Why AI gets your e-commerce numbers wrong, and how to fix it

Connecting Claude to Shopify, HubSpot or your other business tools takes minutes. Getting answers you can actually trust is a different story.

AI has made business analytics more accessible than ever. You can connect Claude to your CRM, online store or marketing platforms, ask a question in plain English and get an answer in seconds, with no SQL, no dashboards to build and no waiting for an analyst. For an e-commerce business that has struggled to get useful information out of its data, that feels like a breakthrough, and in many ways it is.

But an AI assistant that can answer questions about your business is a different thing from an analytics system that answers them correctly and consistently. You notice the gap when your CFO asks why the revenue in your AI-generated report doesn’t match Shopify, when your marketing team sets a budget on a customer acquisition cost that turns out to be wrong, or when two executives ask the same assistant the same question and get different answers.

The AI is rarely the whole problem. Most of the trouble sits in everything that has to happen to your data before AI can give you a trustworthy answer.

A real example: nearly twice as many customers as actually existed

A subscription e-commerce brand’s AI assistant reported nearly twice as many active customers as the brand actually had. We had set out with this brand to measure exactly what changes when you move from a direct AI connection to properly prepared, AI-ready data, and before changing anything, we established a baseline. We asked their AI assistant a simple question: “How many active customers do we have?” It answered in seconds, confidently.

Two things had gone wrong. The assistant was reading an outdated copy of the company’s data, and the table it used held several copies of individual subscriptions, so customers who had cancelled were still counted as active. The AI noticed neither problem. It counted the records and presented the result as fact.

A minute later, we started a new conversation about customer retention. The same assistant, looking at the same table, noticed the duplicate subscription records and removed them from its calculation. Same assistant, same data, two different approaches, and neither answer came with a warning that the underlying data might be unreliable.

That is why connecting AI directly to your business systems isn’t enough. You need reliable data, clear business definitions, and calculations that don’t change depending on how the AI reads your question.

Problem 1: Your AI might not be looking at all your data

A direct AI connection can leave your assistant working from part of your data, and from none of its history. When you connect Claude directly to Shopify, HubSpot or another platform, it usually reads the data through an API. An API is the doorway that lets one piece of software request information from another. Those doorways have limits. Most systems return records in batches, called pages, and cap how many requests you can make in a given period. A data pipeline is the automated process that copies data from your business tools into one place. A proper one is built around those limits: it retrieves every page, handles rate limits, retries failed requests and checks whether the data is complete. An AI connector may not do all of that reliably.

Say you ask Claude: “How many customers purchased from us more than once this year?” Your store has 200,000 customers. If the connector retrieves only some of those records, Claude calculates your repeat purchase rate on an incomplete dataset. You get a percentage that looks plausible, with no sign that thousands of customers were left out.

There’s a second gap. Your AI usually sees your systems as they are today, not as they were six months ago. When a customer changes their subscription, a sales opportunity moves stage or an order gets updated, the previous state may be gone. That makes historical questions surprisingly hard:

• How many active subscribers did we have on March 1?
• How has our sales pipeline changed over the last six months?
• What was our customer lifetime value at the end of last year?

If your systems haven’t kept that history, no amount of clever prompting will bring it back. And waiting makes it worse. AI connectors are improving fast, and it’s tempting to wait six months to see what becomes possible, but better AI won’t recover history your systems have already overwritten. Start keeping subscription changes today and you’ll have six months of history six months from now. You can upgrade your AI later. You can’t always recover yesterday’s data.

The fix is a data warehouse. A data warehouse is a central database that regularly collects information from your business systems, checks that everything arrived and keeps historical changes, so AI has complete, reliable information to work with. Our practical guide to the e-commerce data stack covers the pieces.

Problem 2: Your business systems don’t agree on what the numbers mean

If you’ve ever compared revenue in Shopify, Google Analytics and Meta Ads, you’ve seen three revenue numbers for one business. Each system was built for a different purpose. Shopify records orders and transactions, Google Analytics tracks website activity and conversions, and Meta attributes purchases to ad interactions under its own attribution rules. None of them is necessarily wrong; they answer different questions.

Now connect Claude to all three and ask: “What’s our return on ad spend?” Claude has to decide which revenue to use. Meta’s attributed revenue or actual Shopify sales? Before or after refunds? With or without returning customers? And what about orders that several ad platforms claim? None of these has one technically correct answer. They’re business decisions, and your business needs to make them before the AI starts calculating.

