in 2017; Shopee’s gross merchandise volume reached USD 24.6 billion in Q3 2024, with revenue of USD 5.7 billion. Shopee’s strategic positioning emphasises mobile-first design, vernacular-language localisation across the seven main SEA markets (Singapore, Malaysia, Thailand, Philippines, Indonesia, Vietnam, Taiwan), and aggressive merchant-acquisition incentives that produced rapid scale. The platform’s AI use is comprehensive: search ranking, recommendations, fraud detection, seller-side automation (listing optimisation, image enhancement), customer-service automation (now more cautious post-Klarna), and the demand-forecasting infrastructure for warehousing and logistics. Shopee Live and Shopee Video — the live-commerce and video-shopping surfaces — incorporate computer-vision-based product tagging and recommendation. The platform’s competitive advantages include its parent Sea’s gaming-derived expertise in recommendation systems (the Garena gaming platform has been doing user-engagement modelling at scale for over a decade) and its substantial data position from the cumulative SEA e-commerce market.

Lazada (Alibaba). Lazada, founded 2012 by Rocket Internet, was acquired by Alibaba in 2016 for USD 1 billion (with subsequent additional investments raising the total to over USD 6 billion by 2024). Lazada operates across the same six SEA markets as Shopee (excluding Taiwan). The platform’s scale is below Shopee’s by GMV; analysts (e.g., Bain, Momentum Works) estimate Lazada’s GMV at roughly 30–40% of Shopee’s. Lazada’s AI deployment leverages Alibaba’s broader infrastructure: Alibaba’s Damo Academy provides foundation-model and recommendation-system technology; Alibaba Cloud provides the underlying compute. The platform has been a faster adopter of generative-AI features than Shopee in some respects (Alibaba’s broader generative-AI push under the Tongyi Qianwen / Qwen family produces a strong technical pipeline) but a slower competitor in market-acquisition over the 2022–2024 period. Lazada’s restructuring in 2024 — including significant headcount reductions in operational regions — suggests the company is in a defensive consolidation phase.

Tokopedia + TikTok Shop (GoTo Group + ByteDance). Tokopedia, founded 2009 in Indonesia by William Tanuwijaya, was Indonesia’s largest e-commerce platform by GMV through the 2010s. The 2021 merger with Gojek to form GoTo Group consolidated Indonesian super-app dynamics. The 2023 entry of TikTok Shop into Indonesia produced a regulatory and competitive crisis (the Indonesian government banned TikTok Shop’s e-commerce operations in October 2023, citing predatory pricing concerns); the resolution in early 2024 was a partnership-and-investment structure under which TikTok Shop was integrated with Tokopedia under a TikTok-controlled majority stake. The combined entity competes with Shopee and Lazada as the primary Indonesian e-commerce channel. The technical architecture combines Tokopedia’s mature e-commerce infrastructure with TikTok’s recommendation system — the latter being the structural advantage that has driven TikTok Shop’s rapid growth in Southeast Asia. TikTok’s For You algorithm, when applied to commerce, produces a discovery-driven shopping pattern that the search-and-category-driven competitors find difficult to match.

The platform competitive dynamics. Three observations about the platform dynamics matter for understanding the SEA AI deployment landscape.

First, the take-rate dynamics. Platform commission on each sale (the take rate) is the primary revenue mechanism. Take rates in SEA (typically 1–8% depending on category and seller status) are substantially below those of Amazon (typically 6–20%) or Apple’s App Store (15–30%). The compressed take rates limit how much the platforms can invest in AI infrastructure relative to the GMV they handle, which in turn shapes the AI deployment patterns. Operational efficiency AI (search, recommendations, fraud) gets heavier investment than customer-experience AI (conversational, agentic) where the ROI is harder to demonstrate.

Second, the subsidies-and-acquisition pattern. SEA e-commerce platforms have historically subsidised growth heavily — free shipping, voucher discounts, seller-onboarding incentives. The subsidies are funded by venture capital and platform parent capital, not by current operating margin. The 2022–2024 reset in venture funding has compressed subsidy budgets; platforms have shifted toward profitable growth. The shift affects AI deployment: platforms that previously could afford to deploy AI improvements at marginal cost now must justify deployment against ROI metrics, which slows the pace.

