One phrase from the data world sums up the most common problem in AI projects: “garbage in, garbage out.” An AI model is, at heart, a system that learns patterns from the data it’s trained on. If that data is wrong, incomplete, or unrepresentative, the model learns the wrong patterns — and makes wrong predictions or classifications with complete confidence.
For companies evaluating or starting AI projects, data quality is most often what separates a project that delivers value from one that never reaches production.
Why nobody wants to look at the data first
There’s a practical reason many AI projects start with the most visible part — the model, the interface, the demo — instead of the data: the data is what exposes an organization’s underlying problems.
A serious data inventory turns up uncomfortable things: duplicate records nobody cleaned up, mandatory fields nobody actually filled in, systems that were supposedly integrated but use different keys and can’t be joined, histories that only recently started being recorded properly after years of incomplete data.
This diagnosis is uncomfortable because it reveals data cleaning and governance work that wasn’t in the original budget. But doing it upfront — before committing budget to model development — is far more efficient than discovering it halfway through.
The most common data quality problems
Null or missing values in key variables. If the variable the model has to predict — or the variables it uses to predict — have a high share of missing values, the model struggles to learn robust patterns. Sometimes missing values carry meaning (an empty field on an order may mean the customer had no preference, not that the data is wrong). Handling missing values correctly takes an understanding of the business, not just the data.
Duplicate records. Business systems accumulate duplicates over time — the same customer with two records, the same supplier with name variations, the same product with different codes in different periods. For ML models, duplicates distort the patterns the model learns.
Inconsistencies between systems. A company can run its ERP, CRM, production system, and quality system on data that describes the same entities — customers, products, orders — yet can’t be joined because the systems use different keys or formats. Integrating these sources is often the longest task in a data project.
Historical data that doesn’t reflect the present. Companies change: product portfolios, markets, and processes all shift. Sales history from several years ago may not represent today’s demand. Training on very old data can teach the model patterns that no longer hold.
Bias in training data. Data reflects a company’s past decisions. If credit was historically approved only for customers with a certain profile, the data carries that bias and the model can perpetuate it. Identifying and correcting bias in training data is part of the ethical design of AI systems, and the EU AI Act includes it in its requirements.
How to assess data quality before you start
A good data quality inventory answers four basic questions:
Does the data exist, and can you access it? Not all the data that theoretically exists in a company can be accessed programmatically. Data in unstructured PDFs, scattered Excel files, or legacy systems with no API takes extra work to extract.
Is the data complete? For each variable relevant to the use case: what percentage of records has a value? Do the missing values follow a pattern (certain periods, products, or customers) that could introduce bias?
Is the data consistent? Are the same concepts coded the same way across records and systems? Are units consistent? Are date formats consistent?
Is the history long enough and representative? Does it cover enough complete cycles of what you want to model? Are exceptional events properly flagged?
What to do when quality isn’t good enough
A data diagnosis that uncovers quality problems isn’t a death sentence for the project. You have options:
Clean the historical data. For many quality problems, the answer is a systematic cleanup of the history. Depending on volume and complexity, this can be a project in itself or a phase of the AI project.
Adjust the scope of the first project. If the data of sufficient quality only covers part of the original use case, it can make sense to start with that subset — the part with clean data — and widen the scope once the rest of the data improves.
Start collecting data properly now. If the data you need doesn’t exist at sufficient quality, now is the time to start recording it properly. Sometimes the AI project starts with a phase that sets up data collection, and the model is built once there’s enough quality history.
Frequently asked questions
Are there tools that assess data quality automatically? Yes. Data profiling and data quality tools automatically analyze completeness, consistency, uniqueness, and distribution. They speed up the diagnosis but don’t replace it — you still need a business expert to interpret the findings.
Is data governance the same as data quality? They’re related but different. Data quality is the current state of the data (complete, consistent, correct). Data governance is the processes, policies, and responsibilities that keep quality up over time. Without governance, quality degrades. Without quality data, governance has nothing to govern. Well-designed AI projects address both.
How long does it take to improve a company’s data quality? It depends on the starting point and the scope. Data cleanup for a specific use case can be done in weeks. Company-wide data quality improvement is a long-term initiative. The most practical strategy is to start with the data needed for your first, highest-value use case.