The Most Expensive Mistake In Ecommerce Is Invisible

Okay, I need to talk about something that's been gnawing at me. And no, it's not about another shiny new AI tool that's going to "revolutionise" ecommerce (give me strength). It's about something much less sexy and much more likely to actually wreck your business: the data underneath your AI agents is probably wrong.

Not dramatically wrong. Not "the spreadsheet is full of expletives" wrong. Wrong in the quiet, insidious way where everything looks fine right up until the moment it really, really isn't.

We've All Got a Junk Drawer (Yours Just Makes Pricing Decisions Now)

Here's what's happening across ecommerce right now. Everyone's deploying agents. Repricing agents. Advertising optimisation agents. Inventory forecasting agents. Listing copy agents. The pitch is always the same: hand the task to the machine, free yourself to think strategically, watch the magic happen.

And look, the appeal is real. I get it. Nobody got into ecommerce because they loved manually adjusting bids at midnight. (Well, maybe some of us did at the start, in a Stockholm syndrome kind of way, but we've grown.)

The problem is that these agents are making decisions based on data that hasn't been verified in months — sometimes years. The product cost spreadsheet still reflecting last year's freight rates. The Amazon ad campaigns optimising bids against a target ACoS that was set before your category's competitive dynamics completely shifted. The repricing rules calibrated to a competitor set that includes three ASINs that have been discontinued. The inventory reorder points calculated from demand data that includes two months of a stockout — which the model helpfully interprets as "low demand" rather than "you literally had no stock to sell."

Every one of these is a real data quality failure happening in real businesses right now. And every one of them is invisible until the agent acts on it.

The Human Filter You Didn't Know You Had

Here's the thing nobody talks about: before agents, bad data was annoying but survivable. Your advertising was slightly less efficient. Your pricing was slightly off. Your restock timing was slightly late. But a human sat somewhere in the process — eyeballing the replenishment order, checking the margin calculation, doing that instinctive "wait, that doesn't look right" thing that humans are quietly brilliant at.

You know that background process your brain runs every time you open a spreadsheet? The one where you silently correct for the fact that the COGS figure is stale, or that one competitor delisted last month, or that the keyword data is from a tool you cancelled in February? That process is genuinely sophisticated data quality management. You just never thought of it that way because it felt like common sense.

Agents remove that filter. That's their entire value proposition — they act at speed, at scale, without waiting for a person to squint at a number and think "hmm." They also act with absolutely zero instinct to question whether the data in front of them still means what it says. (Which, if you think about it, makes them the most confidently wrong colleague you've ever had. We all know someone.)

The Failure Mode That'll Keep You Up at Night

This is the part that genuinely unnerves me. AI systems acting on flawed data don't fail loudly. They fail plausibly.

A repricing agent will chase a competitor's price that was scraped incorrectly. An advertising agent will keep scaling spend on a keyword where the conversion data is inflated by duplicate orders in the source file. An inventory agent will trigger a reorder based on demand signals that include returns miscoded as sales. A listing optimisation agent will rewrite your copy anchored to keyword research from eighteen months ago, when the search landscape was materially different.

None of these failures announce themselves with a siren and a flashing red light. The outputs look reasonable. The recommendations sound coherent. The campaign launches. The price adjusts. And then, over weeks, margins erode, ad efficiency declines, stockouts appear in the wrong categories — and nobody can diagnose the cause from the dashboard. Because the dashboard is built on the same compromised data. (It's turtles all the way down, except the turtles are lying to you.)

Research on AI data quality keeps confirming the same structural point: models don't merely inherit the weaknesses of their data, they amplify them at scale. The differentiator between organisations that successfully scale AI and those that don't isn't the sophistication of their tools — it's the maturity of their data practices. Which is the most unsexy competitive advantage in the history of competitive advantages, but here we are.

This Is Not an Enterprise Problem (Sorry)

I know what you're thinking. "Data governance? That's for companies with a data engineering team and a six-figure tech stack." The vocabulary sounds enterprise. The problem is not.

A solo Amazon seller's business runs on data with exactly the same failure characteristics, just in smaller containers. The cost-of-goods spreadsheet that hasn't been updated since the last supplier renegotiation. The FBA fee assumptions baked into a margin calculator that predates the most recent fee restructure. The keyword tracker still pulling search volume estimates from a tool you cancelled three months ago but never disconnected. (We've all got at least one zombie subscription haunting our tech stack. Mine was a heatmapping tool I forgot about for an embarrassingly long time.)

