Overview of AI usecases in e-commerce

By 2026, AI is no longer a separate topic in e-commerce — it permeates every operational area, from tracking to CRM, from the product catalogue to customer acquisition. Email service providers (ESPs) incorporate native predictive analytics, helpdesks offer agent-assisted support, product information management (PIM) systems automatically enrich data, and advertising networks optimise creative content using algorithms. Retailers no longer need to ask themselves whether they should adopt these technologies. The real question is where to start, and in what order.

This guide maps out four categories of use to answer this question: Analytics, Customer Relations, Data Enrichment and Marketing. Each corresponds to a distinct business objective, is aimed at different types of retailers and offers returns on investment (ROI) with varying timeframes — from a ‘quick win’ within one month to an infrastructure project that pays off in nine months.

For each category, we detail three practical use cases: what they’re for; who they’re relevant to; what works, what doesn’t, and what needs to be sorted out before getting started; and what categories of solutions exist — without comparing tools, because that’s neither the aim nor the value you’d expect from an architectural guide.

One overarching principle before we go into detail: AI is not a product; it’s a layer. It connects to the retailer’s existing data. The quality and relevance of tracking, the product catalogue and the CRM determine 80 per cent of the final ROI. A retailer who starts with AI without having cleaned up their foundations ends up with

Family 1 — Analysis

Business objective: to understand what is happening on the website and within the business, to detect issues more quickly, to diagnose them more effectively, and to make decisions based on data rather than intuition.

All retailers fall into this category, but there are two entry points. The €2–10 million retailer needs to automate a function they do not have in-house (no data analyst). Retailers with a turnover of €10–50 million already have reporting systems, but remain stuck at a descriptive level: they know what is happening, but not why. The three use cases below address these two profiles from different angles.

Use Case 1 — Business anomaly detection

The AI continuously monitors a range of critical metrics (CVR by device × channel × category, AOV, successful payment rate, front-end errors, ATC/session) and issues an alert as soon as a metric falls outside its expected statistical range. It distinguishes between a genuine deviation and normal variations linked to seasonality, product mix or the marketing calendar.

Who it’s for: all retailers with a turnover of €2 million or more. This is the AI application with the most immediate and internally justifiable ROI, as it is measured in terms of revenue recovered from incidents that would otherwise have gone unnoticed.

Why it’s critical: a JS bug on iOS 17 that halves the mobile conversion rate for three weeks across 8 per cent of traffic is costly. AI detects this type of anomaly 48 to 72 hours earlier on average. The higher the daily turnover, the easier it is to achieve a return on investment.

Best practice: focus on 10–20 truly critical metrics; don’t monitor everything (alert fatigue sets in and renders the system unusable within four weeks). Always pair the alert with an LLM that summarises the context and highlights the dimensions that are changing (device, channel, category) — otherwise, you’re alerting on the symptom without helping with the diagnosis. Model seasonality explicitly, otherwise thresholds drift during peak season.

Pitfalls: the confusion between detection and diagnosis — an anomaly highlights the symptom, not the root cause. Static thresholds that generate noise during peaks and miss genuine deviations during troughs. Deployment without an associated response process: an alert without a dedicated

Use Case 2 — Semantic analysis of reviews

Accumulated reviews (from the website, Google, Trustpilot and marketplaces) are vectorised, clustered by theme, and summarised by an LLM into recurring pain points and untapped selling points. The results are presented by SKU, by category and over time: this allows you to spot a quality issue before product returns skyrocket, or an unexpected use case that deserves to be highlighted in the product description.

Who it’s for: any retailer that accumulates more than 5,000 reviews a year and only uses them for star ratings. Particularly powerful for product ranges where perceived quality varies significantly (fashion, beauty, equipment, electronics).

Why it’s underutilised: reviews are the richest source of information a retailer has about its products – richer than NPS, richer than satisfaction surveys, and more actionable than analytics. And yet almost all retailers are content simply to display them. Systematic semantic analysis transforms this wealth of data into product intelligence, editorial intelligence (what to write in the product description) and operational intelligence (what needs to be corrected in after-sales service, logistics and quality control).

Pitfalls: relying solely on the average star rating when the signal actually lies in the distribution of themes. Launching the project without removing duplicate reviews (Trustpilot and Google collect reviews from the same customers). Attempting to automate everything without human validation of the initial clusters — the quality of the clustering must be audited by a subject-matter expert for the first two iterations.