Combining systems makes it harder. Take customer acquisition cost. Customer acquisition cost (CAC) is your marketing spend divided by the number of new customers you acquired. Your spend may be split across Meta, Google Ads, TikTok and influencer campaigns, while the customer count comes from Shopify, which may count someone as new even if they bought before under a different email address. Add subscription data in Recharge, customer records in HubSpot and refunds recorded separately, and a simple CAC now means combining several platforms, matching reporting periods, identifying customers across systems and agreeing on what counts as a new customer. Claude can help you work through those decisions. It shouldn’t reinvent them every time someone asks for CAC.

The fix is one consistent set of business definitions on properly integrated data. Data engineers call this a semantic layer. A semantic layer is a shared dictionary that tells every report, dashboard and AI assistant exactly what your business means by revenue, customer acquisition cost, lifetime value or any other metric. Each calculation is defined once and every tool uses the same logic, so your CFO and CMO can discuss what the numbers mean instead of arguing about which number is correct.

Problem 3: Teaching AI your formulas isn’t enough

Better instructions help, but they don’t guarantee the same calculation every time. “Can’t I just give Claude better instructions?” You can, and you should. Write down how your business calculates revenue, which orders to exclude and how to handle refunds, and Claude can reuse those instructions through features such as Skills.

But knowing the formula and executing it correctly every time are two different things. Say your instruction reads: “Calculate net revenue by subtracting discounts and refunds from gross sales. Exclude test orders and cancelled orders.” Now apply it to 150,000 orders, where some have partial refunds, some several refunds, some refunds still pending, some orders cancelled after payment, and some discounts on single products rather than the whole order. Claude has to interpret the instruction, retrieve the data and run the calculation, and when you ask again tomorrow, it may take a different route.

AI can write code to do exact calculations, which beats asking a language model to add up thousands of records in its head. But unless that code is saved, tested and reused, you’re still asking AI to recreate your business logic every time. It’s like hiring a new analyst every morning, handing them your documentation and hoping they read it exactly as yesterday’s analyst did.

A semantic layer gives you one tested calculation that everyone uses. The AI asks for net revenue, and the system underneath runs the agreed calculation. Let AI interpret your questions. Let reliable software calculate your answers.

Problem 4: You can’t always explain where the answer came from

Answers an AI works out mid-conversation are hard to trace back to their source. Picture your quarterly board meeting. Your report shows a 23% increase in customer lifetime value, someone asks what’s driving it, and Claude gives you a convincing explanation. Then your CFO asks to see exactly which customers and transactions were included. If the answer came from an AI conversation that pulled raw records and calculated on the fly, you may not be able to reproduce the calculation, show the underlying records, list the exclusions or confirm the data was complete when it ran.

This is one of the biggest differences between exploratory AI analysis and production-grade business intelligence. A properly designed analytics system gives you a trail back to the source: where the data came from, how it was transformed, which business rules were applied and when it was last updated. It can also run automated checks. If your Shopify data suddenly shows almost no orders yesterday, the system flags a possible failure in the data feed. Without that check, Claude might simply tell you sales were unusually slow, and you might start investigating your marketing campaigns instead of a broken connection.

Problem 5: Your quick AI setup becomes another system to maintain

Connecting an AI assistant to your tools is easy. Keeping it reliable as your business grows is another matter. Connectors and APIs change, permissions get updated, data structures evolve, and new products, markets and sales channels bring exceptions your original instructions didn’t anticipate. Meanwhile, people build their own AI workflows: your marketing manager has one set of instructions, your finance team another, and your CEO asks questions in a separate conversation. Before long you’ve recreated the problem you were trying to escape. Everyone has access to data, and nobody agrees on the numbers.

Security matters too. Direct connectors may expose customer details and other sensitive information to the AI service, and some can change records as well as read them. Analytics rarely needs either. A better setup gives AI only the information it needs, ideally through a read-only connection with the right permissions. Your assistant doesn’t need every customer’s email address to tell you whether your repeat purchase rate went up, and it shouldn’t be able to edit products or customer records while preparing a sales report.

If two or more of these sound familiar, book a free data strategy call and we’ll look at which of your numbers you can trust today.

Should you stop using AI for analytics?

