No.
You do not need perfect data before starting AI initiatives on manufacturing KPIs. In most plants, perfect data never arrives, especially in brownfield environments with mixed MES, ERP, historian, QMS, spreadsheets, and manual logs. If you wait for complete standardization and total cleanup first, the AI program usually stalls.
What you do need is data that is good enough for the specific question you are trying to answer, with known limitations documented up front. That means being explicit about where the data comes from, how the KPI is defined, what is missing, and how much error the use case can tolerate.
A narrow use case with a clear decision point, such as identifying likely causes of recurring downtime, yield loss by routing step, or late order risk.
A stable KPI definition. If each site or function calculates OEE, scrap, cycle time, or schedule adherence differently, AI will amplify confusion rather than reduce it.
Basic data lineage and traceability. You should be able to show what source systems were used, what transformations occurred, and which records were excluded.
A quality baseline. Measure completeness, timeliness, consistency, and known gaps before claiming insight.
Human review. Early outputs should support operations, engineering, and quality decisions, not replace them.
Weak data does not always stop a project, but it changes what is realistic.
If timestamps are inconsistent, sequence and duration analysis may be unreliable.
If master data is fragmented, cross-system KPI rollups may be misleading.
If reason codes are incomplete or operator-entered with poor discipline, root cause patterns may be noisy.
If process changes are not controlled, model performance can degrade without obvious warning.
If labels are subjective or inconsistently applied, supervised learning may not be trustworthy.
In other words, imperfect data is acceptable for some descriptive and prioritization use cases. It is much less acceptable for automated decisioning, closed-loop control, or anything presented as a definitive explanation of process behavior.
Start with bounded use cases where the cost of being directionally wrong is manageable and where results can be checked against known process knowledge. Examples include anomaly triage, downtime categorization support, queue aging analysis, or identifying which data collection gaps most distort a KPI.
This is usually safer than starting with plant-wide optimization claims or full replacement of existing reporting stacks. In regulated, long-lifecycle environments, full replacement strategies often fail because of validation burden, qualification concerns, downtime risk, integration complexity, and the need to preserve traceability and change control across legacy systems.
A more durable pattern is coexistence. Keep the existing MES, ERP, QMS, and historian as systems of record, then add an analytics or AI layer that is tightly scoped, versioned, and governed. That does not remove integration debt, but it limits operational risk and makes validation more manageable.
Starting early creates learning, but it also exposes data defects faster.
Cleaning data first improves confidence, but large cleanup programs often overrun before any operational value is proven.
Using AI on partially manual datasets may still help prioritize improvement work, but results need stronger review and caveats.
Standardizing KPI definitions across sites improves comparability, but can take significant process and governance effort.
The right balance depends on process maturity, integration quality, and whether the output will be used for exploratory analysis, operational management, or regulated evidence. Those are not the same bar.
Do not ask whether the data is perfect. Ask whether it is sufficiently reliable for this KPI, this decision, and this level of consequence.
If the answer is yes, start small and govern tightly. If the answer is no, the first AI use case may need to be data quality monitoring itself.
Whether you're managing 1 site or 100, Connect 981 adapts to your environment and scales with your needs—without the complexity of traditional systems.
Whether you're managing 1 site or 100, C-981 adapts to your environment and scales with your needs—without the complexity of traditional systems.