Solution categories: built-in analytics modules on review platforms (Yotpo Insights, Okendo, Reviews.io), dedicated voice-of-customer tools (Thematic, Reviews Atlas), or a bespoke stack with APIs and an LLM for analysis.

Use Case 3 — Automated CRO diagnostics

The AI combines the CRO’s diverse data sources (funnel, heatmaps, session replays, reviews, anomalies) within a single engine that generates test hypotheses prioritised by expected impact, rather than descriptive reports that nobody has time to read. Rather than being told “your funnel has a 32% drop-off rate at stage 3”, the retailer receives: “here are the three most likely hypotheses; here’s what we suggest testing first; here’s the expected uplift”.

Who is it for: ambitious retailers embarking on a continuous improvement journey who have already explored descriptive tools (GA4, Hotjar, Clarity). They want to move from diagnosis to decision-making. This is the most emerging use case of the three; few tools do it really well – it is precisely the market’s blind spot.

Why this is the bottleneck: all retailers know that their conversion funnel is flawed. Few know where the real losses lie, why they occur, and in what order to tackle them. Traditional analysis identifies symptoms, not hypotheses. Yet in a CRO cycle, what takes time and is costly is not the test itself, but formulating the right hypothesis to test.

Best practice: start by prioritising (which stage of the funnel costs the most in absolute turnover) before delving into the ‘why’. Cross-reference at least three sources for a hypothesis to be considered valid (a drop in the funnel + a cluster review + a heatmap all pointing in the same direction). Keep a human in the hypothesis validation loop — AI makes suggestions, the CRO manager makes the final call.

Pitfalls: confusing hypothesis generation with test execution — AI does not eliminate the need for rigorous A/B testing behind the scenes. Over-investing in tools before resolving the tracking issue upstream (no proper tracking = no diagnosis possible, AI or not). Launching too many hypotheses in parallel without the capacity to execute them.

Solution categories: AI insights modules embedded in CX analytics platforms (Contentsquare, FullStory, Glassbox), dedicated session intelligence tools (Heap AI, Pendo), or a custom stack comprising GA4 events + heatmap API + clustered reviews + Claude as a hypothesis orchestrator for highly technical profiles (a 4–8-week development project).

Achievable ROI — Yes, quickly

Typical payback period: 1–3 months. Once in place, anomaly detection pays off from the very first significant detection. The analysis delivers actionable insights from the very first iteration. Automated CRO diagnostics take longer to master because they depend on the pace of testing, but they speed up the optimisation cycle by 30 to 50 per cent.

Data prerequisites: clean tracking (server-side GA4, consistent data layer), export of orders and sessions to a data warehouse (BigQuery recommended for cost flexibility), and a stable customer ID across front-end and back-end systems. Without these foundations, none of the three use cases will work.

Family 2 — Customer relations

Business objective: to serve customers better, at a lower operational cost, without compromising perceived quality.

The customer relations sector is the one that suffers most from media confusion. Many retailers think “AI in customer relations = chatbot”, whereas these are three distinct applications, with very different ROIs, which need to be implemented in a specific order. Our recommendation: start with the back office, never the front office.

Use Case 1 — Front-end AI chatbot

A conversational assistant on the website that combines a dynamic FAQ section, pre-purchase product advice and information gathering to assess the customer’s needs. The user types their question in natural language, and the AI responds by drawing on the product catalogue, shop policies and FAQs. Depending on its level of development, the bot can also recommend products, manage bookings or escalate the query to a human.

Who it’s for: retailers with complex product catalogues where customers need advice before buying — technical DIY, nutrition with specific dietary requirements, specialised B2B, technical sports equipment, and beauty products with specific routines. For simple product catalogues (basic fashion, accessories), the ROI is marginal and the risk of a degraded user experience is significant.

Why this is nuanced: marketing figures such as “+4× CVR (Conversion Rate) on engaged sessions” are true but misleading. The uplift relates to the segment already engaged with the bot, not to overall traffic. A well-implemented chatbot achieves a 2–8 per cent increase in conversion rate within the segment that uses it (typically 5–15 per cent of traffic), which equates to a 0.3–1.5 per cent increase in the overall conversion rate. A useful addition, but not a game-changer.

