The question that really decides an industrial AI project isn’t which model to use. It’s whether your plant’s industrial data is in a fit state to support it.
Every week, plants come to Deduce Data Solutions that already have a SCADA, MES or ERP system running for years. Even so, they’re unsure whether that “counts” as enough data.
The answer is almost never a flat yes or no. It’s a range, and knowing where on that range your plant sits completely changes how the project is designed.
In this article we share the same 7-question checklist we use internally to diagnose industrial data readiness before starting an Industrial Process Intelligence (IPI) project. We back it up with verifiable data on why this matters so much, and two real cases where far-from-perfect data still made better decisions possible.
AI projects don’t fail because of the model; they fail because of the data
Most conversations about industrial artificial intelligence revolve around the algorithm: which model, which architecture, which vendor. But the most common cause of failure lies one step earlier, in the quality and traceability of the underlying industrial data.
A recent MIT NANDA report, “The State of AI in Business 2025”, analysed dozens of corporate generative AI deployments. It found that 95% of pilots fail to generate a measurable return for the business.
The recurring cause isn’t model quality. It’s the lack of integration with the organisation’s real data flows and processes (MIT NANDA / Fortune, August 2025).
McKinsey reaches a similar conclusion from another angle. In its analysis “AI data readiness: Foundation for scaling enterprise AI” (2026), only 7% of companies have managed to scale AI across the whole organisation.
More than two-thirds of those that do get results point to data, not technology, as the main obstacle. The gap isn’t in ambition; it’s in the foundation.
The pattern repeats in industry. Deloitte’s 2025 Smart Manufacturing and Operations Survey, conducted among 600 manufacturing executives, shows that only 29% use AI or machine learning at plant or network scale.
40% place data analytics among their investment priorities for the next 24 months. The gap between “wanting to do AI” and “being able to” is largely explained by the state of the underlying industrial data.
And that data is almost never entirely missing. The problem usually isn’t the absence of industrial data, but that it exists scattered, without context, or without the traceability a model needs to learn anything useful from it.
Where your industrial data actually comes from
Before getting into the checklist, it’s worth clarifying what we mean by industrial data. Almost no plant starts from zero.
SCADA (Supervisory Control and Data Acquisition) usually provides the real-time pulse of the process: temperatures, pressures, consumption, machine states. MES (Manufacturing Execution System) adds the context of what was made, on which line, in which shift, and with what quality outcome.
ERP (Enterprise Resource Planning) closes the loop with production orders, materials and costs. On its own, each system only tells part of the story.
A useful AI project almost always needs to cross-reference all three. That’s where most of a plant’s industrial data reveals its true state of readiness, not in each isolated system but in the ability to connect them to one another.
Who should answer this checklist?
In our experience, the checklist works worse when it’s filled in by a single person from an office. A technology or operations director can assess the systems side well (which SCADA, which MES, what integration exists) but rarely knows the daily detail of every sensor.
That’s why we recommend completing it together with someone from the plant floor: a shift supervisor, a maintenance manager or a technician with years on the line. That combination is what genuinely diagnoses the real state of that data, not just of your systems.
When both perspectives agree on the 7 answers, confidence in the result is much higher. When they don’t, that very discrepancy is already valuable information about where the project’s risk lies.
7 questions to know if your industrial data is ready
This is the checklist we use at Deduce Data Solutions before proposing a pilot. It doesn’t aim for a perfect “yes” on all 7 questions: it aims for you to know precisely where your plant stands.
1. Do your systems record consistent timestamps across SCADA, MES and ERP?
Without synchronised timestamps you can’t cross-reference a process variable with a production order or energy consumption. It’s the basis of any causal analysis of industrial data.
It’s usually the first place where mismatches appear between systems installed at different times.
2. Is there at least 12-18 months of historical data, not just real-time data?
A model needs to see seasonal variability and complete production cycles, not a snapshot of the last month. Records from a single shift or a single campaign are rarely representative of how the plant behaves over a full year.
3. Are your variable names consistent across lines, shifts and technicians?
If the same variable changes name depending on who logged it, any model will have to reconcile that by hand before it can learn anything. In practice, this inconsistency is one of the biggest time sinks in industrial data projects, well above training the model itself.
4. Do you know what percentage of your data has gaps or out-of-range values?
It doesn’t need to be zero. What’s needed is to know it, in order to decide whether to correct, impute or discard it before training anything.
Undiagnosed uncertainty is more costly than the gap itself.
5. Can you explain what each key variable physically means and why it’s measured?
Data without operational context (what process it represents, in what units, with which sensor) is hard to validate and easy to misinterpret. Industrial data is only useful when someone in the organisation can explain its physical origin, not just its column name.
6. Are stoppages, incidents and maintenance logged in a system, not just in the team’s memory?
These events are the labels a model needs to tell normal operation apart from an anomaly. Without them, even the longest historical record loses much of its predictive value.
7. Is there someone on the plant floor who can validate whether a detected pattern makes operational sense?
Perfect data doesn’t replace human judgement. That validation turns a statistical pattern into a reliable decision, and it’s as much a part of industrial data readiness as any field in a database.
How to read your result
Industrial data readiness is a range, not a pass or fail.
With 6 or 7 “yes” answers there’s a solid foundation: a pilot can be started on this data almost immediately. With 4 or 5, there are gaps, but they’re manageable, and it’s worth starting in parallel while strengthening the data at source.
With 0 to 3 “yes” answers, it’s worth sorting out data capture and traceability first. Industrial data doesn’t need to be perfect.
It does need to be reliable, with its limits known precisely before committing budget to a more ambitious project.
