AI in ecommerce: the right kind of AI for each operations task
AI in ecommerce is a task-by-task choice, made in this order: where an error is costly and irreversible, a person decides; if you set the parameters, a setting is enough, with no AI; text to read goes to an AI assistant; otherwise, a model learns from the data. In our 26 published customer cases, 23 gains came from settings.
The essentials
- If you can set the parameters, a setting is enough, as with a supplier reminder; if the data has to set them, you need a model.
- A task splits into steps: when a supplier emails about a delay, only reading the email takes AI; the rest is a setting or a person's call.
- At a multi-brand lifestyle retailer, configured replenishment, with no AI, cut purchasing time from 2 days a week to 30 minutes.
- Let the confidence score decide who approves: at a motorcycle parts distributor, matches above 96% are auto-approved; between 90% and 96%, a team member reviews them.
- Judge an AI assistant on repeat contacts: in a trial with 5,940 new reps, it sped up service with no significant effect on repeat contacts.
- Set three dials before connecting an AI assistant: read-only at first, drafts that aren't sent, and explicit approval for every write.
In our 26 published customer cases, AI in ecommerce was behind only three of the gains: matching records with no shared code (two cases) and deciding automatically which products to reprice often. The other 23 came from settings.
Under the EU AI Act, an AI system "infers" its outputs from its inputs, and the definition "should not cover systems that are based on the rules defined solely by natural persons" (recital 12). So the test is one question: who sets the parameters? Answer it to place each task in one of four boxes (human decision, setting, AI assistant, or model), each with its own data, check, and autonomy.
Map where AI in ecommerce fits
Read the map top to bottom, then left to right. On the storefront, recommendations go to the model; product copy and the pre-sales chatbot go to your AI assistant.
Retailers have been slower to adopt AI: according to Eurostat, 15.46% of EU retailers with 10 or more employees used it in 2025, compared with 19.95% of all EU businesses of that size. Among EU businesses that use AI, 34.70% use it for marketing or sales, 6.08% for logistics (Eurostat, 2025).
The machine prepares, a person decides
- Supplier spend
- Customer compensation
- Picker evaluation
Check: test on past orders
- Reorder point
- Supplier reminder
- Carrier selection
- Lot tracking and FIFO
Check: the source, repeat contacts
- Where-is-my-order ticket
- Supplier email
- Product descriptionStorefront
- Pre-sales chatbotStorefront
Check: the confidence score
- Matching without EANs
- Private-label equivalents
- Products to reprice
- RecommendationsStorefront
Top to bottom first: if an error is costly and can't be undone, a person decides. Then left to right, by who sets the parameters.
Each box fails in its own way and starts with its own level of autonomy.
| Box | How it fails | Data it needs | Check | Starting autonomy |
|---|---|---|---|---|
| Human decision | Rubber-stamping | The file the machine prepared | Who approved, and when | None for the machine |
| Setting | Set up wrong, the same way every time | Lead time, quantity, order number | Test on a month of orders | Automatic after the test |
| AI assistant | States what it never read | The text and actual order status | Cited source, repeat-contact rate | Drafts that a person sends |
| Model that learns | When data is thin or unusual | Images, titles, attributes, and a validated history | Confidence score, reviewed sample | Set by a confidence threshold |
Spot where AI made the difference
Our 26 published customer cases were sorted by the solution each case page describes: a case counts as AI when that solution names an AI, LLM, or machine-learning component. Twenty cover orders, inventory, purchasing, the warehouse, shipping, or point of sale, and three cover costs, commissions, or price gaps. All of them work on quantities, lead times, statuses, or prices that the team configures, and their 23 gains came from settings. The 3 AI cases are the ones where the data had to set the parameters.