Best practice: use templates for all sensitive responses (prices, returns, lead times, availability) — the AI routes the query to the correct template; it does not generate the sensitive content itself. Provide a clear and quick route to a human agent as soon as a complex case arises. Measure the CSAT (Customer Satisfaction Score) after the conversation, not just the resolution rate. Test the friction involved in the bot’s appearance — an intrusive pop-up destroys more revenue than it generates.

Pitfalls: deploying the chatbot before having mastered the back-office agent support (tickets piling up in the background = a degraded overall customer experience). Misinterpretation of shop policies. Conversational dead-ends that force the user to start again via a human chat. Over-investing in marketing the chatbot’s front-end at the expense of its substance.

Solution categories: dedicated e-commerce sales assistants (Rep AI, Zipchat, Manifest AI), platform-native modules (Shopify Sidekick, more limited), enterprise conversational agents (Certainly, Ada), or a custom RAG stack on Claude/GPT with a catalogue + FAQs + policies in a context window.

Use Case 2 — Back-office agent support

The human agent remains at the centre of the conversation, whilst the AI works behind the scenes: suggesting pre-drafted responses based on the customer’s history and the knowledge base, automatically classifying the ticket, and extracting relevant data (order number, reason, urgency). The agent approves, adjusts and sends the response.

Who is it for: any retailer with a support workload that justifies at least one FTE, which in practice means a turnover of €2–3 million or more. This is the customer relationship AI application with the most measurable ROI and the lowest risk — the one that should be rolled out before any other.

Why this is a priority over the front-end chatbot: a human support agent handles 20–40 tickets per day with an average resolution time of 4–12 hours. Agent-Assist cuts this time by a factor of 2–3 without compromising quality, and standardises performance levels between junior and senior staff. The ROI is direct and quantifiable: a saving of 0.5 to 1 FTE at constant volume, or increased capacity as volume grows. No other AI customer relationship module offers this level of predictability.

Best practice: start with repetitive tickets (order tracking, returns, size exchanges) before tackling complex cases. Maintain the knowledge base as a product deliverable rather than as dead internal documentation — it determines 70 per cent of the quality of the suggestions. Measure the average resolution time before and after, and the agents’ acceptance rate of suggestions: if < 50 per cent, the RAG (Retrieval-Augmented Generation) is poorly calibrated.

Pitfalls: rolling out agent assistance without first cleaning up the knowledge base (result: suggestions that miss the mark, agents disabling the tool within two weeks). Attempting to standardise sensitive policies (pricing, returns, disputes) — always use templates; never let the AI improvise. Measuring productivity alone and not CSAT — if perceived quality drops, the gains won’t last.

Solution categories: AI modules for established helpdesks (Zendesk AI, Freshdesk Freddy AI, Gorgias AI, Crisp, Intercom Fin); e-commerce-specific add-ons (Siena, Maven AGI); or a custom RAG stack based on Claude/GPT connected to the helpdesk via API for technical roles.

Use Case 3 — Post-purchase personalisation

All post-order transactional and relationship-based communications (confirmation, dispatch, delivery, satisfaction follow-up, subscription onboarding, review reminders) are dynamically contextualised: product purchased, customer segment, relationship history, preferred channel, appropriate tone. The customer receives a message that feels bespoke rather than part of a generic workflow.

Who it’s for: retailers with high repeat purchase rates (beauty, nutrition, subscriptions, seasonal fashion) where the post-purchase experience is the real driver of CLV (Customer Lifetime Value). It’s also relevant for retailers with high average basket values, where the perceived quality of the post-purchase experience justifies the price premium.

Why it’s underutilised: the majority of retailers stick to static post-purchase workflows that are set up once and never touched again. Yet the post-purchase phase is when the customer is most engaged, most receptive, and when the marginal cost of communication is zero. Contextualised onboarding increases 90-day retention by 10–25 per cent, based on reliable data.

Best practice: start with a limited number of variants (3–5 templates per flow) routed by AI based on the customer’s profile, rather than free-form generation. Measure CTR and retention by variant, not just the open rate. Maintain strong brand consistency — personalisation must not compromise the editorial voice.

