You probably do not need more data—you need the right data

More data is not automatically better data

Manufacturers often delay an AI project because their data is incomplete, spread across systems, or missing from older equipment. Those are real constraints, but waiting to collect everything can create an expensive holding pattern. The useful question is narrower: what information is needed to improve one decision in one process?

A maintenance model may need operating cycles, alarms, work history, and a technician’s observation—not every signal the plant can produce. A quality workflow may need images, inspection outcomes, material context, and line conditions. Relevance, timing, and trustworthy labels matter more than raw volume.

Start with the decision and work backward

Define who will use the output, what action it could change, and what evidence would justify that action. Then map the smallest set of available inputs that could support the decision. This keeps the project connected to operations and exposes missing context before it becomes a model problem.

Treat operator knowledge as data

Experienced people recognize shifts, sounds, combinations, and exceptions that may never appear in a database. Capturing that knowledge alongside machine and business data often produces a better starting point than pursuing a perfectly clean historical dataset in isolation.

A practical first move

Choose one recurring decision, gather representative examples, document what people look for today, and test whether the available evidence can support a measurable improvement. If the evidence is weak, the pilot should reveal exactly which data is worth collecting next.

Source: NIST — How to Find the Right Balance of Data for Your Industrial AI System