Third, the seller-side AI race. The platforms compete to make their seller-side experience more attractive than competitors’. AI-powered seller tools — automatic listing translation, image enhancement, copywriting assistance, pricing suggestions, demand forecasting for inventory — have become a major battleground. Shopee’s 2023 launch of Shopee SCM (supply chain management) AI tools and Lazada’s parallel investments are responses to this dynamic. The race is largely invisible to consumers but is consequential for SEA SMB seller economics.

The 2024–2026 generative-AI wave at platform level. All three major platforms have launched generative-AI features for both buyer and seller surfaces in 2024–2026. Buyer features include AI shopping assistants (Lazada’s “AI buddy”; Shopee’s AI search), product-description summarisation, image-based search, and personalised landing pages. Seller features include automated listing generation from a single image and product description, AI-powered customer-service drafting for seller-buyer chats, and demand-forecasting tools. The deployment is broad but cautious; the platforms have studied the Klarna case and have been more conservative than the early-2024 hype suggested they might be.

The Carsome case. Carsome is Malaysia’s most-prominent AI-adjacent e-commerce success. Founded 2015 by Eric Cheng and Jiun Ee Teoh, Carsome is a used-car marketplace operating across Malaysia, Indonesia, Thailand, and Singapore. The company’s scale grew rapidly: it became Malaysia’s first unicorn in 2021 (valued at USD 1.7 billion) and reached approximately USD 1 billion in revenue by 2023. Carsome’s AI deployment focuses on three areas: vehicle inspection automation (computer-vision-based damage assessment and condition rating), pricing dynamics (regression-based pricing recommendations for sellers and buyers), and dealer-network optimisation (matching cars to dealer demand). The case is referenced in Chapter 19 (§19.3.7) and Chapter 26 (§26.3.5) for its bottom-up market-sizing methodology and regional-knowledge moat. Carsome’s continued scaling through 2024 — despite the broader fintech-and-e-commerce funding compression — indicates that operational AI in regional markets can produce durable advantages even without frontier-model technology.

8.6 Generative AI in retail operations

The 2022–2026 generative-AI wave has produced a broad-but-shallow deployment pattern across retail operations. Most major retailers and e-commerce sellers now use generative AI in some form; few have used it for the strategic transformation that the early-2023 hype suggested. The deployment pattern is concentrated in five specific operational areas.

Product description and listing generation. Shopify’s Magic, introduced in 2023, uses GPT-4 and similar models to generate product descriptions from a few input attributes. Amazon Seller Central added equivalent functionality via its Bedrock-backed service in late 2023. eBay’s “Magical Listing” feature, launched 2024, allows sellers to photograph an item and receive a complete listing (title, description, attributes, suggested category, suggested price) in seconds. The deployment scale is large — Shopify reported in early 2024 that Magic had been used to generate millions of product descriptions — but the per-merchant value is modest; the technology saves time on a task that was previously low-value labour.

Ad creative generation. Meta’s Advantage+ Creative (2023), Google’s Performance Max with generative-AI extensions (2024), and Amazon DSP’s AI-generated ad creative (2024) all use generative AI to produce ad variations from product feeds. The deployment is operational rather than novel: ad creative testing was previously done with human-designed variants; AI generates more variants faster. The ROI claims are moderate (Meta reports 5–15% ad-performance improvements from Advantage+ Creative versus manual creative); the deployment is universal across major advertisers.

Customer service. The Klarna case (Section 8.4) is the cautionary reference; deployment has continued but with explicit awareness of the lessons. Sephora’s AI customer service (deployed 2024 in cautious phases), Best Buy’s Microsoft Copilot integration (2024), and various other retail-AI customer-service deployments reflect a more-conservative approach than Klarna’s. The pattern is now augmentation-first: AI suggests responses, human agents review and send; full autonomous resolution is reserved for the simplest queries (order status, delivery tracking) where the failure cost is bounded.

Personalisation at scale. The personalisation capability that classical recommendation systems provided is being extended by generative AI to produce per-user personalised content (email subject lines, product descriptions, landing-page copy). Retention.com, Klaviyo, and various marketing-automation platforms have integrated foundation-model capabilities into their personalisation engines. The deployment is mostly invisible to consumers (the email they receive is personalised; they cannot easily compare to the un-personalised version) but represents substantial commercial value at scale.

Search and discovery. Foundation-model-based search — semantic understanding of queries, generating natural-language responses, conversational refinement — is being deployed by major platforms. Amazon’s Rufus shopping assistant (launched 2024 in beta, expanded 2025) is the most-visible example. Google’s AI-powered shopping search (the Search Generative Experience extended into commerce) competes from a different starting point. The strategic question is whether the conversational-search interface displaces the traditional search-and-browse pattern or augments it; as of 2026 the answer is “augments mostly, displaces in narrow contexts.”