Pitfalls: Cosmetic personalisation that merely adds the recipient’s first name and the product name without enriching the message has no measurable ROI. Over-personalisation that reveals to the customer an intimate knowledge of their behaviour, which is perceived as intrusive. Inconsistency across channels (email using a formal tone, text messages using a familiar tone).

Solution categories: AI-enhanced workflows within existing ESPs/CRMs (Klaviyo, Connectif, Brevo, Mailchimp); dedicated post-purchase experience platforms (Wonderment, Malomo); or custom orchestration using Claude + ESP API + CRM data for technical profiles.

Achievable ROI — Yes for agent-assisted support, more mixed for front-end chatbots

Agent-assist: payback period 2–4 months, typically reduces resolution time by 30–50 per cent. Solid and measurable ROI; should be prioritised for roll-out.

Post-purchase personalisation: payback period of 3–6 months, CLV increase of +10 to +25 per cent for repeat customers.

Front-end chatbot: payback period of 4–9 months, real but diluted gains (0.3 to 1.5 per cent overall conversion rate). Should not be over-promoted internally; to be rolled out after Agent-Assist.

Data prerequisites: an up-to-date FAQ database or help center, accessible ticket history (via API export from the helpdesk), a well-maintained product catalog, documented return and shipping policies, and a stable customer identifier shared across both the e-commerce platform and the helpdesk.

Family 3 — Enrichment

Business objective: to create, structure, and maintain catalog content at a scale that would be impossible to achieve manually, while ensuring sufficient quality for SEO, SEA, on-site search, and increasingly, visibility within LLM-powered shopping experiences.

Enrichment is the category that remains the most underserved by the market. The hype around “AI-generated product descriptions” obscures the real challenge: 80% of the value lies not in text generation, but in attribute structuring, image quality enhancement, and taxonomy mapping. These are three distinct use cases that should be addressed separately.

Use Case 1 — Description and Metadata Generation & Optimization

AI generates or rewrites product descriptions, meta titles, meta descriptions, bullet points, and localized variants based on structured catalog attributes, brand voice guidelines, and target keywords. Among more mature merchants, the workflow is often connected directly to Search Console, allowing automatic iteration on underperforming pages.

Who it’s for: any merchant with more than 200 SKUs, as well as businesses operating fast-changing catalogs (seasonal products, supplier restocks, new collections). A particularly important segment is international merchants maintaining 5 to 15 language versions of their catalog—AI can reduce localization costs by a factor of 5 to 10.

Why it matters: the average online store contains roughly 273 products, representing around 550 hours per year of manual description and metadata optimization. With an AI-assisted workflow, this can be reduced to approximately 12 hours. The benefit is not just cost savings; it frees teams to focus on the 20% of products that genuinely require human attention—hero products, new launches, and premium SKUs.

Best practices: feed the model with the brand voice (using the top-performing existing product descriptions as examples), target keywords sourced from Search Console, and a RAG layer built on positive product reviews. Validate performance on a limited set of target pages before scaling. Systematically log prompt_version, model, timestamp, and validated_by fields to enable rollback in the event of SEO regressions.

Common pitfalls: large-scale duplicate content. When hundreds of merchants ask ChatGPT to describe the same Wilson tennis racket using similar prompts, the outputs inevitably converge, and search engines eventually deduplicate them. The solution is to enrich the model with proprietary assets—your customer reviews, buying guides, and brand voice—so the resulting content is genuinely unique. Another risk is gradual SEO degradation following a major Google update. Monitor Search Console performance by page family and keep previous content versions available for rapid rollback when needed. Regulatory content is another critical area: cosmetic ingredient disclosures, toy compliance statements, and health-related claims should never be generated freely by AI. These sections should always rely on pre-approved regulatory templates.

Solution categories: AI enrichment modules embedded in mid-market PIM platforms such as Akeneo AI-Enhanced Enrichment and Plytix; dedicated product-copy platforms such as Describely and Hypotenuse AI; general-purpose content generation tools such as Jasper and Writesonic; or a custom stack combining the Claude API, brand-voice RAG, and CMS write-back workflows for technical teams (typically under €300/month for catalogs of up to 50,000 SKUs).

