Generative AI gets prices wrong: and on the data
.avif)
A general-purpose language model predicts the most plausible word, not the actual price; by default, it has no access to an up-to-date catalog or any uncertainty signal.
Price is a particularly telling example because it can be verified immediately. The real issue is never the model's power, but rather the proprietary data to which it is—or is not—connected.
Ask a general-purpose language model the price of a product, and it will respond confidently, in a well-constructed sentence, without any apparent hesitation. The figure may be correct. It may also be completely wrong, without anything in the tone of the response giving that away.
This isn't a one-off bug. It's a direct consequence of how these models work: they predict the most plausible sequence of words; they don't verify anything. When it comes to a price—a specific, time-sensitive piece of data that’s constantly changing—this limitation becomes immediately apparent. And it reveals a problem that goes beyond the model itself: the lack, upstream, of proprietary data and business rules to which it can be connected.
A guaranteed response, a fake price
This scenario has played out thousands of times since chatbots became part of everyday life: a buyer, a category manager, or someone simply curious asks a general-purpose language model a question about price. The answer comes in a matter of seconds, phrased with the same confidence as if it were an exact answer. The figure seems credible. But it isn’t always.
What is striking is not just the error—every system makes mistakes sometimes—but the absence of any signal that would allow us to detect it. A model that hesitates, that indicates a margin of uncertainty, that says, “I don’t know,” would be easier to correct than a model that asserts a made-up price with the same ease as an exact price.
This behavior is not a teething problem that will disappear with the next version of the model. It is a direct and lasting consequence of what these systems actually do.
Predicting is not the same as verifying
A general-purpose language model (whether it’s ChatGPT or an equivalent competitor) operates on a principle that’s simple to state but less straightforward to intuitively grasp: for every word it generates, it calculates the statistically most likely sequence of words, based on everything it has seen during its training. By default, it does not consult any real-time database. It “knows” nothing in the sense that a spreadsheet or ERP system knows what a cell contains.
For creative writing or rephrasing, this mechanism produces excellent results: the most plausible answer is, in most cases, a good answer. For a price (a specific, date-sensitive figure that has only one correct value at a given moment), the mechanism changes in nature. The model does not seek the actual price: it seeks the number that, statistically, most closely resembles a plausible price for this type of product, based on the patterns it has encountered in its training corpus, which is frozen at a cutoff date that never corresponds to the present moment.
- There is no native access to the actual inventory; the model does not query the point-of-sale system, ERP, or e-commerce site of the relevant retailer.
- A corpus that is frozen in time—even information that was correct at the time of training may be outdated by the time the question is asked.
- There is no inherent concept of uncertainty; in its formulation, the model does not distinguish between a firmly established fact and a statistical extrapolation.
Why the price is a particularly telling example
Many errors made by generative AI go unnoticed: an approximate date, a slightly outdated statistic, a nuance lost in a summary. The price, however, can be verified immediately. All you have to do is visit the retailer’s website or compare it to the receipt to see the discrepancy. That’s what makes the problem visible—it’s no more common than elsewhere, just easier to measure.
This visibility is becoming increasingly important as consumers get into the habit of asking a chatbot for advice before making a purchase rather than comparing prices themselves.
With the use of generative AI applications such as ChatGPT and Google Gemini by French consumers having grown over the past two years, a growing proportion of them now rely on these tools to search for a product or a good deal (IBM Institute for Business Value & NRF, “Own the Agentic Commerce Experience” study, January 7, 2026).
The more consumers rely on a general-purpose conversational assistant to handle part of their price research, the more the gap between the stated price and the actual price becomes a tangible business problem—both for the retailer whose product is misrepresented and for consumer trust in the tool itself.
What an isolated model lacks to be reliable in terms of price
The difference isn't due to the model's size or version. It's due to what surrounds it—or doesn't surround it.
| Component | Isolated Language Model | Model connected to proprietary data |
|---|---|---|
| Source of the price | Statistical pattern of the training corpus | Up-to-date catalog, ERP, or point-of-sale system |
| Freshness | Fixed as of the model's cutoff date | Real-time or near real-time |
| Business Rules | None; no minimum margin requirement; no product line consistency | Integrated and verifiable |
| Uncertainty signal | Even when he's not there, his tone remains confident—even when he's wrong | Displayed confidence score |
To put it this way: it is never the model alone that makes a price estimate reliable; it is what the model is connected to, and the rigor with which that connection is established.