The 80% deployment pattern. A useful framing: by mid-2025, approximately 80% of major e-commerce sellers report using generative AI in some form (per IBM and Salesforce surveys); the remaining 20% are typically smaller operators with limited tooling budgets. The 80% number is high but the depth varies: most use generative AI for one or two operational improvements; few use it as the structural foundation of their business. The pattern resembles other technology adoption waves (mobile, cloud) — broad early adoption for incremental gains, with strategic transformation taking longer to materialise.

8.8 Conversational and agentic commerce

Conversational and agentic commerce — where AI takes the buyer’s or seller’s role in mediated interactions — is the field’s 2024–2026 frontier. The technology builds on the foundation-model wave; the deployment is uneven; the structural questions about the long-run pattern are unresolved.

Conversational shopping assistants on the buyer side. Amazon Rufus (launched 2024 beta, expanded 2025) allows customers to ask questions in natural language about products, compare options, and refine searches conversationally. Walmart Sparky (launched 2024) and Shopify’s Shop AI (2024) provide similar functionality. The deployment is broad but the depth is uneven; most users still default to traditional search-and-browse, with conversational interactions concentrated in specific contexts (gift recommendations; complex multi-attribute searches; comparing products with technical specifications).

The 2025 agent wave. OpenAI’s Operator (launched January 2025), Anthropic’s Computer Use (launched October 2024 in beta, expanded 2025), Google’s Gemini-based agentic browsing (Project Mariner, 2024), and various startup-stage agents (Browserless, Adept, several others) collectively represent the agentic-commerce frontier. The core capability: a foundation-model-driven agent that operates a web browser on the user’s behalf, navigating pages, filling forms, completing transactions. The early deployment showed substantial capability (the agents can complete complex multi-step purchases) and substantial limitations (the agents are slow; they make mistakes; they require trust the user has not yet extended).

The structural challenges of agentic commerce. Five problems define the field’s deployment trajectory.

Trust. Allowing an AI to spend the user’s money requires substantial trust. Users have not generally extended this trust to agents. The 2025 deployment is mostly experimental; commercial agentic commerce at scale (where agents handle the bulk of transactions for non-trivial categories) is more like a 2027–2030 trajectory than a 2025 reality.

Fraud. If the agent can spend the user’s money, the agent is also a target for adversarial attacks (prompt injection from malicious websites; spoofing; misleading product information). The fraud surface for agents is potentially larger than for human shoppers; the mitigation infrastructure is immature.

Payment and authorization. The current payment infrastructure (Visa, Mastercard, the bank-mediated layers) was designed for human-authorised transactions. Agent-mediated transactions require new authorization patterns (delegated tokens; spend limits; reversibility). The major card networks announced agent-payment-token frameworks in 2024–2025 but the deployment is early.

Returns and disputes. When an AI agent purchases something the user didn’t want, who is responsible? The legal framework is undefined. The 2025 cases (a few prominent disputes have surfaced in consumer-protection regulatory channels) are being handled ad hoc; durable resolution frameworks are still emerging.

Platform vs neutral agent. The strategic question of whether agents will be platform-aligned (Amazon’s agent buys from Amazon; Walmart’s agent buys from Walmart) or neutral (the user’s preferred agent shops across retailers, competing on price and quality) shapes the long-run market structure. Platform alignment compresses the consumer benefit; neutral agency compresses the platform’s pricing power. As of 2026 both patterns are observed; the long-run equilibrium is unsettled.

The 2026 state. Agentic commerce is operational for narrow use cases (price comparison, restocking commodity items, simple replenishment) at small scale, and experimental for broader uses. The technical capability is largely sufficient for many use cases the deployment has not yet reached; the binding constraints are trust, infrastructure, and regulatory framework. The trajectory is upward — each year sees more deployment, broader use cases, better capability — but the pace is slower than the 2024 announcements suggested.

8.9 Demand forecasting and inventory

Demand forecasting and inventory management are AI applications with much longer deployment histories than the recent generative-AI wave suggests. Walmart’s analytics infrastructure has been operating since the 1990s; Procter & Gamble’s demand-planning systems for decades. The contemporary pattern combines classical statistical forecasting (ARIMA, exponential smoothing, hierarchical forecasting) with machine-learning approaches (gradient boosting, deep learning, increasingly transformer-based time-series models).