Use Case 2 — Enhanced Product Visuals

AI now supports three distinct visual workflows: generating lifestyle imagery from a product packshot (placing the product in realistic usage scenarios), batch image editing (background standardization, defect removal, multi-marketplace resizing), and automatic generation of alt text for accessibility and image SEO.

Who it’s for: merchants with sizable catalogs where photography production has become a bottleneck, particularly in fashion, home décor, beauty, and furniture. The value is especially high for catalogs with many product variants (e.g., 5 colors × 4 sizes = 20 lifestyle images that would have been prohibitively expensive to produce through traditional photography).

Why it has become mature: product image generation was largely unusable 18 months ago due to artifacts, distortions, and inconsistent product fidelity. The 2025–2026 generation of models has reached a level of accuracy sufficient for most catalog use cases, provided they are supplied with a high-quality product packshot as input. Unit costs have fallen from several euros per image to just a few cents.

Best practices: always start from a clean studio-quality packshot as the reference image—the quality of the output is directly correlated with the quality of the input. Maintain human review for hero products and any visuals featured on homepage placements. Build and maintain a library of validated prompts by category, as the prompt that produces excellent results for a chair is unlikely to work equally well for a handbag.

Common pitfalls: subtle artifacts on critical details such as coat buttons, packaging text, logos, or fabric patterns. These issues often pass a quick review but become immediately apparent to customers. Another challenge is visual inconsistency when every image is generated independently; a coherent visual style must be actively managed across the catalog. Finally, marketplace compliance remains critical: platforms such as Amazon and Google Shopping enforce strict image guidelines, making compliance audits essential before publication.

Solution categories: dedicated e-commerce lifestyle image generators such as Pebblely, Photoroom, and Booth.ai; batch editing and image-standardization tools such as Claid.ai and Cutout.pro; embedded image-generation features such as Shopify Magic; or custom workflows combining image-generation infrastructure (Replicate, fal.ai) with vision-capable LLMs for automated alt-text generation and catalog enrichment.

Use Case 3 — Attribute Structuring and Taxonomy

AI extracts, normalizes, and enriches product attributes (material, dimensions, compatibility, usage, target audience) from heterogeneous sources such as legacy descriptions, supplier sheets, and images. It also maps internal taxonomy structures onto external standards (Google Product Taxonomy, GS1, marketplace-specific taxonomies) and detects inconsistencies such as misgrouped variants or free-text attributes that should be structured fields.

Who it’s for: any merchant operating with a PIM or CMS where product attributes are incomplete, inconsistent, or stored as free text. This is the most strategic of the three use cases, yet also the least visible. A poorly structured catalog simultaneously degrades SEO performance, breaks Google Shopping feeds, reduces on-site search relevance, weakens recommendation systems, disrupts merchandising logic, and increasingly limits visibility in LLM-based shopping assistants.

Why it is the underestimated core issue: even with excellent product descriptions and high-quality visuals, a catalog with weak attribute structure will not rank properly, will trigger feed disapprovals, will confuse internal search, and will produce irrelevant recommendations. Data structure is a hidden strategic asset. This becomes even more critical in 2026, as LLM shopping interfaces (ChatGPT Shopping, Perplexity, Copilot) rely primarily on structured data rather than marketing copy.

Best practices: start by mapping the internal taxonomy to a standard such as Google Product Taxonomy or GS1 before any other enrichment initiative—this is the foundation layer. Define required attributes per category and track catalog completeness as a core KPI (target: 90%+). Maintain full change logs to enable rollback of enrichment decisions. Regularly audit Google Shopping feeds and marketplace exports to ensure structural integrity over time.

Common pitfalls: semantic drift in attribute values, where AI treats equivalent concepts as distinct (e.g. “navy blue” vs “dark navy”). Another critical risk is loss of data ownership clarity—uncertainty over whether the PIM, CMS, or ERP is the source of truth. This must be resolved before deployment. Finally, organizational complexity is often underestimated: attribute enrichment impacts SEO, SEA, marketing, supply chain, and finance teams simultaneously, requiring strong cross-functional governance.

Solution categories: AI enrichment modules embedded in PIM systems such as Akeneo Unifai and Plytix; feed management and mapping tools such as Productsup and Shopping Feed; native e-commerce platform features such as Shopify; or custom stacks combining LLM-based classification (Claude/GPT), supervised taxonomy mapping, and GS1-aligned canonical data models for enterprise-grade setups.

