On-premises or cloud AI: Where Do Your Costs Come From? When does pricing use an LLM?
.avif)
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.
Your teams are already using LLMs. Your pricing software may be using them as well—without you knowing where the prices and margins it sends them are coming from.
This article is intended for pricing, sales, and IT departments that are evaluating an AI solution. It compares cloud-based AI and on-premises AI, and then explains the choice Booper made for its agent, Booper Mind.
What an LLM Sees in Your Prices
When pricing software uses a large language model (LLM), it sends it data to work with: product descriptions, prices, and sometimes margins or volumes.
So the question isn't just "What can the model do?" It's also "Where is it running, and who sees what we send it?"
Cloud-based AI processes your data on a model provider's servers via an API. On-premises AI runs the model on the publisher's or client's infrastructure, without involving a third party.
For a retailer, this data is not insignificant. A file containing prices and margins by SKU describes the sales strategy better than any presentation.
Some of the security and privacy professionals surveyed are concerned about inadvertently sharing sensitive information—either publicly or with competitors—through generative AI.
(Cisco, Data Privacy Benchmark Study, 2,600 respondents, April 2025)
The issue, therefore, goes beyond compliance. It concerns the confidentiality of what gives you a competitive advantage, as we explain on our data security and compliance page.
Cloud-based AI or on-premises AI: two architectures
Neither of them is inherently bad. They address different constraints, and the right choice depends on the nature of the data being processed.
Cloud-based AI provides access to the most powerful models without the need to manage infrastructure. It’s pay-as-you-go, and each call sends data outside your network.
On-premises AI requires dedicated computing servers and a team to operate them. In exchange, the data stays on your premises, and costs don't increase with volume.
| Criterion | Cloud-Based AI (API) | Local AI |
|---|---|---|
| Where is the model being filmed? | At the model supplier's | On the publisher's or client's infrastructure |
| Data Sent to a Third Party | Yes, every time I call | No |
| Cost | Variable, billed based on usage | Landline, the server's number |
| Model Power | The latest models | Open-source models, more compact |
| Infrastructure to Manage | None | Computing Servers and Dedicated Team |
| Model Change | Depends on the provider | Can be done without touching the screens |
For occasional use with public data, the cloud remains a convenient option. For repeated processing of an entire catalog of prices and margins, however, the situation changes.
Booper's Choice: AI Hosted by Us
At Booper, this decision was made right from the start of our AI project, which was presented internally in May 2026. Our agent is called Booper Mind. It is called directly by our MPS platform.
The language model is open source and runs on our own computing server, which is equipped with four dedicated graphics cards. No data is sent to an external model provider.
We ruled out using a cloud-based LLM for four reasons: security and the GDPR, the confidentiality of financial and strategic data, independence from a single vendor, and a cost structure that is not usage-based.
The template is running on our server. Switching to a different open-source template does not require any changes to the screens.
Wajdi Tarchi, Chief Information Officer at Booper
This choice has a less obvious consequence: the model is interchangeable. If a more powerful open-source model becomes available, we can switch to it without changing the screens used by the teams.
That's also why our customers see only one name: Booper Mind. The underlying engine is never displayed, since it may change.
Cost Per Use: The Blind Spot in AI Pricing Projects
A cloud-based LLM is billed based on the volume of text processed. Based on a test of a few hundred products, the bill appears to be negligible.
Retail pricing doesn't operate on this scale. A customer database we analyzed in May 2026 contained 239,160 products. These analyses are repeated every time there is a change in the product line, pricing, or competition.
With on-premises AI, the cost is limited to the server, which is a fixed expense. Running another analysis on an entire bill of materials costs nothing extra, which changes the way teams use it: they feel free to experiment.
Some companies have abandoned, delayed, or scaled back an AI project out of fear that sensitive data might leave their premises or jurisdiction.
(Cloudian, Enterprise AI Infrastructure Survey, 203 IT decision-makers, 2026)
Our belief: When it comes to pricing, an AI tool that you hesitate to use again because it costs money every time you call it ends up being underutilized. The AI’s business model matters just as much as its power.
One agent, with rights limited to each use
Hosting the model on your own server isn't enough. An AI agent can perform actions using tools: read a file, query a database, send an email. The more tools it has, the more safeguards are needed.
Booper Mind therefore uses separate profiles. Each user is granted only the permissions they need—and nothing more—along with their own settings in MPS.
