Online Retail With AI: Eleven Workflows From Purchasing to Product Copy, and Where It Pays Off

An online shop is a chain of many small workflows: buying stock, setting prices, managing inventory, reconciling payments, answering questions, running campaigns, writing copy. Most of them repeat every day, and that is exactly what makes online retail a good field for automation.

AI is not the right tool everywhere, though. It is strong where forecasts, language or patterns in large amounts of data are needed. Stock synchronisation, mandatory information and payment rules, on the other hand, are handled more reliably and more cheaply by fixed rules. This article goes through eleven areas: what AI takes over, where people decide, what evidence exists and what the law says.

The starting point

According to the HDE Online-Monitor, online retail of goods in Germany generated 92.3 billion EUR netU in 2025, 3.9 percent more than the year before. That is 13.5 percent of total retail. For 2026, the German Retail Federation (HDE) expects growth of 4.3 percent. 56.7 %U of online sales went through marketplaces. Amazon.de, with its own retail and marketplace combined, accounts for 63.3 %U.

On AI adoption, retail sits just below the average across all sectors. According to Eurostat, 23.4 %A of retail companies in Germany with at least ten employees used AI in 2025, compared with 26 percent across all sectors. In EU retail, the share was 15.5 percent. Among mid-sized and large retailers, the share is higher: in an HDE survey of 175 companies, 47.3 %U said they use AI, twice as many as in 2023.

Customers have changed too. According to the HDE, 35 % of peopleU in Germany already use AI chatbots to research products.

Overview: eleven areas

The workflow

The areas are connected. What the purchasing forecast proposes determines stock and price. What comes in through support and returns feeds back into product copy and purchasing. The workflow shows this loop.

Clean master data is the foundation of everything: item number, dimensions, material, manufacturer, safety information, images. An AI that creates copy, listings or forecasts is only as good as this data. Many retailers keep it in a PIM system (product information management) or directly in the inventory system. What matters is that there is exactly one source that feeds all channels.

1. Purchasing: forecasts instead of gut feeling

Purchasing determines how much money is tied up in stock and how often items sell out. AI models estimate demand per item from sales history, seasonality, promotions, prices and other signals. The result is a purchase proposal that the buying team reviews.

The M5 forecasting competition, based on Walmart sales data, shows how much better methods can deliver. The best method, a machine-learning approach, was 22.4 % more accurateS than the best statistical benchmark. Only 7.5 % of teamsS beat that benchmark at all. The lesson: good forecasts are possible, but simple methods are a tough yardstick. Anyone introducing an AI tool should test it against the method used so far.

From practice, the evidence is mostly vendor claims. At Otto, which has worked with Blue Yonder for years, 90 % of goods orderedH sell within 30 days, according to a case study.

Forecasting is only now arriving in standard tools. Since October 2026, JTL has been testing its agent Juno in a closed beta; it checks stock levels and prepares purchase proposals for approval.H Shopify says Sidekick forecasts demand and can create purchase orders, without disclosing its method or accuracy.H In other inventory systems, purchase proposals still work with fixed or dynamic reorder points.

How to put it into practice:

• Classify items: Forecasting AI pays off for items with a sales history and meaningful revenue. New and slow-moving items stay with rules or comparisons with similar items.
• Keep approval: The AI proposes, purchasing orders. The system flags deviations above a threshold for review.
• Measure: Compare forecast error, stock-outs and days of stock before and after the introduction.

How purchasing itself gets leaner, with ordering process and suppliers, is covered in Procurement With a System.

2. Pricing strategy: guardrails first

Prices in online retail often change several times a day. AI can suggest prices per item based on demand, stock, competition and margin. Strategy stays with people: minimum margin, price thresholds, maximum change per day, promotion periods.

A study by Zalando shows what this can deliver. In 23 A/B tests across twelve countries with around 6.2 million items, profit rose by 6.0 %S with the new pricing system. Revenue and sales after returns, by contrast, did not change measurably. The study comes from Zalando itself and has not yet been peer-reviewed.