None of this matters much while you're making every decision yourself, because you compensate instinctively. The moment you hand any of it to an agent, that compensation disappears. The agent takes the spreadsheet at face value. It doesn't know what you know. It doesn't even know that it doesn't know.

And there's a second failure mode that cuts in the opposite direction: overloading context. The instinct, when setting up an AI agent or a Claude Project, is to feed it everything — every export, every historical report, every document — on the theory that more context produces better decisions. It doesn't. Relevance is a data quality dimension in its own right. What agents need isn't a bigger archive, it's an accurate, current picture of how your business actually operates now. Curation, not accumulation. (Think of it less like stocking a library and more like packing a carry-on. Everything in there should earn its place.)

What to Actually Do About This

Right. Practical bit. The good news is this isn't a six-month transformation project. It's an audit, and for most sellers it's genuinely achievable in a day or two.

For solo sellers and small teams: Start by identifying the core datasets any agent would touch — product costs and margins, inventory levels and lead times, advertising benchmarks, keyword and competitor data, customer purchase history. Date-stamp each one. Anything that hasn't been verified in the last quarter gets checked before it gets connected to any automation. Deduplicate your product data — if you're tracking the same SKU across multiple spreadsheets with slightly different cost figures, you don't have two sources of truth, you have zero. Update your FBA fee assumptions and landed cost calculations. And before building any agent, write down — in plain language — what data it should act on and what it should ignore. That document does more for agent reliability than any amount of prompt engineering.

For mid-sized brands and growing teams: The priority shifts from auditing data to governing it. Someone specific needs to own the data layer your agents depend on — with the ability to answer one question at any time: what are our agents acting on, and when was it last verified? In most ecommerce businesses, this ownership currently sits in a gap between marketing, operations, and whoever manages the tech stack. Which means it sits nowhere. That gap is manageable when humans review every decision. It becomes a structural vulnerability the moment agents start executing. Establish precedence rules for conflicting data. When your ad platform reports one conversion figure and your Shopify backend reports another, which does the agent trust? These questions need answers before the agent encounters them — not after it's already acted on the wrong number.

For larger brands, aggregators, and agencies: The specific risk here is agent-to-agent workflows, where one agent's output feeds another agent's input. A demand forecasting agent that overestimates sell-through because of flawed historical data triggers an inventory agent to over-order, which triggers a pricing agent to discount to clear excess stock, which triggers an advertising agent to scale spend against artificially low prices. Each agent performed its function correctly. The cascade was wrong because the first input was wrong. Oversight needs to cover the full workflow, not individual tools — audit trails, drift monitoring, and human checkpoints at the decisions where the cost of a confident error is highest.

The Bottom Line

The industry's attention is fixed on what agents can do. The more consequential question is what they're standing on.

An agent is an amplifier. Point it at accurate, current, well-structured data and it compounds the quality of that foundation. Point it at the data nobody's checked in two years and it compounds that instead — faster than any human process ever could, and more quietly.

The appealing project is building the agent. The valuable project is the unglamorous one that precedes it: knowing what your data says, when it was last true, and who's responsible for keeping it that way.

• ChatGPT Ads rolls out an AI feature that auto generates ad variations from your website and campaign settings, with review, edit, and activate controls. OpenAI aims to speed ad creation and grow revenue on ChatGPT.

• OpenAI launches GPT-Live, full duplex voice models that listen and speak at once for natural conversation. GPT-Live-1 and mini roll out globally, delegate to GPT-5.5 for complex tasks, and reach the API soon.

• OpenAI expands its ad pilot across Europe weeks after its UK launch, hiring an EMEA ads solutions manager to scale the business. The role builds a regional team and sets country and partner strategy.

• Anthropic launches Reflect in beta, a dashboard showing when you use Claude most and on what, with quiet hours and break nudges. It excludes incognito chats and connected tool files.

📲 Arcads - Generate UGC-style video ads from a script using 1,000+ AI actors.

⚙️ Relay.app - Workflow automation with built-in human-in-the-loop approval steps.

🗣️ Cartesia - Ultra-low-latency text-to-speech and voice cloning for real-time apps.

🧑‍💼 Beautiful.ai - AI presentation maker with auto-designed, on-brand slides.

📧 Clay - Enrich and scrape web data to build hyper-personalized prospecting lists.

😍It was great, all I needed

🙂Good, not great

🙄It sucked