ChatGPT, Claude and Gemini can read your retail data. That doesn’t mean they understand your business

Key takeaways:

Picture this. A Head of Insight or IT tells the board that connecting a general-purpose AI chatbot straight to their data warehouse could cut weeks of analyst time down to minutes. And to a point, they’d be right. The model can inspect tables, write queries and summarise performance in seconds.

But being able to read retail data isn’t the same as understanding a retail business. The biggest risk isn’t an answer that’s obviously wrong. It’s one that sounds intelligent but is based on the wrong version of the truth.

Your retail data doesn’t explain itself.

A general-purpose AI agent doesn’t automatically know what your organisation means by net sales, whether returns are included, or when a customer counts as lapsed. It hasn’t sat in your trading meetings, and it hasn’t learned which definitions finance trusts.

An experienced analyst learns these distinctions over time, building up an understanding of the business through documentation and previous decisions. ChatGPT, Claude and Gemini haven’t had that experience, so the model works out how your business operates from table names and whatever documentation is lying around. It’ll still give you an answer. But that doesn’t mean it’s the right one.

The mistake is treating the agent as the whole solution. A stronger model can’t recover context that was never provided and can’t tell whether a fall in sales is a serious issue or an expected seasonal dip unless that logic exists somewhere for it to use. The question isn’t which chatbot sits on top of the data. It’s what intelligence sits beneath it.

The most dangerous answer is the believable one.

Consider a familiar retail question: why is margin down in womenswear? A general-purpose agent might identify the products with the largest decline and return a professional-looking chart. The answer could be fast. It could also be wrong – perhaps using cash margin when the board expected margin percentage, or leaving closed stores in the comparison.

Nothing about the response would necessarily look broken, which is exactly what makes it dangerous. A commercially plausible answer can survive long enough to influence a decision, leading to an unnecessary price change or a genuine trading issue being missed. Once a visibly wrong answer reaches the boardroom, trust in the tool disappears fast, and every questionable answer afterwards has to be checked, adding another layer of noise rather than removing the bottleneck.

A semantic layers gives AI the language of your business.

So how can retailers mitigate against this? A semantic layer provides AI with your business’s language, sitting between the raw data and the AI and translating technical structures into business meaning. It defines what net sales means, how products roll into categories, and captures the language people actually use, so the system can work out whether one person’s lost customers and another’s lapsed customers mean the same thing in that business, and which agreed definition should apply.

Without a sematic layer, the agent interprets the business for itself and confidence breaks down. With one, teams can ask questions in their own language while getting answers grounded in shared definitions, and that consistency is the foundation of trust.

This isn’t just our view. Microsoft recommends giving agents clear business terminology, focused data schemas and verified answers, since some context can’t be understood from the underlying data alone. Snowflake makes a similar point: when definitions don’t match how the business operates, the problem isn’t necessarily the model; it’s the context it’s been given.

But a semantic layer still isn’t the whole answer. It can tell the AI how margin is calculated, but it doesn’t teach it how an experienced retailer investigates a decline, examining product mix, markdowns or competitor pricing. A decline isn’t automatically a problem either: perhaps the retailer has deliberately invested in entry-level prices, or lower margin is simply the cost of clearing seasonal stock. The number alone won’t tell you what to do about it. Retail expertise is what turns it into a useful decision.

Retail expertise has to be designed into the system.

Think of an AI agent as a highly capable new employee. You wouldn’t give someone access to every company system on day one and ask them to set next week’s prices; you’d train them first. AI needs the same support: access to the signals that shape a decision, depending on the question, which might include:

• Price elasticity
• Customer value
• Promotional response
• Stock position

This is the difference between placing a chatbot on a warehouse and building a retail intelligence capability. One retrieves information. The other connects it to a commercial decision, understanding what to examine when category sales decline and when a movement is a strategic risk rather than normal trading noise. AI can do much of the analytical work and present the available evidence, while the retailer applies experience and stays accountable for the decision.

More answers aren’t the prize.

The model alone isn’t your competitive advantage.

ChatGPT, Claude, Gemini and the next generation of models will keep improving, and all of them can be valuable parts of a retail intelligence stack. But access to them isn’t scarce. Your competitors can choose the same model and build a similar-looking interface. The more defensible advantage comes from what surrounds the model: trusted data, agreed definitions and an understanding of how prices, promotions and customer behaviour interact. That understanding only counts for something once it’s put to work turning an answer into a decision.

Before placing an AI agent on your retail data, ask whether it understands your agreed definitions, recognises how your data connects, and can explain the evidence behind its answer, so your colleagues know when to trust it, challenge it or override it. If those things aren’t in place, you may have built an impressive interface but you haven’t built retail intelligence.

Our new agentic insight tool helps retailers bring their business definitions and retail expertise into the way AI interrogates performance. Find out more about our Agentic Advantage approach, or get in touch at contact@hyperfinity.ai.