Walmart’s forecasting infrastructure. Walmart operates one of the world’s largest retail forecasting infrastructures, with hundreds of thousands of items forecast at thousands of stores at daily granularity. The infrastructure underlies Walmart’s competitive position in inventory turn and stock availability. The 2010s evolution was from rules-based forecasting (heuristics with manual planner overrides) to ML-based forecasting (gradient boosting on a wide feature set including sales history, weather, local events, promotions, competitor pricing). The 2020s evolution incorporates deep-learning approaches and increasingly graph-neural-network representations of the supply-chain structure.

Zara and fast-fashion data integration. Zara’s parent Inditex operates a fast-fashion supply chain that integrates point-of-sale data with manufacturing and distribution decisions in near-real-time. The integration allows Zara to design, manufacture, and ship new product to stores in 2–3 weeks (versus competitors’ 6+ months). The data infrastructure underlying the integration is the source of the competitive advantage; AI applications in pricing, allocation, and trend identification compound on the data foundation.

Amazon inventory placement. Amazon’s fulfilment-network design requires forecasting demand at fine geographic granularity (where will demand for product X be in 30 days?) and placing inventory accordingly. The placement decisions involve trade-offs between transportation cost, holding cost, and stock-out risk. Amazon’s optimisation infrastructure for these decisions is among the most sophisticated in retail; the inventory turn improvements that the optimisation enables are a substantial contributor to Amazon’s operational efficiency.

The bullwhip effect and ML mitigation. The bullwhip effect (Lee, Padmanabhan, and Whang, 1997) describes how demand-signal distortions amplify upstream in supply chains, producing volatility that exceeds underlying demand volatility. Classical mitigations (information sharing across the supply chain; smaller batch sizes; reduced lead times) have been augmented by ML-based demand-sensing approaches that detect signal earlier and cleaner. Companies with end-to-end visibility (Amazon, Walmart, the SEA platforms) have an inherent advantage in bullwhip mitigation; companies with fragmented visibility (most independent retailers) struggle more.

Cold-chain and perishables. Forecasting and inventory management for perishable goods (groceries, pharmaceuticals, fresh-flower products) is structurally harder than for non-perishables. The combined constraints — perishability, demand volatility, and inventory shrinkage — produce a multi-objective optimisation problem that ML approaches address well. Specific deployments at Walmart Grocery, Tesco, and the European retail chains (Carrefour, Sainsbury’s) have produced measurable shrinkage reductions and stock-out improvements. The Australian context — Coles and Woolworths — has had similar deployments, with the additional constraint of geographic remoteness for some store networks.

8.10 Fraud, abuse, and trust-and-safety

Fraud detection in retail and e-commerce is among the most-mature AI deployment domains. The fraud landscape includes payment fraud (stolen credit cards used for purchases), account takeover (compromised customer accounts used by third parties), refund and return abuse (fraudulent return claims), counterfeit listing fraud (sellers listing fake products), review fraud (paid or fake reviews), and various forms of platform abuse.

Stripe Radar. Stripe Radar, the fraud-detection engine within Stripe’s payment platform, is among the most-deployed retail-fraud-detection systems globally. Radar uses a combination of classical machine learning (gradient boosting on transaction features) and deep learning (sequence models over user behaviour). The system processes Stripe’s full transaction volume — over USD 1 trillion annually as of 2024 — and produces real-time risk scores. The data flywheel here is substantial: every transaction, every chargeback, every dispute is a training signal that improves the system’s discrimination over time.

Sift, Forter, and the fraud-platform vendors. A specialised fraud-detection-platform industry has matured around the recognition that most retailers cannot build in-house fraud detection competitive with specialised vendors. Sift (founded 2011), Forter (founded 2013), Riskified (founded 2013), and Signifyd (founded 2011) provide fraud-detection-as-a-service to retailers. The competitive positioning differs: some vendors emphasise the breadth of the cross-merchant data network they have access to; some emphasise the ML model architecture; some emphasise specific verticals or use cases. The market has consolidated somewhat through 2022–2024 (acquisitions, consolidation of smaller vendors); four major platforms account for the bulk of mid-and-large-merchant fraud-detection spend.