One important nuance: a low score doesn’t mean “wait a year before touching AI”. It means the first project should be smaller and more focused on sorting out industrial data capture than on deploying a complex model from day one.
In fact, many plants with 0 to 3 “yeses” manage to rise to 5 or 6 within a few months simply by standardising variable names and enabling systematic logging of stoppages. It’s usually not a budget problem, but one of priority and method.
The most expensive mistake: waiting for “perfect data”
No plant has perfect industrial data, not even those that have been digitised for years. A recent analysis of the “dark data” phenomenon in industry (Forbes Technology Council, December 2025) is revealing on this point.
It reports that a large share of industrial companies and plants use only around 32% of the data they already hold. The rest (sensors, historical records, maintenance logs) goes unexploited, according to Seagate’s “Rethink Data” report cited in the analysis.
Waiting to clean everything before starting only delays what really needs to be learned: what’s missing, which sensor is lying, and which process needs to be better understood. In our experience across very different plants, progress is faster another way.
It’s better to work with what you have, scope a pilot and correct as you go, rather than wait for a total clean-up that almost never arrives.
Two real cases: imperfect data that still worked
When a SCADA’s historical data was enough to save gas
Reinosa Forgings & Castings (RFC), a century-old forge in Cantabria that produces parts of up to 70 tonnes, didn’t start from perfect industrial data. It started from what it already had: the SCADA of its heat-treatment furnaces and years of operating records.
From that starting point, Deduce Data Solutions developed Hefesto. It’s a system that learns from the furnaces’ real behaviour to anticipate and reduce their gas consumption, one of a forging plant’s biggest energy costs.
The system works alongside the existing SCADA, without replacing infrastructure. It’s organised into three layers: a predictive consumption model, a furnace-occupancy planner, and a deviation monitor for preventive maintenance.
The project was developed with support from the Government of Cantabria’s INNOVA 2024 programme, managed by SODERCAN. It was industrially validated across five furnaces on a main commercial forging line.
The first tests confirmed a high potential for energy savings and performance improvement. On top of that comes deviation detection that used to only be discovered when the bill arrived.
“Hefesto will not only let us reduce energy consumption and emissions, but also gain precision, reliability and operational control,” sums up Javier Cordón, Technology Director at Reinosa Forgings & Castings.
None of this required waiting for the plant’s industrial data to be ideal. It required understanding what was there, what was missing, and designing the system around that reality.
You can see the full technical detail in our success story on how AI reduced gas consumption in industrial furnaces.
When the data was already enough to decide better, not just to predict
In manufacturers of automotive components and tyres, DDS has come across the opposite challenge: the industrial data existed, but operational decisions were still being made on conservative, “by experience” criteria.
Which production sequence to use, how to distribute load between furnaces and cupolas, what trade-off to accept between cost, energy and lead time: all of that depended on individual judgement, not a quantitative comparison of alternatives.
With that same process data, DDS developed Pythia. It’s a system that simulates operating scenarios and quantifies their impact on cost, energy, lead time, quality and capacity before a decision is made.
Instead of delivering a single opaque recommendation, it shows the alternatives side by side and explains why one is preferable to another. The starting point wasn’t more sensors or more industrial data.
It was a layer of understanding built on the data the plant was already generating. It’s the same logic as this article’s checklist, applied to a different problem.
Common mistakes when assessing your industrial data
There’s a mistake we see repeated across very different plants: confusing “having lots of data” with “having useful industrial data”. A huge raw historical record, without traceability or context, can be less usable than a modest but well-documented one.
Another common mistake is treating data quality as a separate project, prior to and apart from the AI project. In practice, understanding a plant’s industrial data is already the project’s first phase, not an administrative requirement before starting.
It’s also common to overestimate the sensors needed. Before installing more sensors, it’s worth checking whether the ones that already exist, often underused within the SCADA itself, are enough to answer the real business question.
Finally, some organisations expect the AI team to “guess” what each variable means without involving anyone from the plant floor. Data with no one to interpret it is, in practice, only half the data.
What to do if your score wasn’t a 6 or a 7
If your checklist came out in the low range, the next step isn’t a large-scale AI project. It’s a short, scoped diagnosis of the industrial data you already have.
At Deduce Data Solutions we’re a scientific team specialising in Industrial Process Intelligence (IPI). We turn the data your plant already generates into physical, operational and actionable understanding, before proposing any model.
We work in brownfield mode, integrating with the systems that already exist on the plant floor. We don’t require large investments in new instrumentation.
Understanding before optimising always starts by understanding the industrial data that already exists, not by waiting for it to be ideal.
Before writing the first technical requirement of your next AI project, go through this article’s 7 questions with your plant team. Note down, honestly, how many “yeses” you get on your own industrial data.
That number, more than any commercial proposal, is the real starting point for your project.
Sources consulted
- MIT NANDA, “The State of AI in Business 2025” (The GenAI Divide), reported by Fortune, 18 August 2025
- McKinsey & Company, “AI data readiness: Foundation for scaling enterprise AI” (2026)
- Deloitte, “2025 Smart Manufacturing and Operations Survey”
- Forbes Technology Council, “Analysis Of The Dark Data Phenomenon In Industrial Practice” (Adrian Stelmach), 12 December 2025, citing Seagate’s “Rethink Data” report
- El Faradio, “Reinosa Forgings & Castings and Deduce Data Solutions complete the final phase of ‘GASAI'”, 2 September 2025
- Deduce Data Solutions, Hefesto/GASAI success story
Shall we talk about your data and AI project?
At Deduce Data Solutions we turn data into decisions for industry, with data governance and compliance from day one.