There are four legal points to note, based on German and EU law:

• Price reductions: Anyone announcing a price reduction must state the lowest total price of the previous 30 days (Section 11 PAngV). According to the CJEU, percentages and advertising claims must refer to exactly this price. A price that an algorithm briefly raises and then advertises as reduced therefore does not make a permissible reduction.
• Personalised prices: If a price is adjusted for individuals on the basis of automated decision-making, the shop must say so (Art. 246a Section 1 EGBGB). Dynamic prices that apply equally to everyone are not covered.
• Competition law: According to the European Commission's Horizontal Guidelines, companies cannot escape liability on the grounds that their prices were set by algorithms. Coordination by software is still a cartel.
• What is coming: The European Commission plans to propose the Digital Fairness Act in the fourth quarter of 2026. It is expected to address unfair personalisation too. No draft has been published yet.

3. Inventory and merchandise management: spotting anomalies

Synchronising stock between shop and marketplaces is a job for rules, not AI. AI helps elsewhere: it spots anomalies, such as items with a sudden drop in sales, unusual stocktaking differences or goods that sit too long. And it answers questions in plain language, such as: "Which items will last less than 14 days at current sales?"

The major inventory systems are building in such features now. Xentral offers a co-pilot that searches ERP data. According to Xentral, AI is included in every plan, and retailers decide how much they use it.H At plentyONE, reports can be requested in a chat with the AI assistant since May 2026, initially as a limited-time trial.H

What matters is that the AI reads from the database and shows its query. A number that the model merely "estimates" has no place in a warehouse decision.

4. Payments and accounting: matching and checking

In German online retail, according to EHI, PayPal leads with 28.7 %U of revenue, followed by purchase on invoice with 26.1 %U. Direct debit (14.4 percent) and cards (13.7 percent) come next. The high share of invoice purchases makes fraud checks and dunning important.

AI helps in three places:

• Fraud risk: Payment providers score orders with models trained on large amounts of data. Stripe cites an average fraud reduction of 32 %H for its Radar product, without method or period.
• Matching: Payouts from marketplaces and payment providers often contain lump sums with fees, returns and chargebacks. AI suggests which orders and fees belong to a payout. Accounting confirms.
• Chargebacks: AI sorts chargebacks by reason and compiles the evidence for disputing them.

Two rules change the framework. Since 9 October 2025, banks in the euro area check before transfers whether the IBAN and the payee name match. This protects, for example, against fake supplier invoices with someone else's account number. And e-invoicing applies to business-to-business transactions in Germany: all companies must be able to receive them since 2025; companies with more than 800,000 EUR prior-year revenue must issue them from 2027, everyone else from 2028. This does not apply to invoices to consumers.

How incoming invoices are booked semi-automatically is shown in From Invoice to Ledger. For open invoice purchases, the workflow in Semi-Automated Dunning helps.

5. Customer support: status automated, complaints handled personally

According to the EHI E-Commerce-Report 2026, 40.8 %U of the 1,000 most important online shops in Germany offer a customer chat. In 77.9 %U of these, AI is involved. Customers are more sceptical: in a Bitkom survey of online shoppers, 62 percent prefer a quickly available human, 36 percent a chatbot. According to Gartner, 87 %U of customers say companies using AI in customer service must offer a way to reach people.

A series of field experiments on a large cross-border retail platform shows that well-designed assistance can increase sales. An AI chatbot for pre-purchase questions raised sales there by 16.3 %S. The study has not yet been peer-reviewed.

A proven workflow:

• Classify: The AI sorts every request: delivery status, return, defect, invoice, product question, complaint.
• Answer status questions: Questions such as "Where is my order?" are answered automatically from order and shipping data.
• Draft: For product questions and returns, the AI writes a draft, a person reviews and sends it.
• Hand over: Complaints, goodwill, legal questions and angry customers go straight to the team.