Marketplace abuse — counterfeits, listing fraud, review fraud. Marketplace platforms (Amazon, eBay, the SEA platforms) face structural abuse problems that pure-payment fraud detection does not address. Counterfeit listings, fake reviews, listing manipulation (re-listing under different SKUs to avoid history), and seller fraud all require domain-specific detection systems. Amazon’s Project Zero (2019) and Brand Protection programmes use computer-vision-based and text-based detection to identify counterfeit listings; the deployment scale is large but the cat-and-mouse dynamic with bad actors is ongoing. The 2023–2024 wave of generative-AI-produced fake reviews has produced a new detection challenge: when fake reviews are written by GPT-4-class models, the linguistic-analysis approaches that previously detected fake reviews are weaker; new approaches use behavioural-pattern analysis (does the reviewing-user account look genuine across non-text features?) to compensate.

Chargeback management. Chargebacks — disputes filed by customers via card networks — produce both direct cost (the disputed amount, plus chargeback fees) and indirect cost (excessive chargebacks can lead to merchant-status downgrades or terminations). Chargeback-management platforms (Chargebacks911, Verifi, the broader Visa/Mastercard programmes) use ML-driven dispute-resolution workflows to reduce chargebacks and to win the disputes that reach the network.

The arms-race dynamic. Fraud detection operates in an adversarial environment; bad actors adapt their techniques to evade detection. Foundation models’ availability has accelerated the adversarial side: AI-generated fake reviews, AI-generated counterfeit listings, AI-generated phishing materials, AI-driven account-takeover attacks. The fraud-detection platforms are responding with their own AI tooling, but the asymmetry between offence (low-cost; immediate experimentation; few legal-or-regulatory constraints) and defence (must minimise false-positive rate; legal liability) creates a structural advantage for the offence side that the defence side must compensate for through scale and data.

8.11 Operations — warehousing, last-mile, returns

Retail operations — particularly warehousing, last-mile delivery, and returns processing — are large-scale labour-intensive activities where AI and automation produce substantial cost savings.

Robotic warehousing. Amazon’s Kiva acquisition (2012) and subsequent robotic-warehouse deployment is the field’s reference case (Section 8.3). The pattern has spread broadly: Walmart’s Symbotic partnership (acquired full ownership 2023), Carrefour’s automation programme, Tesco’s Ocado-derived warehousing, the Asian e-commerce platforms’ direct or partner-led automation. Symbotic (publicly-listed in 2022) provides the warehouse-automation platform to Walmart and other large customers; the company’s revenue grew from USD 350 million in fiscal 2022 to over USD 1.5 billion in fiscal 2024. AutoStore (Norwegian, IPO 2021) provides cube-based storage automation. Locus Robotics (private, ~USD 800 million revenue 2024) provides collaborative robots for human-augmented warehousing. The combined market for warehouse-automation hardware and software exceeded USD 30 billion in 2024.

Last-mile delivery automation. The last-mile delivery problem — the cost-intensive final step from local hub to customer doorstep — has received substantial AI investment with mixed deployment results. Cruise, the General Motors-backed autonomous-vehicle subsidiary, exited the autonomous-taxi market in late 2024 after multi-year losses; the autonomous-vehicle path to last-mile delivery has been slower than 2017-era projections suggested. Nuro (founded 2016) operates an autonomous local-delivery vehicle programme; the deployment scale is modest but operational at specific markets. Starship Technologies (founded 2014) operates small sidewalk-delivery robots in college campuses and specific urban contexts. Drone delivery (Wing, Zipline, Amazon Prime Air) has reached limited commercial scale. The 2026 picture: last-mile autonomy is operational at narrow scales, broadly deployed at the last 100m (sidewalk robots in campuses; delivery lockers; in-store pickup), but the autonomous-vehicle-to-doorstep general case remains unsolved.

Returns management. The returns problem — particularly for apparel, where return rates of 20–40% are typical — produces substantial reverse-logistics cost. AI applications include returns-prediction (estimating the probability that an order will be returned, used to optimise inventory placement and refund processing), returns-routing (deciding whether to ship the returned item to the original warehouse, a returns-processing centre, or directly to a different customer), and return-fraud detection. Returnly (acquired by Affirm 2021) and Loop (founded 2017) provide returns-management platforms; the integration with retailers’ broader operational systems is the deployment pattern.

The reverse-logistics problem. Returns generate environmental cost (transportation emissions; packaging waste; product disposal) that has become a regulatory focus in some jurisdictions. The EU’s Right to Repair regulations (2024) and various national waste-reduction programmes overlay returns management with environmental-compliance considerations. The combined cost — operational, environmental, regulatory — has produced a return-reduction-via-better-shopping push, in which AI-driven sizing recommendations, AR try-on, and improved product information are positioned as primary preventive interventions rather than reactive returns-processing improvements.

