The Addition Fallacy
Walk through any industrial technology conference today and the same three words appear on nearly every stage: data, science, and AI. Most organizations have responded the way org charts encourage them to, a data team building pipelines, an R&D or engineering function advancing domain science, and an AI program running use cases. The result, as is fragmented adoption: some functions are still standing up data foundations and governance, others are piloting use cases, and only a smaller set is driving toward scale, with siloed legacy systems among the persistent obstacles.
The spending, meanwhile, keeps rising faster than the returns. Deloitte's State of AI in the Enterprise research, based on a survey of 3,235 leaders, found that while two-thirds of organizations report productivity and efficiency gains from AI, only 20 percent are achieving revenue growth from it, against 74 percent that hope to. Activity is abundant; compounding value is rare.
The explanation, in our experience engineering AI systems for asset-intensive industries, is rarely that any single initiative is weak. It is that the three domains are managed as if they add. They don't. They multiply.
The Logic of Multiplication
Why multiplication? Because in an industrial environment, each domain is a precondition for the others to matter. A capable model trained on fragmented, ungoverned data produces confident answers no one can trust. Pristine data without scientific validation produces correlations no engineer will act on. Deep data and rigorous science locked inside decades-old monolithic applications produce insight that never reaches a workflow. In each case, the weak factor doesn't merely reduce the outcome, it zeroes it. A system that operators won't rely on delivers the same operational value as no system at all.
That is why the diagnostic question matters more than the investment question. Not "how much are we spending on AI?" but "which factor in our equation is closest to zero?"
Factor One: Data, The Most Common Zero
In industrial settings, the data factor fails in a specific way: subsurface measurements, equipment telemetry, operational records, and engineering documents live in separate systems, formats, and eras, and none of it was collected with machine consumption in mind. Gartner predicts that through 2026, organizations will abandon 60 percent of AI projects unsupported by AI-ready data, and reports that 63 percent of organizations either lack, or are unsure they have, the right data management practices for AI. AI-ready data is not a warehouse; it is data aligned to specific use cases, governed at the asset level, and continuously quality-assured. Everything downstream inherits this multiplier.
Factor Two: Science, The Factor Everyone Skips
Here is the factor that gets the least attention and deserves the most. In consumer applications, a wrong AI answer is an inconvenience. In industrial operations, it is a safety event, a misallocated well, or a capital decision built on a hallucination. The difference between a demonstration and a system people stake decisions on is not model size, it is scientific credibility.
The energy industry understands this better than most. The International Energy Agency notes that oil and gas has been an early adopter of AI precisely in its most science-intensive work, making resource evaluation more reliable, reducing predrilling uncertainty, and optimizing production. The sector's computational heritage tells the same story: IEA analysis shows oil and gas companies operated 24 of the world's 500 fastest supercomputers by 2024, with computing capacity growing nearly 70 percent annually ,capacity built overwhelmingly for scientific simulation, not chatbots.
That credibility must be engineered into AI systems deliberately. Models must respect physics, geology, chemistry, and engineering constraints rather than merely fit historical patterns. Their predictions must be validated the way scientific results are validated: against first principles, against lab results, against the judgment of domain experts who can articulate why an answer is wrong. An AI system that cannot survive that scrutiny will be quietly abandoned by the engineers it was built for, the most common and least reported failure mode in industrial AI.
The payoff of treating science as a multiplier rather than an afterthought is measurable. When a leading global oil services company needed to accelerate cement mix design, facing data overload, low model precision, and costly physical lab experiments, the answer was not a generic model but an AI-powered virtual lab grounded in materials science, built by FPT around the domain's own experimental logic. The result was 95 percent accuracy in thickening-time prediction and an 80 percent reduction in experimentation costs, performance that outperformed human engineers in most cases precisely because the AI was multiplying the science, not bypassing it.
Factor Three: AI, The Interface That Exposes Everything Else
The newest factor is best understood not as intelligence but as a new interface to an organization's data and science. A model can only reason over data it can reach; an agent can only call functions that have been exposed to it; both can only be trusted to the degree their outputs survive scientific scrutiny. This is why the AI factor so often reveals the zeros elsewhere in the equation: put a capable model on top of siloed data and unvalidated science, and it will surface those weaknesses at scale and with confidence. The constraint, in most industrial organizations, was never the model.
Engineering the Multiplication Sign
If data, science, and AI are the factors, the multiplication sign itself is engineering: the discipline that connects governed data to validated models to systems people use. The evidence that this connective work is the real constraint comes from the most neutral observer available: the IEA's analysis of why proven AI applications fail to spread across the energy sector cites barriers that are almost entirely system-level ,lack of access to data, interoperability concerns, gaps in skills, inadequate digital infrastructure, and resistance to change ,rather than shortcomings in the AI itself.
The prize for getting the system right is equally well documented. McKinsey's research on digital transformation in energy found that where digital has been applied successfully, companies achieved 2 to 10 percent improvements in production and yield and 10 to 30 percent improvements in cost, worth $2 to $12 per barrel in upstream operations. Those returns do not come from any single factor. They come from the multiplication.
The Executive Equation
The multiplication frame suggests a different set of questions than the usual AI checklist. Not "how many AI use cases do we have?" but "which factor is our zero, and who owns fixing it?" Not "which model should we buy?" but "would our engineers stake a decision on its output, and if not, what validation would change that?" Not "do we have a data team, a science function, and an AI program?" but "who owns the multiplication between them?"
Organizations that answer honestly tend to find that their constraint is neither ambition nor algorithms. It is the engineering that turns three respectable initiatives into one compounding system. In industrial intelligence, addition produces activity. Multiplication produces outcomes.