- Retail, France2.94% · 9.27% · 9.93%
- All businesses, France5.88% · 9.91% · 18.16%
- Retail, EU6.08% · 11.29% · 15.46%
- All businesses, EU8.06% · 13.48% · 19.95%
- SettingsQuantities, lead times, statuses, prices: set by the team
- 23
- AI componentRecords with no shared code, products to reprice often
- 3
Left: retail stays below the average, and the gap widened in 2025. Right: our 3 AI gains came from tasks where the data sets the parameters.
Triage the task in three questions
Ask them in the map's order.
| Question | If yes | If no |
|---|---|---|
| Is an error costly and irreversible? | Human decision | Next question |
| Do you set the parameters yourself? | Setting | Next question |
| Is the input text to read? | AI assistant | Model that learns |
A supplier reminder written from a late purchase order stops at the second question: its inputs are fields (order number, promised date, quantity), and you set the timing and the message template. The supplier's reply is text to read.
Replenishment stops at a setting too. At a multi-brand lifestyle retailer, supplier lead times, safety stock, and the sales windows the system compares with last year were configured; the buyer approves the proposed orders. Purchasing time dropped from 2 days a week to 30 minutes, and stockouts on top sellers from 15-20 a month to zero within the first quarter, with no AI. The ecommerce replenishment guide covers those thresholds, and the ecommerce automation guide covers what to automate first. The Boostmyshop purchasing module calculates the reorder points and drafts the purchase orders.
Hand the reading and drafting to your AI assistant
Your AI assistant is a chat assistant connected, read-only at first, to the system that holds your orders and inventory. It reads each message, pulls up the orders by number, and drafts a reply that a person sends. Without them, it would fill the gap with something plausible. Our MCP server connects it to your data in Boostmyshop: orders, warehouse, shipping, and purchasing.
Check the cited source for every fact taken from a document, and track repeat contacts on tickets. At a software company's support desk, issues resolved per hour rose 15% on average with suggested replies (Brynjolfsson, Li, and Raymond, 2025). In a randomized trial in January and February 2024 with 5,940 new reps at a large marketplace (Ni et al., 2026), service sped up with no significant effect on repeat contacts for the same issue within three days. Count them against contacts handled, before and after.
Bring in a model when the data has to set the parameters
The most basic model is a statistical forecast: it takes its parameters from your sales history. In the M5 forecasting competition (2022), the winning team's machine-learning method beat the best statistical benchmark by 22.4%, but the top methods' gains were concentrated at aggregate levels: at the product-store level, standard statistical models stayed competitive. For a single SKU, compare your configured thresholds with a standard statistical forecast on a month of orders before going further.
You need a richer model when no shared identifier links your records to competitors' listings, there are too many variables for a readable setting, and the volume rules out a full review. Its confidence score becomes the autonomy rule. A motorcycle parts and equipment distributor needed to match its 200,000 SKUs to competitors' listings without reliable EAN codes. The engine compares images and text, then an LLM arbitrates and scores each match. The published case record sets the rule: "Above 96% confidence, matches are auto-approved. Between 90-96%, a team member reviews them weekly." Matching accuracy reached 99%. "After 6 months of validated matching, they activated automated pricing." That is how myPricing competitive intelligence works: every match carries a score, and you set the thresholds for auto-approval and review.
200,000 SKUs to match at a motorcycle parts distributor
- Read
The engine compares images and text
- Propose
Each match gets a confidence score
"matches are auto-approved"
"a team member reviews them weekly"
- Validated matching
- 6 months
- Automated pricing
"After 6 months of validated matching, they activated automated pricing."
The score decides who approves each match; six months of validated matching earned a second step: automated pricing.
The other two AI cases each show part of the mechanism: an organic cosmetics seller and manufacturer gets scored equivalents for its private-label products, and a high-tech marketplace seller lets AI decide automatically which products to reprice at high frequency, about 30% of the catalog. The system reclassifies products continuously, so a wrong call can be undone and that task stays below the human-decision line.