ROI achievable — Yes, but with a lag (3–9 months)

Unlike analytics and customer service, catalog enrichment pays off indirectly through SEO, SEA, search, and recommendations. The gains compound over time but are not visible at day 30. This is an initiative that must be positioned internally as an infrastructure investment, not a quick win.

Typical gains: time-to-publish –60% to –90%, attribute coverage increasing from 40–60% to 85–95%, +10% to +30% organic impressions on product SEO within 3–6 months, Google Shopping disapproval rate reduced by –50% to –80%, and the equivalent of 1 FTE content manager freed for every 10k–30k SKUs per year.

Critical data prerequisites: a stable internal taxonomy with 3–5 levels mapped to Google Product Taxonomy and/or GS1 standards, typed attributes (no free-text values for structured fields), product images ≥ 1000px with proper alt text, a daily exportable feed (API or CSV), and a clearly defined source of truth (PIM or ERP, not the CMS).

Family 4 — Marketing

Business intent: : deliver the right message, to the right person, at the right time, through the right channel — with decision-making driven by ROI rather than mass sending.

AI marketing in e-commerce suffers from a “copy-first” mindset: the vast majority of educational content focuses on generating subject lines or email copy, even though copy typically accounts for only 10–20% of performance at best. The real performance driver is targeting and timing — in other words, the predictive and operational layer, not the editorial layer. Three core use cases stand out for moving in that direction.

Use Case 1 — Predictive CRM and Behavioral Segmentation

AI calculates predictive scores for each customer (historical and future CLV, category-level purchase propensity, churn risk, and expected next purchase date) and feeds these insights back into the CRM as dynamic attributes. These scores then become segmentation and message-routing criteria, enabling brands to deliver the right content to the right customer at the right time. The logic is driven by behavioral signals and predicted outcomes rather than static rules.

Who it’s for: merchants with a meaningful active customer base (typically 5,000–10,000 customers who have placed at least one order, providing enough data to generate reliable predictive signals). Recurring DTC businesses—such as beauty, nutrition, subscription-based products, and seasonal fashion—tend to see the greatest impact. A particularly important case is merchants already paying for platforms such as Klaviyo, Bloomreach, or similar solutions without having activated the built-in predictive capabilities. This is the majority of the market—and often the easiest and fastest win described in the entire guide.

Why it is the number one lever: a segment built on predicted CLV, expected next purchase date, and category purchase propensity typically outperforms a static RFM (Recency, Frequency, Monetary) segment by 3 to 5 times, while using exactly the same copy (all the messaging used to communicate and enhance the merchant’s offer).
For a €10 million merchant generating 30% of its revenue through email and SMS, moving from RFM-based targeting to predictive targeting can generate an additional €200,000 to €400,000 in annual email revenue—without any increase in platform spend..

Best practices: use predictive scores as segmentation criteria (more stable) rather than as flow triggers (more volatile). Reduce send volume while improving targeting at the same time — both deliverability and platform costs benefit from it. Cross at least two scores for strategic segments (high CLV + churn risk = top retention priority). Retrain models at least quarterly.

Pitfalls: cold start on a dataset that is too small — below ~500 active customers, predictive models in platforms like Klaviyo simply will not activate, and that is actually a good thing: without sufficient statistical signal, predictive scores would produce essentially random recommendations. The churn risk score in isolation — in the “mid-risk” range, 88 to 97% of customers flagged will still churn in practice, which makes the signal weak for prioritizing actions. It must always be combined with a second signal (high CLV, product category, or customer tenure). Over-personalization harming deliverability: fragmenting campaigns into too many micro-segments reduces the email providers’ ability to build sender reputation with inboxes (Gmail, Outlook), causing emails to land in spam even for engaged customers.

Solution categories: native predictive features in ESP/CRM platforms (Klaviyo Predictive Analytics from ~500 customers, Bloomreach Loomi for brands above €10M revenue, Ometria for UK retail/beauty, Splio for French mid-market omnichannel), predictive modules within CDPs (Imagino, Twilio Segment), or custom scoring using BG/NBD + Gamma-Gamma models in BigQuery, pushed to the ESP via API every 24 hours for more technical teams.