Automated Processing
No tools, no external access. A structured response, controlled by MPS. Example: AI-driven chain management.
Analyses and Reports
Permitted tools: viewing exports, drafting, and sending via email. Example: a selection analysis report.
Same base
Same server, same local model. A new use case in MPS reuses the existing one.
Source: Booper, AI Architecture, November 2026.
This principle aligns with our stance onagentic pricing: the key issue isn't the sophistication of the algorithm, but rather what we delegate to it.
What Local AI Doesn't Solve
A local AI is still an LLM. It can be confidently wrong, and it doesn't know your data until you provide it.
These limitations are the same as those described in our article on pricing errors in generative AI. Hosting addresses the issue of privacy, not that of reliability.
Reliability comes from elsewhere: scheduled checks on each response, a clear explanation, and human validation before any action is taken. That is the essence ofexplainable AI.
Our initial tests confirmed it: a local model can sometimes make mistakes, just like any LLM. So we added automatic checks for each response, along with a recovery mechanism when a response is incomplete or corrupted.
No proposal is implemented in MPS without the user's consent. These rules apply to all future use of Booper Mind.
Wajdi Tarchi, Chief Information Officer at Booper
We applied it to our first production use case—managing product matching and assortment chaining —where the AI suggests groups that the user can accept or reject.
Using a general-purpose LLM outside of this context raises other issues, which we detail in our comparison of general-purpose AI versus specialized pricing solutions.
Three Questions to Ask Your Publisher
Before entrusting your pricing data to software that incorporates an LLM, three questions are all it takes to clarify the architecture.
- Where is the language model hosted, and is my data being sent to a third-party provider?
- Does the cost of AI depend on the number of calls or products processed?
- What controls the model's responses before they are incorporated into my prices?
If you're unsure about any of these three points, it's worth digging deeper. To learn more about how the system works as a whole, read our article on how an AI pricing engine works.
Are you looking for AI that won't drive up your prices?
Our experts will demonstrate Booper Mind using your own data, complete with its controls and human validation.
Let's discuss your AI projectFrequently Asked Questions About On-Premises AI for Pricing
On-premises AI is an artificial intelligence model—often a large language model—that runs on the software vendor’s or customer’s infrastructure, rather than on the servers of a model provider accessible via API.
The data sent to the model therefore remains within this scope. The cost depends on the computing server, not on the number of calls. On the other hand, dedicated graphics cards and a team capable of utilizing them are required.
Cloud-based AI is accessed via an API: each request is sent to the model provider, which processes it on its servers and bills for it on a pay-as-you-go basis. It provides access to the most powerful models without the need for infrastructure.
Local AI runs the model at the publisher's or the client's site. No data is sent to a third party, and the cost remains fixed. The models used are generally open source and more compact.
It depends on its architecture, which is why it's a good idea to ask the software vendor. Software that calls upon an LLM in the cloud sends the provider the data needed for each processing step: product descriptions, prices, and sometimes margins or volumes.
At Booper, the Booper Mind agent uses an open-source language model hosted on Booper's server. No data is sent to an external model provider.
Booper Mind is Booper's artificial intelligence agent, integrated into the MPS pricing platform. It is based on an open-source language model hosted by Booper, which can be swapped out without modifying the screens.
Its first application in production is AI-driven product line management: it suggests, supplements, and corrects product groups within the same product line. Each suggestion is verified by the program, explained in French, and then accepted or rejected by the user.
Open-source models that you can host yourself are generally more compact than the latest models from major providers. For a well-defined task, such as recognizing the same brand and format in labels written in different ways, this difference is negligible.
The system's effectiveness stems primarily from its framework: well-prepared data, precise instructions, automatic checks on each response, and human validation.
The cost depends primarily on the AI's business model. Cloud-based AI is billed based on the volume of text processed: for a catalog containing several hundred thousand items that is analyzed regularly, the bill increases with usage.
Local AI integrated into the software represents a fixed cost for the software vendor, with no per-use billing. At Booper, Booper Mind is included in the platform: running an analysis again costs nothing extra.
Yes. The hosting service ensures privacy, not reliability. A local model is still an LLM: it can be confidently wrong, omit a product, or produce an incomplete response.
That is why every Booper Mind response is checked by the program: unknown products, omissions, damaged or cut-off responses. If there is a problem, the response is corrected or the analysis is redone in small batches, and then the user approves it.
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.
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.