Keep a person on the calls that cost the most
A decision stays human when an error is expensive and can't be undone: approving supplier spend, granting customer compensation, evaluating a picker. The machine does the prep work; the person who already approves that kind of action makes the call. Set three dials before you connect anything.
| Dial | Start with | What to log |
|---|---|---|
| Read | A dedicated, read-only user | Data opened for reading |
| Prepare | Drafts and proposals, nothing sent | The draft and its source |
| Act | An explicit yes for every write | Who approved, and when |
An AI agent is an AI assistant whose "act" dial is turned on: the same dials apply. The OWASP Top 10 for LLM applications (2025) traces excessive agency to excessive functionality, permissions, or autonomy, and recommends that a human approve high-impact actions before they're taken. The EU AI Act has a name for rubber-stamping: automation bias. To counter it, send the person the uncertain band, like the 90-96% weekly review at the motorcycle distributor, and log what they reject. With the Boostmyshop MCP server, a dedicated user's role determines what your AI assistant can read and do. Read: you can choose read-only access when you sign in. Prepare: a preview comes first. Act: nothing is created without your approval, and sensitive actions require separate confirmation.
Run one task through the grid
Take a supplier email pushing back a delivery. Only reading it takes AI.
| Step | Box | Why |
|---|---|---|
| Read the email and propose the new date | AI assistant | Free text |
| Find the purchase order and the backorders tied to it | Setting | Fields linked by order number |
| Enter the date on the purchase order | Setting, once the buyer approves | A write, so explicit approval |
| Tell each customer the confirmed date | Setting | A template filled from fields |
| Pay for express shipping to keep the date | Human decision | Spend you can't undo |
The free program "AI, up and running" (8 webinars through March 31) works through the method: checking your AI assistant's work (session 4), choosing the box (session 5), connecting AI to your tools (session 6), and giving the whole team access safely (session 7). Sessions run live in French, with replays and materials in English. Session 1 helps you pick your first task and bring your team along, with a starter checklist: register for Wednesday, November 25, 2026, at 11 a.m. Paris time (5 a.m. ET).
Sources
- Eurostat, Artificial intelligence by NACE Rev. 2 activity, 2026, and Use of artificial intelligence in enterprises, 2025.
- Regulation (EU) 2024/1689, the AI Act, 2024, amended by Regulation (EU) 2026/1744, 2026.
- Makridakis et al., M5 accuracy competition, International Journal of Forecasting, 2022.
- Brynjolfsson, Li, and Raymond, Generative AI at Work, Quarterly Journal of Economics, 2025.
- Ni et al., Generative AI in Action: Field Experimental Evidence from Alibaba's Customer Service Operations, arXiv, 2026.
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency, 2025.
- Boostmyshop, 26 published customer cases, 2026.
Frequently asked questions
On the storefront, it recommends products, writes descriptions, and answers pre-sales questions. In operations, it matches catalogs that share no product code and drafts replies to support tickets, while settings still run reorder points and carrier selection. At one motorcycle parts distributor, matching reached 99% accuracy across 200,000 SKUs.
Pick by task, not brand, in this order. A costly, irreversible error: a person. Parameters you set: a setting. Text to read: an AI assistant. Parameters only the data can set: a model. So carrier selection is a setting; matching a catalog with no shared EAN is a model.
Not in EU retail so far. Eurostat data, updated June 15, 2026, puts 2025 AI use at 15.46% of EU retailers with 10 or more employees, compared with 19.95% of all EU businesses of that size. Among EU businesses using AI, logistics is the least-cited purpose.
Yes: a statistical forecast is the most basic model, and machine learning's gains concentrate on aggregate demand (M5 competition). For a single SKU, compare the forecast with your thresholds: at a multi-brand lifestyle retailer, configured thresholds, with no AI, cut purchasing time from 2 days a week to 30 minutes.
Not necessarily. Your AI assistant reads data where it already lives. If your orders, inventory, and purchasing run in Boostmyshop, our MCP server gives your AI assistant access. Start with a dedicated user and choose read-only access when you sign in, then widen that user's role step by step.