Use Case 2 — Advertising Creative at Scale

AI generates and iterates on advertising assets (static visuals, short-form video formats, headlines, descriptions, and CTAs) to enable a volume of creative testing that is impossible to achieve manually.

On mature ad platforms (Meta Advantage+ and Google Performance Max), auction algorithms require creative diversity to optimize performance. A merchant running only 3 creatives per campaign quickly hits a performance ceiling, while a merchant supplying 30 creatives continuously feeds the algorithm, enabling it to discover and capture audience segments that the initial few assets fail to reach.

Who it’s for: merchants with a significant paid media budget (> €50K/year) and a strong reliance on algorithmic ad platforms (Meta, Google, TikTok). This is especially critical for DTC brands whose acquisition depends on Meta for 60%+ of revenue — where post–iOS 14 performance declines can only realistically be offset through creative output and iteration.

Why this has become structural: advertising platforms have shifted from a “targeted audience + single creative” model to a “broad audience + multiple creatives optimized by the algorithm” model. In this new paradigm, creative is effectively the new targeting. A merchant that does not produce creatives at scale becomes blind to the opportunities the algorithm could otherwise capture. Manual production can no longer keep up with the required iteration speed, making AI a functional prerequisite rather than a nice-to-have.

Best practices: structure production around controlled variations rather than random generation (avoid producing 100 unrelated visuals; instead use a framework like 10 messaging angles × 5 visual executions = 50 coherent variants). Run tests on clean A/B cohorts with sufficient budget thresholds (otherwise results are noise rather than signal). Maintain a “creative learnings” library documenting which message structures perform best by audience segment and product type, and use this knowledge base to continuously inform and improve generation prompts.

Pitfalls: over-generation without a strategic framework—producing 200 creatives without underlying hypotheses dilutes media budget and prevents any meaningful learning. Loss of brand visual identity when each variant is treated independently, leading to inconsistent creative expression. Creative non-compliance with platform rules (Meta rejects visuals with excessive text, Google has its own specifications) — assets must always be validated before launch. Finally, AI-generated copy can sound generic and underperform compared to carefully crafted human copy, especially when differentiation and brand tone are key performance drivers.

Solution categories: dedicated ad creative generation platforms (AdCreative.ai, Pencil, Omneky), visual variation tools for product/catalog content (Pebblely, Photoroom — also part of the broader “enrichment” tooling family), native ad platform features (Meta Advantage+ Creative, Google Asset Library AI), or fully custom pipelines combining AI briefing (e.g. Claude), visual generation via tools like Replicate or Midjourney, and API-based deployment to ad platforms for highly technical teams.

Use Case 3 — Demand Forecasting and Predictive Merchandising

AI forecasts future sales at a product level, week by week, over a 1 to 4-month horizon. It combines historical sales data, seasonality, promotional calendars, external events, and trend signals to produce forecasts that are not a single number, but a range of scenarios (median case, conservative case, and a 90% coverage interval). These forecasts directly inform day-to-day operational decisions: what to reorder and in what quantities, how to allocate inventory across warehouses, when to schedule promotions, and which products to prioritize in paid acquisition campaigns based on available stock levels.

Who it’s for: any merchant selling physical products with a catalogue of at least 100 SKUs and real inventory constraints—long supplier lead times, capital tied up in stock, storage costs, or stockouts that directly reduce revenue. This is the least discussed use case in the guide, yet often the most profitable for product-based businesses: demand forecasting delivers value through unlocked cash flow, improved margins, and significant time savings in day-to-day inventory management.

Why it is underexploited: good forecasting typically reduces stockouts by 30–70% and excess inventory by 15–30%. Concretely, for a merchant holding €1M in inventory, this translates into €50K–€150K in freed cash flow and an additional 3 to 8 margin points. No one talks about this on LinkedIn because it is not “sexy” or narrative-driven—and precisely for that reason, few merchants invest in it. Those who do gain a durable operational advantage over competitors, especially in highly seasonal sectors such as fashion, toys, gardening, and beauty.