The Real Issue: Data, Not Model Power
When faced with an incorrect answer, the natural instinct is to look for a “smarter” model. This is rarely the right approach. A more powerful model, without a connection to a source of truth, will produce an incorrect answer that is phrased even more convincingly, which exacerbates the problem rather than solving it.
A seminal academic study on the factuality of language models illustrates this point well, using an exercise that is far removed from pricing but reveals the same mechanism: the generation of long-form biographies.
This is the factual accuracy score measured for ChatGPT on the generation of long-form biographies, according to the FActScore method, which breaks down each response into basic facts and verifies each one against a reliable source. More than four out of ten claims, which are not grounded in verified data, do not hold up to verification (FActScore, Min et al., EMNLP 2023).
The figure does not pertain to prices, but the mechanism is applicable: an isolated language model, no matter how advanced it may be, produces a mix of correct facts and plausible extrapolations, without drawing a clear line between the two. In the context of a biography, the issue is often merely anecdotal. But when it comes to a price shown to a customer or used to make a business decision, it is no longer just an anecdote.
What This Reveals About the State of Data in Companies
The issue goes far beyond language models. It touches on a longer-standing and more structural problem: most organizations do not have data that is clean, structured, and accessible enough for AI (whether generative or not) to rely on it reliably.
AI projects not supported by “AI-ready” data should be abandoned by the end of 2026, according to Gartner, which also notes that 63% of organizations do not have—or do not know if they have—data management practices suitable for AI (survey of 248 data leaders, Q3 2024) (Gartner, press release, February 26, 2025).
A model that gets a price wrong is therefore not just a modeling issue. It is often the most visible symptom of an underlying problem: the lack, upstream, of a reliable product reference system, a clean price history, and formalized business rules.
At Booper: Connecting AI to the truth, not to its memory
The AI Center: A Solution Rooted in the Retailer's Actual Data
Booper’s AI Center (the “BOOPER AI Assistant”) provides natural-language responses regarding pricing, competitive matching, promotions, and monitoring. The difference between it and a general-purpose chatbot isn’t that it uses a larger model; rather, it’s that its responses are based on the retailer’s actual data, not on the general knowledge of a model trained on web data as of a specific point in time.
The GENIUS Link module provides a concrete example of what it means to “properly connect” an AI component to the real world: using NLP (natural language processing), it matches a retailer’s products with tracked competitor items—even when the product names differ—and displays a confidence score for each match, rather than a binary answer that would mask the uncertainty. This is the exact opposite of the behavior of a standalone language model: here, uncertainty is shown, not masked.
Four Things to Consider Before Trusting a Price Provided by AI
- Check the source, not just the number. Explicitly ask where the data comes from. A system connected to a single source of truth can provide an answer; a standalone model can generally only describe its own behavior.
- Never publish a price without human verification. A price generated by generative AI, no matter how convincing it may be, remains a hypothesis that must be verified against the actual source before any commercial or client use.
- Be wary of a lack of hesitation. A system that displays a confidence score or acknowledges its limitations is more reliable—not less—than a system that always responds with the same level of confidence.
- Check the freshness of the data. Retail prices are constantly changing. An answer that was correct yesterday may be wrong today if the system isn't connected to an up-to-date source.
Checklist before trusting an AI-generated price estimate:
- Is the system connected to a live data source, or is it responding solely from its training data?
- Does the answer indicate a level of confidence, or is it just a number presented as a certainty?
- Is the underlying data up to date, or does it date back to a previous training session?
- Are there any business rules (margin thresholds, product line consistency) that are applied before the price is displayed?
- Does a human verify an AI-generated price before it reaches a customer or is used in a business decision?
Is your data ready to power a reliable AI pricing system? Spend 30 minutes with our team to assess where your data stands today and what needs to be fixed before you integrate AI into it. → Let’s schedule a meeting.
FAQ
Do you still have questions? Here are the answers to the most frequently asked questions on this topic.