No. I think AI can make business analytics far more useful. Traditional dashboards have a built-in limit: someone has to anticipate the questions people will ask. When someone wants revenue without wholesale, or first-time versus returning customers, the dashboard has no answer until someone changes it or writes a custom query. With AI, business users can simply ask:

• Which products had the largest decline in repeat purchases?
• How did our acquisition costs change after we increased Meta spending?
• Which customer segments are becoming less profitable?
• Why did our subscription revenue increase while total revenue stayed flat?

To answer those reliably, AI needs data that has already been collected, cleaned, connected and defined. The future of analytics combines AI with traditional business intelligence.

What reliable AI analytics looks like

Here’s the approach I recommend for businesses that want to use AI seriously.

1. Bring your business data together. Instead of having AI pull raw records from Shopify, HubSpot, Meta, Google Ads and your subscription platform for every question, bring the data into a central warehouse such as Databricks, BigQuery or Snowflake. Automated pipelines collect it, handle API limits, keep it current and preserve history your source systems would otherwise lose.
2. Clean and connect the data. This is where the real work happens: excluding test orders, handling refunds consistently, matching customers across systems, standardizing time zones, connecting marketing spend to sales and combining subscription activity with purchase history, so the data reflects how your business actually operates.
3. Define your metrics once. Revenue, gross margin, CAC, customer lifetime value, repeat purchase rate, subscription churn: each gets one clear definition, agreed by the people who use it, and becomes a reusable calculation in a semantic layer. Dashboards, reports and your AI assistant then work from the same numbers.
4. Connect AI to the trusted data. Claude requests the metrics it needs instead of reading thousands of raw records. The database does the calculating, the semantic layer supplies the definitions, and Claude handles the conversation, so you can ask follow-ups and investigate surprises without rebuilding the logic each time.
5. Add monitoring, security and accountability. Check that data arrives on time, flag unusual changes, control access to sensitive information and keep calculations traceable. You should be able to explain a number six months from now, not just five minutes after Claude produced it.

That may sound like a lot of infrastructure for a few questions, and for some questions it is. If you only need to look up a customer, check a recent order or summarize a handful of CRM records, a direct AI connector may be enough. If you’re using AI to track performance, compare marketing channels, calculate profitability or make decisions involving millions of dollars, the requirements change. You also don’t need to throw away what you’ve built. Maybe you have a warehouse but no consistent metric definitions, dashboards nobody uses because they’re too rigid, or Claude connected to everything and numbers that don’t match. Sometimes the fix is a small improvement to your existing setup, sometimes a rethink of how your systems exchange data, and sometimes simply finding out which numbers you can trust today.

Five questions to ask before you trust your next AI report

1. Is the data complete? Do I know whether AI retrieved every relevant record, or only a portion?
2. Are the metrics clearly defined? Would my CFO and marketing team agree on what the numbers include?
3. Are the calculations repeatable? Will I get the same answer tomorrow if the underlying data hasn’t changed?
4. Can I trace the results? Could I explain exactly where a number came from three months later?
5. Is the setup secure and maintainable? Can my team rely on it without constantly fixing connectors, instructions and inconsistent reports?

If you can’t answer these confidently, your AI analytics may be giving you a false sense of certainty, and there’s room to make it much more valuable.

Let’s find out how reliable your AI analytics really is

I’m Corinna Zennig, founder of LazyAnalyst. I’ve spent more than a decade helping businesses turn messy, disconnected data into reporting they can trust, and AI hasn’t changed what the hardest part of analytics is: knowing whether an answer is correct.

Business leaders shouldn’t have to wait for an analyst every time they want to explore a question, and they shouldn’t have to wonder whether their AI assistant counted the same customer three times. We help e-commerce companies move from experimenting with AI analytics to a foundation they can rely on, without replacing everything they’ve built and without recommending an expensive data platform when something simpler would work. We start with what you have today: the questions you’re asking AI, where the data comes from, how your metrics are defined, and which answers you can already trust. Then we fix what actually needs fixing, whether that’s your pipelines, your business definitions, your semantic layer or the way AI reaches your data.

You don’t need to give up the convenience of AI to get trustworthy numbers. You need to give AI something trustworthy to work with.

---

About the author. Corinna Zennig leads LazyAnalyst, an e-commerce analytics consultancy, and works as a fractional Chief Data Officer. She has 15 years in data analytics and has delivered more than 30 business intelligence projects. More about Corinna