LLMs: why what they know is just as important as what they can do

In 2026, the world seems obsessed with using AI to automate things. As LLMs have improved, their ability to do tasks that were hard or tedious for humans to do has changed the way many people work (albeit with mixed results according to this MIT report).

This is all well and good, but simply focusing on getting AI to do overlooks what is arguably an even more revolutionary benefit of frontier LLMs: using what they know.

In the last blog post, we spoke about how retailers have to predict human behaviour in every decision they make. The losses they make from getting these predictions wrong are astronomical. Take our fictional retailer Fellside again. As a £250M-revenue retailer, they might lose £5–10M by getting their long-range prediction of next season’s sales wrong, and another £5–10M by setting prices wrong or timing markdowns poorly. Millions more might be lost targeting their marketing to the wrong people, sending products to the wrong locations, or developing products nobody wants. All of these losses occur because they don’t know, and struggle to predict, how their customers will behave.

But LLMs do know how people behave. Utilising the knowledge contained in these LLMs is paramount to retailers fixing these mistakes.

Why do LLMs know how people behave?

It’s a bold claim for me to say, “LLMs know how people behave.” It’s worth clarifying exactly what I mean by this.

No LLM is ever going to predict exactly how an individual person is going to behave. It doesn’t know what Andy Burnham is going to have for breakfast tomorrow or what Michelle Obama is going to do this afternoon. And I’m not claiming that it can predict exactly what you are going to buy when you enter a store.

But what an LLM can predict is how a given persona is likely to behave. It can predict that a middle-aged woman living in Chelsea will probably buy more premium items. It can predict that, in the run-up to Glastonbury Festival, a 22-year-old is probably more likely to buy a cheap tent.

It can do this because LLMs know the likely patterns of human behaviour as a by-product of their training.

In its simplest terms, an LLM is just a model of the probability of the next word given the words already in its “context window”. If you give it the words “the cat is”, it predicts the next word is “black” with a high probability, and “a kleptomaniac” with a low probability.

AI labs build these models by “pre-training” them on basically the entire corpus of what people have written online. All this means is that they take any text they can get their hands on, give the LLM part of it, and ask it to predict what comes next. Then, they nudge the model after each sentence to make it more and more accurate.

This means the model has read thousands of instances of different personas behaving and has learnt how to reproduce them. It has read a Mumsnet thread between Chelsea mums about how it’s better to buy premium products and has been trained to mimic how they’re likely to write. It knows the mums in Chelsea will tend to buy more premium products; the LLM knows something about human behaviour.

On top of this, LLMs know the likely behaviours of personas even if those personas haven’t written online. Take a retired 77-year-old grandmother from the Lake District living on a small pension. People like her likely don’t engage in chat forums very often, nor do they leave Google reviews. But the LLM can still emulate their behaviour.

How? Because the way the LLM has learnt what words are likely to come next is by placing every word on a map based on how similar it is to other words. Because it’s learnt from millions of articles about pensioners living on fixed incomes and struggling with the cost of living, “pensioner” and “fixed income” sit close to each other on the LLM’s internal map. In millions more articles, people have explained that when they’re on a limited income, they hunt for bargains and are more price-sensitive. The words “pensioner” and “price-sensitive” are now close together, so the model predicts the 77-year-old in Keswick is likely to be responsive to promotions, even though she’s never written about her budget struggles anywhere.

The fact that LLMs know things about human behaviour, and can emulate it, is proven in practice. Ashokkumar et al.¹ present the most comprehensive proof of this to date. They take 70 different survey experiments (for example a survey of how likely different people are to get the flu vaccine) performed in real life, then ask LLMs seeded with different personas the same questions. The synthetic, LLM-generated responses had a correlation of 0.90 with the real ones. The LLMs almost perfectly simulated how real humans would respond.

Moreover, Ashokkumar et al. generated these simulations on surveys taken after the model’s training cut-off date. There’s no way the model could have memorised the results beforehand; the only explanation for their accuracy is that LLMs know how different personas tend to behave and can replicate it.

What’s more, these results come from GPT-4. This is the same model that told you that you should walk to the car wash and that the word “strawberry” has two r’s in it. The models have already improved drastically and will likely continue improving.

Why does this matter?

This matters because the knowledge contained within LLMs solves so many of the problems retailers face on a daily basis.

Take Fellside. They’re launching a new trail-running range and need to decide how many of each product to put where. They need to understand who the shoppers are in each area and whether those shoppers will buy the new range. They need to know their customers’ behaviour.

The same problem arises when deciding whether to mark down slow-selling products. If shoppers are price-sensitive, you should mark down; if not, you’re losing revenue by doing so.

Every pricing, marketing and inventory decision is a prediction of how customers will behave in different scenarios. Most of them are currently made from inaccurate demand forecasts, expensive surveys or gut feel (which for many retailers is the most accurate prediction method to date). They know they lose millions, they know it’s not perfect, but until now there’s been no way other than gut feel and demand forecasts to predict human behaviour.

Now there is. The information on how your customers are likely to behave already exists, and is contained within an LLM. Retailers could save millions by getting that information out and using it to their advantage.

Yet, the industry is focused on how they can utilise AI to do things for them. They want to “automate manual processes”, reduce headcount and make handy new tools that save each of their employees time.

At Clinchr, we believe the more fundamental benefit is extracting what LLMs know.