8.12 Australian and Malaysian retail context

The Australian and Malaysian retail-and-e-commerce contexts differ from the US and Chinese patterns in scale, structure, and the AI deployment pace.

Australian retail. The Australian retail market is dominated by a small number of large players: Wesfarmers (parent of Bunnings, Officeworks, Kmart, Target), Woolworths Group, Coles Group, JB Hi-Fi, Harvey Norman, Myer, David Jones. The structure is more concentrated than US retail; the large players have substantial AI deployment capacity. Coles and Woolworths (the supermarket duopoly, ~65% of grocery market) have invested heavily in supply-chain AI, loyalty programme personalisation, and store-operations optimisation. The AI deployment is operational rather than customer-facing; both companies’ app and website experiences remain fairly traditional.

The Iconic, Catch, Kogan. Pure-play Australian e-commerce includes The Iconic (apparel; founded 2011; profitable scale by 2018), Catch (founded 2006; acquired by Wesfarmers 2019; struggling profitability), and Kogan (founded 2006; ASX-listed; mid-scale). All three use ML-based recommendation, search, and personalisation, with deployment depth comparable to US mid-scale retailers. The Iconic has been a particular leader in AI-driven sizing recommendations for apparel.

Cotton On Group. Cotton On (founded 1991, Geelong-based, expanded internationally) operates approximately 1,500 stores globally with the Cotton On, Cotton On Kids, Cotton On Body, and Typo brands. The company’s AI deployment includes inventory optimisation, store-allocation forecasting, and customer-segmentation. The scale of operations and the international footprint make Cotton On one of the larger Australian retail-AI deployments.

Afterpay and BNPL. Afterpay (founded 2014, Block acquisition 2022 for USD 29 billion) is the Australian-originated BNPL leader. The acquisition by Block (formerly Square) integrated Afterpay with the broader Square commerce stack, including Square Online and the Cash App. Afterpay’s AI use is concentrated in credit-decisioning (assessing whether to extend BNPL credit at point of sale) and fraud detection. The 2022–2024 BNPL reset compressed the category’s growth; Afterpay’s contribution to Block’s broader business has been more modest than the acquisition price implied.

Canva’s commerce extensions. Canva (Section 19.3.8 reference) is Australia’s largest single technology company by valuation. Canva’s commerce extensions — print products, branded merchandise, collaborative design for shoppable assets — represent a small fraction of overall revenue but substantial growth investment. The company’s AI deployment in design tools (the Magic Studio suite, launched 2023) is among the most-extensive consumer AI deployments by an Australian-domiciled firm.

Malaysian retail. Malaysian retail has a similar structure to other middle-income SEA markets: the major modern retail (Tesco/Lotus, AEON, Mydin, KK Mart, 7-Eleven Malaysia) operates alongside extensive traditional retail (kedai runcit, pasar). E-commerce penetration has grown rapidly over the 2018–2024 period, accelerated by COVID, but remains lower than Singapore, Australia, or US markets. Local e-commerce platforms — historically PG Mall, 11Street Malaysia (now restructured), Mudah.my (Schibsted-owned classifieds) — have largely been displaced by Shopee and Lazada at the dominant-platform level.

Carsome. Already covered in Section 8.5; the Malaysian unicorn used-car marketplace.

iPrice. iPrice Group (founded 2014) is a regional price-comparison platform operating across SEA. The platform aggregates listings from Lazada, Shopee, and other retailers; the company’s AI use is concentrated in product matching across listings (identifying that two listings are the same product) and price-tracking. iPrice’s commercial position has been pressured by direct platform integration: as Shopee and Lazada have improved their own search and comparison features, the third-party meta-search use case has compressed.

Boost and the e-wallet ecosystem. Boost (Axiata-backed) and Touch ‘n Go eWallet (CIMB-backed) are the major Malaysian e-wallet platforms. While not strictly retail, the platforms’ integration with retail (in-store payments, online checkout, BNPL features) makes them retail-adjacent. AI deployment includes fraud detection, transaction-risk scoring, and merchant-side analytics.

AirAsia’s super-app pivot. AirAsia, originally a low-cost airline, has pursued a super-app strategy from 2020 onward (rebranded as Capital A in 2022), with the airasia.com platform extending into food delivery, e-hailing, payments, and various retail-adjacent services. The super-app strategy has produced mixed financial results but is one of the more-ambitious Malaysian-originated technology platform efforts. AI use is operational across the platform; the strategic distinctiveness comes from the cross-service data integration that the airline-and-travel core enables.