To take it a step further:AI draws conclusions based on the data it is given, not on what the company knows.
A general-purpose language model predicts the statistically most likely word, not the actual current price. Without a connection to an up-to-date data source—as explained in our article on how a real AI pricing engine works—it generates a plausible figure based on patterns observed during its training, which may resemble a real price without actually being one, particularly because its dataset has a cutoff date and never reflects the present moment.
This mechanism is not a temporary bug that will disappear in the next version: it stems directly from the very principle of statistical prediction. For a creative text, the most plausible answer is often the correct one; for a price, which has only one correct value at any given moment, this logic changes in nature.
Three structural shortcomings account for the discrepancy, as detailed in the comparative table in this article: no native access to the actual catalog, a corpus that is frozen in time, and no native conceptof uncertainty; in its formulation, the model does not distinguish between a firmly established fact and an extrapolation.
For a retailer, the stakes go beyond a mere technical detail: with a 90% increase over two years in the use of chatbots by French consumers to check prices, a made-up figure presented with confidence can directly skew perceptions of an offer.
No, not by default. A standalone language model has no native access to a product catalog, an ERP system, or a point-of-sale system. It must be explicitly connected to this data in order to provide accurate responses; that’s the difference between a generic AI assistant and what a pricing tool connected to a retailer’s actual product database really is.
The comparison table in this article directly contrasts these two configurations: a standalone model relies on a statistical pattern from the training corpus, frozen as of its cutoff date, without business rules or a confidence score; a connected model draws from the catalog, the ERP system, or the up-to-date cash register, with verifiable rules and a displayed confidence score.
This lack of connection is not unique to pricing: it more broadly affects the way generative models handle any precise, dated factual data, as demonstrated by the FActScore method discussed later in this article in relation to a different exercise that nevertheless reveals the same mechanism.
In practical terms, connecting an AI to a retailer’s actual product database (rather than asking it to guess) is the only way to turn a plausible answer into a reliable one that can be used to inform a business decision.
FactScore is an academic method that breaks down a text generated by AI into basic facts and measures the proportion of those facts that can actually be verified in a reliable source. A landmark study (EMNLP 2023) found that ChatGPT scored 58% on biography generation; a significant portion of the claims it produced—which were not grounded in a verified database—did not stand up to verification.
This figure does not pertain to prices, but the mechanism is applicable: an isolated language model, no matter how advanced it may be, produces a mix of correct facts and plausible extrapolations, without drawing a clear line between the two—a limitation documented in our article on data quality, which represents a glass ceiling for AI pricing.
In a biography, the stakes are often merely anecdotal. But when it comes to a price quoted to a customer or used to make a business decision, that’s no longer the case: more than four out of ten unverified claims, according to the same study, illustrate the magnitude of the risk if there is no reliable source to correct the model.
This finding debunks the natural tendency to look for a “smarter model” when faced with an error: a more powerful model, without a connection to a source of truth, formulates its incorrect answer in an even more convincing way, which exacerbates the problem rather than solving it.
Because a price can be verified immediately and objectively, unlike a vague answer or an opinion. The discrepancy between the price provided by generative AI and the actual price is immediately apparent within seconds, highlighting a broader issue that often goes unnoticed elsewhere: the lack of proprietary data and business rules behind the response—the subject of our article on what truly constitutes artificial intelligence in pricing.
Many errors made by generative AI go unnoticed—an approximate date, a slightly outdated statistic, a nuance lost in a summary. The price, however, can be verified by visiting the retailer’s website or comparing it to the receipt, as this article points out.
This visibility is becoming increasingly important as consumers grow accustomed to consulting a chatbot before making a purchase rather than comparing prices themselves; according to IBM and the NRF, the use of these tools by French consumers has grown by 90% over the past two years.
For a retailer, this discrepancy is more than just a technical curiosity: the more consumers rely on a general-purpose assistant to compare prices, the more a poorly represented product becomes a tangible business problem—one that goes beyond mere trust in the tool.
By connecting it to a source of truth—a reliable product repository (see how to ensure reliable product matching), an up-to-date catalog, and business rules—rather than by seeking a more powerful language model. The problem is almost never the size of the model; it’s the lack of reliable proprietary data to connect it to.
