Entrusting the chaining of its ranges to an LLM: what we've learned
.avif)
An LLM can match differently worded descriptions and suggest groups of products from the same line. It can also be confidently wrong.
Booper's AI-driven chaining system combines AI, programmed controls that correct its errors, and human validation. No group is created without the user's consent.
Grouping together the formats and variants within a single product line is painstaking work. With a catalog containing hundreds of thousands of SKUs, it’s a task that’s never truly finished.
We entrusted this task to a large language model, within a strict framework. This article describes what we learned: what the model does well, where it goes wrong, and why the final decision remains a human one.
Chaining a scale: a labor-intensive task
Range linking groups together the different sizes and variants of the same product so that their prices change in tandem. A yogurt sold in a 4-pack and an 8-pack, or three flavors of the same dessert cream, form a group.
It should not be confused with the chaining that occurs in the sense of continuity over time, which links an old reference to the one that replaces it. Both serve to ensure consistency across the product line.
Today, these groups are often created manually or based simply on similarities in product descriptions. With a catalog containing several hundred thousand items, the work is time-consuming and never-ending.
products and 4,126 existing groups in a retail client’s product catalog analyzed by Booper. This is the actual scale of a chaining project.
(Booper, analysis of a client’s product catalog, May 2026)
The consequences are evident on the shelves: missing sizes, incomplete sets, and inconsistent prices within the same product line. The large size ends up costing more per kilogram than the small one.
Why an LLM, and not a rule on wording?
Let's look at two actual product names from our test of dessert creams: "DANETTE CHOCOLATE 4X125G" and "DANET.CHOCO.HAZELNUT FLAVOR 4X125G."
A text similarity rule runs into trouble with abbreviations. A large language model recognizes the same brand and format, even when they are written differently.
That's exactly what an LLM does: read, match, and classify labels. It also understands instructions in French, such as “don’t mix brands.”
But it has three limitations. It can be confidently wrong. It doesn't know your data unless you give it to it. And it doesn't act—it just generates text.
These limitations explain why we didn't integrate a chatbot with the catalog. We built a process.
An agent-based process, not a chatbot
A chatbot answers a question. An agent performs a sequence of actions using tools to achieve a goal. An agent-based process goes a step further: it combines AI-driven steps with programmed steps, with checks in between.
The AI-powered workflow management, integrated into our MPS platform, falls into this third category. It is triggered by a button, not by a conversation.
Booper Mind offers
Reads the items in a bill of materials and prompts the user to create, add to, or correct groups, along with their master item.
MPS Control
Check each answer, correct any errors or rerun the analysis if necessary, and then rank the options.
The user decides
Reads the rationale for each group, then accepts, rejects, or changes the leader. Nothing is created without his approval.
Source: Booper, AI-driven chaining control, November 2026.
The AI, Booper Mind, sees only what it needs: the product ID, EAN code, description, brand, and price of each product. It has no tools, no access to the web, and no access to files.
This architectural choice—and the reason why the model runs on our servers—is explained in detail in our article onon-premises vs. cloud AI.
What the tests have taught us: AI makes mistakes, but the program corrects them
Early versions relied on the model's response. Testing quickly revealed the limitations of this approach.
The model sometimes listed a product that didn't exist in the bill of materials. It identified a master product that didn't belong to the group. It omitted part numbers. For large bills of materials, its response would arrive damaged or cut off.
The program checks each response for unknown products, master items outside the group, omissions, damaged responses, or cut-off responses. A damaged response is repaired. A cut-off response is recreated by brand, in small batches.
Quality control and remediation steps were added after the tests. The system is also covered by 82 automated tests, which run without launching the application.
Wajdi Tarchi, Chief Information Officer at Booper
Two steps were therefore added after the tests. The first checks each response. The second corrects it or restarts the analysis by brand, in small batches.
Bills of materials that are too large are also automatically broken down into sub-BOMs. An LLM works best when dealing with a scope it can grasp at a glance.
These errors are not unique to our model. They are typical of generative AI, as we explain in our article on pricing errors in generative AI.
Classify, explain, get approval
At the end of the analysis, the proposals are categorized into three groups: new groups, modifications to existing groups, and cases requiring further review.
AI doesn't just create groups. It supplements and corrects existing ones, and selects a lead product for each group, which serves as a price benchmark.
Each group is assigned a reason, in French. For example: “Same brand, same packaging, prices aligned.” A tab also lists the products that have not been assigned to a group, along with the reason.
The user accepts, rejects, or modifies the master product. No group is created in MPS without the user's consent. The options and instructions for each analysis are tracked.
This requirement for an explanation aligns with our position onthe explainability of price recommendations: a proposal that is not understood should not be implemented.