Hermo and the Malaysian beauty e-commerce category. Hermo (founded 2012) is Malaysia’s largest dedicated beauty e-commerce platform. The category is structurally amenable to AI personalisation — beauty products have complex attribute spaces (skin type, undertone, finish, ingredients) that map well to ML-driven recommendation. Hermo’s AI deployment is at mid-scale; the company has been profitable but has not pursued the venture-funded growth trajectory of similar regional firms.

8.13 The 2026 frontier and the path ahead

The retail-and-e-commerce AI landscape in 2026 sits between the broad-but-shallow generative-AI deployment of 2023–2024 and the agentic-commerce frontier that is emerging but not yet at scale. Five trajectories define the path ahead.

Trajectory 1 — agentic commerce maturation. As discussed in Section 8.8, agent-mediated transactions are operational at narrow scale. The 2026–2030 question is whether the trajectory steepens (broad agentic commerce by 2030) or flattens (narrow agentic commerce, with traditional shopping patterns persisting). The infrastructure investments — payment-token frameworks; agent-identity standards; trust mechanisms — are happening; the commercial deployment will follow.

Trajectory 2 — personalisation at scale. The combination of foundation-model capabilities and accumulated data flywheels is producing personalisation depth that prior recommendation systems could not achieve. Per-user personalisation of product descriptions, search results, marketing emails, and landing pages is becoming routine. The strategic question is whether this depth produces meaningful welfare improvements (better matching of customers to products) or whether it primarily extracts value through more-effective price discrimination and targeted persuasion.

Trajectory 3 — the dark-pattern concern. Personalisation depth combined with foundation-model capability for persuasive content production creates what may become a dominant retail-ethics concern through the late 2020s. Per-user persuasive content — emails, ad creative, on-site copy — that is individually optimised to maximise conversion has substantial commercial value but raises concerns about consent, manipulation, and consumer-welfare impact. The regulatory response, particularly in the EU under the Digital Services Act and the AI Act, will shape how the trajectory unfolds. The 2024 EU AI Act explicitly identifies “subliminal techniques beyond a person’s consciousness” as prohibited; the boundary between legitimate personalisation and prohibited subliminal manipulation is unsettled and will be litigated through 2026–2030.

Trajectory 4 — platform vs merchant power dynamics. The major platforms (Amazon, Shopee, Lazada, TikTok Shop) are extending AI capabilities that benefit their own ecosystems, often at the expense of merchant autonomy. Amazon’s algorithms decide which sellers get visibility; the algorithms’ design choices shape merchant economics. As AI capabilities become more central to platform operations, the platforms’ power over merchants grows. The countervailing force is independent infrastructure (Shopify, BigCommerce) that allows merchants to operate outside platform control; the platforms’ competitive responses to this independent layer will shape the late-2020s industry structure.

Trajectory 5 — the foundation-model / data-flywheel interaction. The 2026 question that has the longest-term implications is how foundation-model capability interacts with platform-accumulated data flywheels. The platforms have data; the foundation-model providers have capability; the integration produces personalisation and operational efficiency that neither party could achieve alone. The question is whether the integration is mutually-beneficial (both parties capture value) or whether one side dominates over time. Amazon’s Anthropic investment, Google’s vertical integration with Shopping, Microsoft’s OpenAI relationship and Bing Shopping integration are all bets on different versions of this question. The answer will substantially shape both the foundation-model industry’s economics and the e-commerce platform industry’s structure through 2030.

The retail-and-e-commerce landscape has been the longest-standing AI deployment domain. The 2024–2026 period represents a meaningful inflection — generative AI extending the operational deployment, agentic commerce opening new modalities, the regulatory framework catching up to the practice. The structural lessons from Stitch Fix, Klarna, Amazon, and the SEA platforms provide the most-detailed contemporary case material for the playbook discipline that this textbook’s Part V develops. The integration of those lessons with the analytical frameworks of Parts I–IV is what produces graduate-level competence in the field.

References for this chapter

Foundational recommendation systems

  • Linden, G., Smith, B., and York, J. (2003). Amazon.com recommendations: Item-to-item collaborative filtering. IEEE Internet Computing 7(1): 76–80.
  • Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer 42(8): 30–37.
  • Covington, P., Adams, J., and Sargin, E. (2016). Deep neural networks for YouTube recommendations. RecSys ’16.
  • Smith, B. and Linden, G. (2017). Two decades of recommender systems at Amazon.com. IEEE Internet Computing 21(3): 12–18.

