Recently, I came across a post where someone had asked one of the popular AI chatbots to solve a classic optimization problem. While problem involved some basic tradeoffs and constraints, it wasn’t a complex problem and it had a fairly clear correct answer. Of course, the AI chatbot produced an answer that sounded very authoritative and highly credible.
The problem was that the answer wasn’t even close to being right!
Now, it doesn’t really matter which AI chatbot was being used, because they all would’ve done about the same thing. The simple truth is that the Large Language Models behind the AI chatbots that are getting so much attention these days just aren’t built for doing that kind of work. They aren’t mathematical engines or optimization solvers. They’re not doing calculus, running regressions, or executing multiple tradeoff simulations under the hood.
Yes, these Large Language Models (LLMs) are indeed predictive models. But they’re designed to predict language…words…and not mathematical results. And as you might imagine, this particular distinction is pretty darned important when we’re talking about something as numbers-intensive as pricing!
When you ask an LLM a question that involves optimization or advanced math, it doesn’t actually “solve” the problem in any traditional sense. It predicts what a smart-sounding answer to that kind of question might look like, based on everything it’s read before. Rather than trying to compute the most accurate number, it’s trying to predict the most likely words a response might contain.
That’s why even wildly incorrect answers will often sound highly plausible, logical, credible, and authoritative. After all, the LLM is designed to predict the most logical and credible words. It’s not that these LLM-based chatbots are broken in some way. No, they’re actually working exactly as intended. It’s just not the right kind of AI tool for that kind of problem.
And this is where so much of the current confusion comes in…
As we discuss in the Making Sense of Artificial Intelligence for B2B Pricing training session, the term “AI” is being used to describe a wide range of technologies that are very different. A Large Language Model is a linguistics engine designed to understand, manipulate, and generate language. In sharp contrast, a price optimization model is mathematical and designed to evaluate tradeoffs, run scenarios, and calculate the best answers to satisfy a defined set of objectives.
While both are “AI” in the broadest sense, they are designed to do very different things, and expecting one to do the other’s job is a recipe for disaster.
This doesn’t mean that LLMs have no place in pricing. They can be very useful as translators or facilitators. In fact, a well-configured LLM-based workflow or agent can call a math-based optimization model, feed it the necessary parameters and inputs, move the results into other systems, and even explain the results in natural language for decision-makers and end-users. That’s really powerful stuff!
But the key phrase here is “well-configured.” An LLM won’t automatically know which optimization engine or analytical model to call or when to call it. You’ll have to tell it through the workflow you design. Without that direction, the LLM will just try to sound like it knows the correct answer. And as everyone knows…or should know…sounding right is not actually the same as being right.
While the hype around Generative AI and LLMs makes it seem like our analytical problems are about to vanish, the truth is that a lot of what we do in pricing still depends on quantitative logic. When we’re calculating elasticity curves, running “what-if” simulations, balancing margins and unit volumes, or optimizing prices across multiple segments, we need models that compute.
We need math-driven AI, not word-driven AI.
So the smart move is to use each type of AI for what it does best. Use the language models to synthesize insights from text, summarize sales feedback, automate reporting, manage repetitive workflows, or help communicate recommendations in plain English. Use the mathematical models to do the heavy lifting around robust analysis, simulation, and optimization.
And wherever appropriate, of course we’ll connect them together. We just don’t want to confuse them. Because when we mix up “talking about the problem” and “sounding right” with actually “solving the problem” and “being right,” things can get ugly or go sideways really fast.
Making Sense of AI for B2B Pricing
There's no doubt that artificial intelligence will transform nearly every aspect of sales and revenue operations. How do we cut through the hype and harness AI to generate improved pricing results and business outcomes?
Dynamic Price Segmentation
How do you build a price segmentation model that aligns to deal-by-deal price sensitivities while still being manageable? In this session, we illustrate why dynamic, attribute-based segmentation is best practice.
Working With "Bad" Pricing Data
In B2B, there are so many moving parts that always having a perfectly accurate and complete dataset just isn't very realistic. So we've got to figure out how to do the best we can with the data we've got.
