A retail chain in the Midwest spent $1.2 million building a custom AI recommendation engine in 2025. The system was supposed to analyze purchase history and predict what customers would buy next. It worked beautifully in testing. In production, it kept recommending snow shovels to customers in Florida and baby formula to households with no children. The model was fine. The customer data it was trained on had not been cleaned since 2019, when a system migration had scrambled zip codes and merged duplicate accounts incorrectly.

This story is not unusual. It is the norm. Gartner estimated in 2024 that poor data quality costs organizations an average of $12.9 million per year. That number was calculated before most of those organizations started feeding their data into AI systems that amplify every error at machine speed. Bad data in a spreadsheet is a nuisance. Bad data in a machine learning model is a multiplier.

The conversation about AI in business has been dominated by model selection — which platform, which provider, which architecture. Almost nobody is talking about the data those models consume. And the data is where every AI project succeeds or fails.

The principle is decades old: garbage in, garbage out. What AI changes is the scale of the garbage and the confidence with which it is presented. A flawed spreadsheet looks like a flawed spreadsheet. A flawed AI output looks like a polished, professional analysis. It arrives formatted, cited, and articulate. The errors are buried inside the confidence.

A logistics company in Germany fed three years of shipping data into an AI optimization tool in early 2025. The tool reduced route planning time by 60% and projected annual fuel savings of $800,000. Six months later, the actual savings were $120,000. The discrepancy was traced to the training data: the shipping records included pandemic-era routes that reflected temporary road closures, port delays, and fuel price anomalies. The model had optimized for a world that no longer existed.

Staleness. Data has a shelf life. Customer preferences change, markets shift, regulations update. A model trained on 2023 data will produce 2023 answers. If your business environment has changed — and it has — your training data is lying to you quietly.

Duplication. Most CRM systems contain between 10% and 30% duplicate records. When an AI model analyzes a customer base with 25% duplicates, it overweights the preferences and behaviors of duplicated customers. The result is a model that serves a quarter of your customers twice as well and the rest not well enough.

Selection bias. The data you have is not the data you need. It is the data you happened to collect. A B2B company that tracks only closed deals will build an AI model that understands what winning looks like — but has no information about why it lost the deals it lost. The model cannot optimize for what it cannot see.

Start with an audit. Not a technology audit — a data audit. Take the three most important datasets your business relies on and answer four questions about each: When was this data last cleaned? What percentage is duplicated? How current is the oldest record in active use? What data are we not collecting that we should be?

Most companies cannot answer all four questions for even one dataset. That gap is your AI risk.

Next, establish a cleaning cadence. Data hygiene is not a project. It is a maintenance schedule, like changing the oil in a truck fleet. Quarterly deduplication, annual schema reviews, and real-time validation on new entries will prevent most of the errors that derail AI implementations.

Choosing between GPT and Claude, between TensorFlow and PyTorch, between building and buying — these are technical decisions with marginal business impact. The decision with enormous business impact is whether you feed your AI clean data or hope for the best. One path leads to useful output. The other leads to confident, well-formatted nonsense that your team will trust because the presentation is convincing.

Fix the data first. Everything else is optimization.