“We have lots of data, but we’re not making the most of it.” We hear this all the time. We also hear the opposite: “We don’t have enough data for AI.” Oddly enough, both sometimes come from the same company, depending on who you ask.
The truth is that the data AI needs depends entirely on the type of project. There’s no universal threshold. What there is, though, is a set of clear patterns for judging whether the data you have is enough for a specific use case — and what to do when it isn’t.
The “big data” myth
One of the most widespread misconceptions is that AI requires massive amounts of data. It comes partly from the marketing of recent years, which associated AI with Google or Amazon projects processing billions of records.
Business reality is different. A sales forecasting model for a midsize company can work perfectly well with enough order history — not necessarily huge volumes, but data that represents the business. An anomaly detection system on a production line can be trained on the sensor history you already have. A document classifier can be accurate with hundreds of labeled examples, not millions.
What matters isn’t absolute volume but whether the data covers enough of the variability in what you want to model. One year of data in a seasonal business is worth far less than three years covering several full cycles.
The three dimensions that really matter
Relevance: The data must contain information about the variables that drive the outcome you want to predict or automate. It sounds obvious, but it isn’t: many companies have plenty of data on what’s easy to measure (process times, units produced) and little on what drives the outcomes that matter most (why customers cancel, the root cause of quality defects).
Quality: Data can exist and still be unusable if it has high rates of null values, systematic entry errors, duplicates, or inconsistencies between systems. A rule of thumb: if more than 20–30% of values are missing in key variables, data cleanup comes before the AI project.
Time span: For predictive models, the history needs to cover enough time for the model to learn the relevant patterns. If your business has annual seasonality, two or three years of data are far more useful than six months.
Types of project and their data requirements
Automation with generative AI (email classification, document summaries, drafting replies): No training data of your own required. Large language models come pre-trained. What you need are the documents or texts that provide context (manuals, catalogs, FAQs). Any company with internal documentation has enough.
Supervised prediction models (demand, churn, pricing): You need history of the variable you want to predict, plus the relevant explanatory variables. The minimum volume depends on the problem: you need enough examples of the event you’re predicting for the model to learn the pattern. The rarer the event, the more history you need. More data is better, though with diminishing returns beyond a certain point.
Computer vision (quality control, image classification): You need images labeled by category. The minimum volume depends on the complexity of the problem, but a common reference is hundreds of images per class. Transfer learning reduces this requirement significantly.
Anomaly detection (predictive maintenance, fraud): Can work with mostly “normal” data if the anomaly is rare. The models first learn normal behavior and then detect deviations. A history of failures or fraud helps, but isn’t always essential.
What to do when the data isn’t enough
This is the most common situation for companies starting out with AI, and it doesn’t have to be a blocker.
Start collecting data systematically. If the data you need isn’t recorded today, now is the time to start. Many successful AI projects began six or twelve months before a single model was trained, by building the data collection infrastructure.
Fit the use case to the data you have. Instead of going after the ideal problem, go after the one your current data lets you solve. A less ambitious model in production creates more value than a perfect model that never exists.
Data augmentation. In computer vision, there are techniques to artificially multiply the volume of training data (rotations, lighting variations, and so on). In other domains, data synthesis techniques can expand a small training set.
Transfer learning. Models pre-trained on large datasets can be adapted to specific tasks with relatively little data of your own. This technique has sharply reduced training data requirements for many kinds of problems.
A data inventory as step one
Before deciding whether you have enough data for an AI project, it’s worth taking a systematic inventory: what data you record, which systems it lives in, how good it is, how far back it goes, and whether it can be exported and joined with other sources.
This inventory — part of what an AI audit does systematically — rarely produces extreme surprises either way. Typically you find enough data for some priority use cases, data that needs improving for others, and clear gaps that show where to invest in data infrastructure.
Frequently asked questions
Can supplier data or external data complement our own? Yes, in many cases. Weather, economic, or traffic data can enrich forecasting models. Industry data or benchmarks can provide context. The key is making sure that integrating external data complies with applicable privacy law, and that the external data has the quality and update frequency the model needs.
How long does it take to “clean” data for an AI project? It depends on where you start. When the data sits in well-maintained systems, preparation is relatively straightforward. When it’s spread across multiple sources with mixed formats and quality problems, preparation can be by far the longest task in the whole project.
What about customers’ personal data? Working with personal data in AI projects has implications under the GDPR and, depending on the type of project, under the EU AI Act. Data minimization — using only the data you strictly need — and anonymizing or pseudonymizing training data are good practices that reduce both regulatory risk and privacy risk.