Best practices: start by ensuring supplier lead times are reliable before implementing forecasting—an accurate forecast based on incorrect lead times is effectively useless.
Always operate with forecast ranges rather than single-point estimates, focusing on the conservative scenario (covering ~90% of outcomes) across the replenishment horizon to avoid stockouts during slightly higher-than-average demand spikes.
Maintain the ability to quickly override or adjust forecasts when unexpected events occur (a product going viral on TikTok, press coverage, or a competitor stockout).
Finally, align paid media calendars with inventory availability—driving ad spend toward products that are about to go out of stock damages both customer experience and ROAS as much as it reduces revenue opportunities.ROAS.

Pitfalls: less than 18 months of sales history leads to unstable forecasts—early-stage merchants are better off using simple threshold-based systems rather than building models that will overreact to the first demand spike. Using an LLM (ChatGPT, Claude) directly for numerical demand forecasting is also a mistake: LLMs perform poorly on time-series analysis. They should be used as supporting tools (e.g. labeling products, enriching attributes, synthesizing merchandising notes), but not for generating or driving the forecasts themselves.

Solution categories: specialized e-commerce demand planning platforms (Inventory Planner by Sage, Cogsy, Prediko for entry-level use cases), enterprise-grade solutions for complex demand patterns (Lokad for highly irregular demand products, ToolsGroup), forecasting modules embedded in ERP and PIM systems, or custom pipelines built on BigQuery using open-source forecasting libraries for technically mature teams with in-house data capabilities.

Achievable ROI — Yes, very directly through activation of existing predictive capabilities

CRM prediction: near-immediate payback from activating features that are already paid for (e.g. Klaviyo Predictive and equivalents). Typically +5 to +15% attributable email/SMS revenue, alongside a –10 to –20% reduction in sending volume. This is the most defensible quick win in the entire guide.

Scaled creative production: 2–4 month payback, with –10 to –25% CPA improvements for merchants with mature acquisition systems.

Demand forecasting: 3–6 month payback, with +3 to +8% EBIT impact, –30 to –70% stockouts, and 10–30 hours per week of operational time saved.

Data prerequisites: a stable customer_id across front-end and back-end systems, properly timestamped order exports into a data warehouse, consistent email matching, up-to-date CRM consent status tracking, and clean supplier lead times for forecasting use cases.

Decision synthesis

A framework for deciding where to start.

Recommended sequence for a merchant with €5–30M revenue starting out

Month 1–2 — Quick wins
Activate the already-paid predictive CRM (Family 4, Use Case 1). Set up minimal anomaly detection (Family 1, Use Case 1). Deploy agent-assist support if volume exceeds 50 tickets/day (Family 2, Use Case 2).

Month 3–6 — Foundational projects
First wave of catalog enrichment on attributes and taxonomy (Family 3, Use Case 3) — this is the foundation that unlocks everything else. Start demand forecasting on the top 20% of SKUs (Family 4, Use Case 3). Deploy semantic analysis of reviews (Family 1, Use Case 2).

Month 6–12 — Industrialization
Systematic generation of product descriptions for long-tail catalog (Family 3, Use Case 1). Scaled ad creative generation if media budget justifies it (Family 4, Use Case 2). Automated CRO diagnostics if testing cadence allows it (Family 1, Use Case 3). Advanced post-purchase personalization (Family 2, Use Case 3). Front-facing chatbot if the catalog is complex and back-office processes are under control (Family 2, Use Case 1).

To be deferred beyond 12 months:
Advanced 1:1 personalization (not enough signal below 500k monthly sessions), AI-driven multi-touch attribution (ROI diluted below €5M revenue), front-end agentic autonomous agents (still experimental in production in 2026).

The principle that emerges from the 4 families

The four families share three counterintuitive characteristics that explain why most merchants fail when entering AI.

The highest-ROI use cases are the least spectacular. Activating an already-paid Klaviyo feature, cleaning up a product catalog, detecting a CVR drop 48 hours earlier, clustering 20,000 reviews—none of these can be demoed in 30 seconds. All of them outperform a viral chatbot.

Clean data matters more than the model. A merchant with broken tracking and a messy catalog will get noise from any AI, even the most advanced. A merchant with clean tracking and a structured catalog will see meaningful gains even from a relatively average model.

Don’t buy AI, buy the problem being solved. Generative AI shines because it offers solutions before clearly defining the problem. The four families in this framework reverse that logic by forcing the problem to be defined first—which is exactly why they deliver ROI. The question is never “how do I use AI?” but “what is costing me the most and why?” AI comes afterward.