This article suggests four practical steps to take before trusting a price generated by AI: check the source—not just the number—never publish a price without human verification, be wary of a lack of hesitation, and question the recency of the underlying data.
A reliable system must also display a confidence score rather than a binary answer that would mask uncertainty; this is exactly what Booper’s GENIUS Link module does for the comparison of competing products mentioned above.
This correction is as much structural as it is individual: according to Gartner, 60% of AI projects not supported by “AI-ready” data are expected to be abandoned by the end of 2026, which reframes the issue of made-up prices as a much broader fundamental problem than just the model itself.
No. Booper’s AI Center responds in natural language but draws on the retailer’s actual data (pricing, competitive matching, promotions, monitoring), not on the general knowledge base of a model trained on the web. It is a conversational assistant connected to a single source of truth, not an isolated model that makes extrapolations.
The difference does not stem from a larger model; rather, it stems from the fact that its results are based on the retailer’s actual data rather than on a statistical pattern frozen at a cutoff date—which is precisely the discrepancy described in the comparative table in this article.
The GENIUS Link module provides a concrete example of this approach: it uses NLP to match a retailer’s products with the tracked competitor SKUs, even when the product names differ, displaying a confidence score for each match rather than a response that would mask the uncertainty.
For a retailer, this architecture changes the nature of the trust placed in the tool: uncertainty is revealed, not masked, which ties into the broader question of where to draw the line between AI that makes decisions and AI that executes them, and allows a category manager to know when a recommendation can be applied directly and when it requires verification.
Yes, just like any statistical model: an elasticity model can be wrong, particularly when dealing with noisy or incomplete data, or in the face of unprecedented situations (stockouts, new competitors, or exceptional events).
This is precisely why the process described above never stops at a recommendation: business rules govern what the model can propose, and the simulation phase makes it possible to quantify the impact on revenue, margin, and inventory before the solution goes live.
The final step—validate and execute—always allows a category manager to validate, adjust, or reject the recommendation, with full traceability of each decision for audit purposes; it is not a “black box” that makes decisions on its own, contrary to what the category might sometimes suggest.
This architecture—which involves a recommendation followed by human validation —is what makes the tool deployable in stores: a system that ran without safeguards or explanations would directly expose the margin and price image to model error.
To learn more about this topic: AI and pricing, what truly falls under the umbrella of artificial intelligence (pillar), AI that decides vs. AI that executes, and data quality as the glass ceiling of AI pricing. To assess the reliability of your pricing data, check out our price diagnostic.
The model is the easy part. A recommendation engine can be up and running in a few months; what takes years is the knowledge base, competitive matching, the rule set, explainability, and compliance. According to Exclaimer (2025), 71% of in-house developments are eventually abandoned, and 83% in regulated sectors: the breaking point is almost never deployment—it’s maintenance.
The hidden cost is a cost of continuity: an in-house tool relies on two or three people, and if they leave, the asset becomes a liability. The build remains useful within a narrow and stable scope, as a complement to a solution rather than a replacement for it. The deciding factor isn’t the team’s expertise; it’s their ability to last ten years.
Data and context are two distinct assets: data describes what happened, while context explains what we are allowed to do and what would be absurd. This context is implicit by nature: it has never been written down because there was never a need to do so as long as decisions were made by humans who shared it.
Four categories cover the essentials: price benchmarks, relational constraints, unwritten rules, and commercial objectives by category. Formalizing them provides lasting value, because the context outlasts the models: algorithms are replaced every two or three years, but a written context remains a company asset.
A business case for a pricing solution combines two types of return on investment that are completely different. Friction ROI refers to the hours saved: easy to measure, but capped. Decision ROI refers to the margin points gained from better-set prices: difficult to measure, and unlimited.
Confusing the two explains the gap identified by IBM in 2025: 66% of executives report productivity gains from AI, but only about one in five has met their return on investment (ROI) goals. The ROI of a decision isn’t measured by comparing “before” and “after,” but rather against a control group; otherwise, seasonal factors, inflation, or a competitor’s actions will be credited to the tool. The test to apply to a business case: if you remove all the “time saved” lines, is the project still profitable?