Stitch Fix

  • Lake, K. (2018). Stitch Fix’s CEO on selling personal style to the mass market. Harvard Business Review 96(3): 35–40.
  • Iansiti, M. and Lakhani, K. R. (2020). Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World. Harvard Business Review Press.
  • Stitch Fix Inc. (2017–2024). Annual reports (Forms 10-K filed with the SEC).
  • Klingenberg, B. (2018–2023). Stitch Fix Algorithms team blog and conference talks.

Amazon and AWS

  • Lerman, R. (2022). Amazon’s Alexa losing $10 billion a year as it loses ground to Microsoft, Google. Reuters, 21 November 2022.
  • Amazon.com Inc. (2024). Annual report (Form 10-K).
  • Amazon.com Inc. (2023, 2024). Anthropic investment announcements (multiple, totalling USD 8 billion).

Klarna

  • Klarna AB (2024). Klarna AI assistant handles two-thirds of customer service chats in its first month. Press release, 27 February 2024.
  • Klarna AB (2024). NYSE IPO prospectus, November 2024.
  • Siemiatkowski, S. (2025). Bloomberg interview, May 2025; reversal of full-AI customer-service strategy.
  • Bloomberg News (2025). Klarna chief regrets going too AI heavy. Coverage of Siemiatkowski statements, May 2025.
  • Financial Times (2025). Multiple reports on the Klarna AI customer-service reversal, mid-2025.

Southeast Asian platforms

  • Sea Limited (2024). Annual report and Q3 2024 earnings release.
  • Alibaba Group (2024). Annual report; Lazada operating disclosures.
  • GoTo Group (2024). Annual report and TikTok Shop integration disclosures.
  • Bain & Company (2024). Southeast Asia e-Commerce report.
  • Momentum Works (2024). Ecommerce in Southeast Asia 2024.
  • Carsome Group (2023, 2024). Annual reports and investor materials.

Generative AI in retail operations

  • Shopify (2023, 2024). Magic feature announcements and product documentation.
  • Meta Advantage+ (2023, 2024). Documentation and case studies.
  • IBM Institute for Business Value (2024). Generative AI in retail.
  • Salesforce (2024). State of Commerce Report.

Visual search and CV

  • Radford, A., Kim, J. W., Hallacy, C., et al. (2021). Learning transferable visual models from natural language supervision (CLIP). ICML.
  • Pinterest Inc. (2024). Annual report and product disclosures.
  • Google (2017–2024). Google Lens product documentation and feature announcements.

Agentic commerce

  • OpenAI (2025). Operator launch announcement, January 2025.
  • Anthropic (2024). Computer Use beta announcement, October 2024.
  • Visa (2024). AI-Ready Commerce framework announcement.
  • Mastercard (2024). Agentic-commerce token framework.

Demand forecasting and supply chain

  • Lee, H. L., Padmanabhan, V., and Whang, S. (1997). Information distortion in a supply chain: The bullwhip effect. Management Science 43(4): 546–558.
  • Fisher, M. and Vaidyanathan, R. (2014). A demand estimation procedure for retail assortment optimization. Management Science 60(10): 2401–2415.

Fraud and trust-and-safety

  • Stripe (2024). Stripe Radar product documentation.
  • Visa (2024). Visa Risk Manager and AI fraud platform. Documentation.
  • US Federal Trade Commission (2024). Fake reviews and the AI implications. Policy discussion paper.

Operations

  • Amazon Robotics (2024). Sequoia automation announcement.
  • Symbotic Inc. (2022, 2024). S-1 and annual reports.
  • AutoStore Holdings (2021–2024). Annual reports.

Australia and Malaysia

  • Inside Retail Australia (2023, 2024). Australian retail AI coverage.
  • Wesfarmers Limited (2024). Annual report.
  • Coles Group, Woolworths Group (2024). Annual reports.
  • Block Inc. (2022, 2024). Afterpay integration disclosures.
  • Canva Pty Ltd (2023, 2024). Public communications and Magic Studio launches.
  • Capital A Berhad (2024). Annual report and super-app strategy disclosures.

Regulation

  • European Union (2022). Digital Services Act.
  • European Union (2024). Artificial Intelligence Act.
  • US Federal Trade Commission (2024). Endorsement guides update — AI-generated reviews.