Assign the following customer request to exactly one category: DELIVERY_STATUS, RETURN, DEFECT, INVOICE, PRODUCT_QUESTION, COMPLAINT, OTHER. Use the attached order data. Return JSON: category, order_number (or null), mood (calm, dissatisfied, angry), needs_human (true or false) with a one-sentence reason. Set needs_human to true for COMPLAINT, for legal topics and for an angry mood. Do not invent order data.

Since 2 August 2026, a chatbot must disclose under Art. 50 of the AI Act that it is an AI, unless that is obvious. There is no general right to a human contact in the EU yet. What the numbers show about AI assistance in customer service overall is covered in AI-Assisted Customer Service.

6. CRM: segments and predictions

In customer management, AI takes on three tasks. It builds segments from a plain-language description, such as "customers who bought running shoes twice and have not ordered for 90 days". It predicts customer value, churn risk and the likely next order date. And it picks the send time for each person.

Such predictions need data. Klaviyo, for example, requires at least 500 customers with orders and 180 days of order history for its predictions.H At Brevo, AI-based send-time optimisation is included from the Standard plan, AI segments only in the Professional plan.H

Personalisation works, but often differently than expected. In a field experiment published in Marketing Science, simply adding the recipient's name to the subject line raised opens by 20 %S and enquiries by 31 percent. Unsubscribes fell by 17 percent.

In Germany, email advertising to existing customers without consent must meet four conditions under Section 7(3) UWG: the address comes from a purchase, the advertising is for the seller's own similar products, the person has not objected, and they were told about their right to object when the address was collected and in every email. All other emails need consent. How AI gets notes and follow-ups into the CRM in day-to-day sales is shown in Call Notes, CRM, Follow-Up.

7. Process and sales analytics: questions instead of dashboards

AI is changing how retailers work with their numbers. Instead of building reports, they ask questions. Shopify Sidekick writes the queries, creates charts and exports reports. According to Shopify, Sidekick only makes changes to the shop after approval.H

Analysing free text is particularly worthwhile: return reasons, reviews, support requests. The need is great. According to the Returns Management Research Group at the University of Bamberg, around 542 million parcelsS were returned in Germany in 2025, almost one in four parcels in consumer online retail. Over 80 percent of them are fashion. A European retailer survey by the research group put the cost in 2022 at an average of 2.85 EUR per returned itemS, for transport and processing alone. According to EHI, 60.6 %U of the retailers surveyed now cover the full cost of return shipping.

Analyse the attached return reasons and reviews for one item. Form no more than six categories, for example "runs small" or "colour differs from the photo". For each category, give the count, two typical quotes and one concrete action: adjust the size chart, correct the product copy, replace the photo, talk to the supplier. Count only what is in the data and do not estimate numbers.

Here too, numbers come from the database, not from the language model. Have the underlying query shown for every analysis.

8. Marketplace management: one data set, many channels

Marketplaces are the largest sales channel for many retailers. Each marketplace requires different categories, attributes and text lengths. AI drafts the matching listings per channel from one master data set. Fixed rules then check whether all mandatory fields are filled.

The marketplaces offer their own AI tools. According to Amazon, sellers created more than 12 million listingsH with generative AI in 2025. Back in 2024, sellers accepted the suggested attributes nearly 80 percent of the time, according to Amazon. eBay offers private sellers in Germany an AI feature that creates a listing from a photo.H

Two obligations apply regardless of the tool:

• Product safety: Since 13 December 2024, online offers must show, under Art. 19 of the General Product Safety Regulation (GPSR), among other things the manufacturer with postal and email address, for manufacturers outside the EU the responsible person in the EU, a product image and the safety information in the local language. An AI can check every listing for this information before publication.
• Trader information: Under Art. 30 of the Digital Services Act, marketplaces must collect and verify the identity of their traders. Keep this data up to date, or you risk being suspended.

9. SEO: visible in search and AI answers

Search is changing fast. Google has shown AI Overviews in Germany since March 2025 and AI Mode since October 2025. According to Pew Research, users clicked a traditional result in only 8 %U of searches with an AI summary, compared with 15 percent without one. Analyses by agencies show sharply falling click-through rates on queries with AI Overviews. Google itself stresses that the remaining clicks are of higher quality.