The person who already approves that kind of action. The EU AI Act requires this oversight only for high-risk systems: from December 2, 2027, for Annex III systems such as employee evaluation, and from August 2, 2028, for Annex I. For other systems, OWASP recommends human approval of high-impact actions.
In the EU, yes. Since August 2, 2026, Article 50 of the EU AI Act requires the provider to design the chatbot so people know they're talking to an AI, unless that's obvious. If you build your own chatbot and run it under your name, you're that provider.
Continue exploring
Stock availability, picking accuracy, and shipping times affect marketplace sales. Jonathan Coloma, COO of Repiauto, explains how his team went from 300 to 1,000 orders a day in two years. He describes the limits of paper-based work, ERPs built for different needs, a go-live estimate of fifteen days, and changes to his daily routine.
To hold up under peak volume, split the order queue into waves based on real operating constraints, then decide where barcode checks belong. At the packing bench, the operator verifies the parcel. Carrier rules select the service and print the label, so the next order can start without another decision.
Industrializing your purchasing means ordering on rules, not guesswork: at the right time (the reorder point), in the right quantity (your coverage period), from the right supplier, with a purchase order validated in one click. Available on replay.
On December 16, set up a project with instructions and reference documents, then produce a Monday operations brief from an export with no customer data. You will leave with the workspace and a record of the requests that work for you. The live session is in French; the replay and training materials will be available in English.
A well-interpreted search reveals what customers want, even without a sale. This Boostmyshop × Venture.ia replay shows how Findy turns that signal into conversions, then how Boostmyshop reserves stock and opens its data to Héhocom so picking and delivery can keep up with order growth.
On January 20, find out in five minutes who already uses which AI tools, prepare a one-hour awareness session and get two volunteers started. You will leave with the outline for that session, a plan for the volunteers and three rules covering company tools, a fixed-price subscription and customer data.
On March 31, compare your project with its baseline for time and errors, and work out what it costs. Decide what to scale, what to stop and when to reduce review. You will leave with a before-and-after comparison, a 90-day plan, the method on one page and the outline of your next project.
On February 17, score 10 tasks for operational pain, difficulty and whether the result can be checked. For each one, decide whether it needs a setting, your AI assistant or a process change, or whether it can wait. You will choose one verifiable task and record its baseline with 5 timed cases and the previous month’s errors.
On March 3, decide what your AI assistant reads, what it prepares, what it must not change and who approves. You will leave with the connection map for your project: the sources, a dedicated read-only account to start and the person responsible for approval.
A successful migration does not replace everything at once. It starts by deciding what to keep, replace or connect, then sequences each move around the work people actually do. On December 10, Boostmyshop and Dedi Agency will map those choices from the storefront to day-to-day operations.
AI still feels like a black box when teams do not know what it actually does. On November 25, David Butin will draw on lessons from three years of using AI at Boostmyshop. He will answer a broad question: where do you start if you want to explore AI, get your organization moving and bring your team along? He will explain, without jargon, what an AI assistant can and cannot do and why critical thinking still matters. The practical perspective will come from e-commerce business leaders at the roundtable, including Arnaud Darsonval of La Ferme de Sainte-Marthe. The live session will be in French. The replay and training materials will be available in English, so English speakers can register.
On March 17, complete the inventory of AI use across your team, give each AI assistant its own account and choose between a subscription and usage-based billing. You will leave with a one-page AI policy and a test to check whether an internal tool is still open on the web.
On February 3, start from a risk-free exercise: the Monday operations brief from step two or your volunteers’ task. Write it down exactly as the team performs it, add three concrete checks and test the result on known cases. You will leave with a tested procedure: its instructions, its checks, a column for discrepancies and around ten cases whose answers you already know.
Let’s review your order workflow
In 20 minutes, we’ll identify bottlenecks and ways to save time. No slides.
Free 20-min call · No commitment








