A few seconds to a few minutes per bill of materials
The most obvious benefit is time savings. Based on our measurements, a bill of materials consisting of 9 items is processed in 7 seconds.
to analyze a bill of materials consisting of 103 products and propose product groups, compared to 7 seconds for 9 products, on the Booper server.
(Booper, measurements of AI-driven sequencing, 2026)
The duration increases with the size of the bill of materials, which is why larger ones are automatically split up. The analysis can be monitored in real time and stopped at any time.
Since the model is running on our server, running another analysis with different parameters doesn't cost anything extra. Teams can test several instructions on the same bill of materials before making a decision.
AI-driven control complements the automatic rule-based chaining and the product matching and chaining features of our offering; it does not replace them.
Our position: AI makes suggestions, the program monitors, and humans make the decisions
AI creates, supplements, and corrects the groups. The program checks and corrects its answers. The user retains the final decision, group by group.
These rules apply to all future use of Booper Mind in MPS: human validation, AI without tools when they aren't needed, verification of each response, maximum duration, the option to stop, and traceability.
Wajdi Tarchi, Chief Information Officer at Booper
This division of roles isn't just a matter of being cautious under the circumstances. It's what makes it possible to entrust a massive amount of work to AI without losing control over costs, as we explain in our article onAI that decides and AI that executes.
The chain control is our first AI screen. Future screens will use the same foundation and the same rules.
Do you want to improve the reliability of your product lines more quickly?
Our experts will launch the AI-driven dependency tracking feature on one of your bills of materials and show you the suggested groups, along with explanations and verification.
Let's discuss your product linesFrequently Asked Questions About AI-Powered Chaining
Product line linking involves grouping together the different sizes and variants of the same product—for example, multiple package sizes or flavors of the same brand—so that their prices change in tandem.
Each group has a master product, which serves as a reference. When its price changes, the other products in the group follow suit according to established rules. Price linking prevents price inconsistencies within a product line, such as a large size costing more per unit than the small size.
Yes. A large language model can match text that is phrased differently and recognize the same brand and format, even when abbreviations are used.
But it can make mistakes with confidence: mention a nonexistent product, omit a reference, or choose a master from outside the group. In the AI-driven management of Booper’s chaining, each response is therefore checked by the program, corrected if necessary, and then submitted to the user, who accepts or rejects each group.
This is a screen from Booper's MPS platform that tasks Booper Mind, Booper's AI agent, with creating, completing, and correcting product groups within a product line, bill of materials by bill of materials.
The user selects the bill of materials and can add instructions in French. The suggestions are sorted, explained, and verified. The user accepts, rejects, or modifies the master product, then exports the result.
A chatbot answers a question with text. An agent is given a goal and performs a series of actions using tools—such as reading a file or sending an email—until the goal is achieved.
An agent-based process chains together AI-driven steps and programmed steps, with checks between them. It is triggered by a button, a schedule, or an event. AI-driven control of the chaining falls into this third category.
Based on Booper's measurements, a bill of materials with 9 items is analyzed in 7 seconds, and one with 103 items in 53 seconds. The time required increases with the number of items.
The largest bills of materials are automatically broken down into sub-BOMs. The user can track progress in real time and can stop the analysis at any time. This process takes anywhere from a few seconds to a few minutes per bill of materials.
No, it complements it. Rules remain useful when the logic is simple and consistent, such as a standard EAN code or a well-maintained naming convention.
AI adds value when the labels are inconsistent, when certain formats have been omitted, or when existing groups are incomplete. It makes suggestions, the program verifies them, and the user retains control over what is applied.
Each proposed group provides, in French, a list of its products and its master product, along with the reason for its proposal. The user can accept it, reject it, or change the master.
No group is created in MPS without the user's consent. Products that have not been assigned to a group are listed along with the reason, and the options and instructions for each analysis are retained, making it possible to trace each decision.
An LLM integrated into pricing software processes sensitive data: prices, margins, and volumes. Where the model runs determines who has access to it.
On-premises AI keeps the data at the publisher's or the client's location and keeps costs fixed. It does not eliminate the need for checks on each response or for human validation.
A market signal describes what competitors are doing. A profitable decision tells you what you need to do, and how it will affect sales, margins, and inventory. The two are not part of the same tool.
The decision-making process relies on five key components that external data does not provide: sales forecasts, price elasticity, cannibalization, business rules, and impact simulation. A signal infrastructure feeds these components; it does not replace them.
When choosing sales forecasting software, there are five key points to check before conducting any tests: demand hidden by stockouts, the forecast granularity, an explanation of each figure, the relationship with prices and promotions, and an honest assessment of accuracy.
A demonstration doesn't prove anything on these points. Only a test using your own data, compared to a simple baseline, will show whether the forecast truly improves your decisions.