What does this mean for shops?

• AI copy is allowed, mass production is not: According to Google, "appropriate use of AI" is not against its guidelines. Generating many pages with generative AI without adding value for users, on the other hand, counts as spam.
• Maintain the data: According to Google, AI Overviews and AI Mode need no special markup. But structured data must match the visible text, and the information in Merchant Center must be up to date.
• Shopping through chatbots: OpenAI launched purchases directly in ChatGPT in the US in September 2025 and scaled them back in March 2026. In January 2026, Google presented the open Universal Commerce Protocol with Shopify, Etsy, Wayfair, Target and Walmart. As far as our research shows, there is no checkout directly in a chatbot in Germany yet. What counts everywhere is complete and correct product feeds.

10. Performance and online marketing: automation with control

The big ad platforms are automating bids, audiences and increasingly the ads themselves. Google states that Performance Max delivers on average 27 % more conversionsH or more conversion value at a similar cost. At Meta, automated advertising solutions reached an annual run rate of more than 75 billion USDH in the second quarter of 2026, according to the company. According to the Wall Street Journal, Meta aims to reach the point by the end of 2026 where brands only supply a product image and a budget.

Independent evidence is rarer. The measurement provider Haus analysed 640 tests on the incremental effect of Meta campaigns. For 58 %D of brands, manually managed campaigns delivered more incremental revenue per advertising euro than the automation. In the field experiment on the retail platform, AI-optimised ad titles brought no measurable benefit either.

This leads to a clear division of labour:

• The automation steers: bids, budget allocation, ad variants.
• People set the frame: margin targets instead of revenue targets, exclusions, brand rules, approval of new creatives.
• Test the effect: Measure regularly with control groups, such as regions without campaigns, how much revenue is truly incremental.
• Consent: For users in the EEA, Google has required consent signals via Consent Mode since March 2024; without them, measurement and personalisation are lost.

How AI speeds up feedback and campaign work in marketing is covered in AI in Marketing Ops.

11. Product content: write faster, check the facts

Product copy, translations and alt text for images are among the most obvious use cases. In the field experiment mentioned above, with around 4.8 million customers, AI-generated product descriptions increased sales by 2.1 %S. With large ranges, that adds up.

To get the quality right, the AI writes only from the master data:

Write a product description for the online shop. Use only the attached product data. Structure: one sentence on the benefit, then three to five bullet points on material, dimensions and care, then a short paragraph on use. No more than 120 words, informal tone, no superlatives. Also write an alt text for the main image of no more than 125 characters. If information such as material or dimensions is missing, write "MISSING: [item]" instead of inventing anything. Do not mention any features that are not in the data.

Four rules apply to product content:

• Accessibility: Since 28 June 2025, online shops for consumers in Germany must be accessible under the Accessibility Strengthening Act (BFSG). Images need text alternatives. Micro-enterprises with fewer than ten employees and no more than 2 million EUR in turnover or balance sheet total are exempt, as far as they provide services.
• Mandatory information: The information required by the General Product Safety Regulation belongs in every offer, see section 8.
• Product images: Google Merchant Center requires AI-generated images to keep a corresponding label in their metadata. The main image must show exactly the item being sold.
• AI Act and competition law: An AI image that shows a product differently from how it really looks can count as a deepfake under the AI Act and must then be labelled. It is misleading, and therefore unfair competition, in any case. Swapping only the background of real product photos is less sensitive, as long as the product itself stays unchanged.

Limits and risks

• Data before models: Incomplete master data leads to wrong copy, listings and forecasts. Most of the work goes into maintaining the data.
• Automation without brakes: Pricing and advertising algorithms react to each other. Without limits on price changes and budgets, errors can escalate within hours.
• Vendor claims dominate: Most figures on AI in retail come from vendors. Independent evidence such as the M5 competition or the Pew study is rare; the Zalando study and the retail platform study have not yet been peer-reviewed.
• Data protection: Customer data in AI tools requires a data processing agreement, a legal basis and data minimisation. Additional rules apply to profiling and personalised advertising.
• Dependency: The more revenue runs through marketplaces and automated advertising, the more their rules shape the business. Your own customer data and your own shop remain important.

How to start

1. Take stock: For each of the eleven areas, note how many hours per week it costs and where errors occur. Choose three areas with high effort and clear rules.
2. Check master data: Fill in missing mandatory information, attributes and alt text before an AI works with it.
3. Quick wins: Start with support classification including automatic status replies, product copy drafts with approval and the analysis of return reasons.
4. Set guardrails: Define minimum margins, maximum changes and budget limits for pricing and advertising before you automate.
5. Measure the effect: Choose one metric per area, such as forecast error, margin, first response time, return rate or conversion, and test against a control group where possible.

Sources

1. UHDE Online-Monitor 2026 (data: IFH Köln)
2. AEurostat: Use of artificial intelligence in enterprises
3. UHDE: KI im Handel 2025 (AI in retail)
4. SMakridakis et al.: M5 accuracy competition, International Journal of Forecasting 2022
5. HRetail Systems: Otto and Blue Yonder
6. HJTL: AI agent Juno
7. HShopify: Enhanced inventory management
8. SZalando: High-Frequency Pricing at Scale for E-Commerce (preprint 2026)
9. RGerman Price Indication Ordinance (PAngV), Section 11
10. RCJEU, C-330/23 (Aldi Süd), judgment of 26 September 2024
11. RIntroductory Act to the German Civil Code (EGBGB), Art. 246a Section 1
12. REuropean Commission: Horizontal Guidelines 2023
13. REuropean Parliament: Legislative Train, Digital Fairness Act
14. HXentral: Pricing and plans
15. HplentyONE: Product updates
16. UEHI: Zahlungssysteme im Einzelhandel 2026 (payment systems in retail)
17. HStripe Radar
18. RRegulation (EU) 2024/886 (instant payments, verification of payee)
19. RGerman VAT Act, Section 27 (e-invoicing transition rules)
20. UEHI: E-Commerce-Report Deutschland 2026
21. UBitkom: Customer service in online shopping
22. UGartner: 87 percent say GenAI customer service must offer access to a human
23. SFang et al.: Generative AI and Sales Productivity, Field Experiments in Online Retail (preprint)
24. REuropean Commission: Guidelines on transparency obligations under Art. 50 AI Act
25. RGerman Act Against Unfair Competition (UWG), Section 7
26. HKlaviyo: Predictive analytics
27. HBrevo: Plans and features
28. SSahni, Wheeler, Chintagunta: Personalization in Email Marketing, Marketing Science 2018
29. SReturns Management Research Group, University of Bamberg: return parcels 2026
30. SUniversity of Bamberg: First European retailer survey on returns management
31. UEHI: Returns study 2026
32. HShopify Help: Sidekick
33. HAmazon: Small Business Empowerment Report 2025
34. HAmazon: Generative AI for product listings
35. HeBay: Magical listing
36. RRegulation (EU) 2023/988 on general product safety, Art. 19
37. RDigital Services Act, Art. 30
38. HGoogle: AI Mode launches in Germany, Austria and Switzerland
39. UPew Research Center: Clicks when an AI summary appears
40. DSeer Interactive: AI Overviews and CTR, September 2025
41. HGoogle Search Central: AI features and your website
42. HGoogle Search Central: Spam policies
43. MDigital Commerce 360: OpenAI shifts checkout plans
44. HOpenAI: Product feed specification
45. HGoogle: Universal Commerce Protocol
46. HGoogle Ads Help: Performance Max
47. HMeta: Earnings call Q2 2026
48. MReuters: Meta aims to fully automate advertising with AI (citing the Wall Street Journal)
49. DHaus: Lessons from 640 incrementality experiments on Meta
50. HGoogle Tag Manager Help: Consent mode
51. RFederal Accessibility Office: Online shops and the BFSG
52. RGerman Accessibility Strengthening Act (BFSG), Sections 2 and 3
53. HGoogle Merchant Center: Image requirements