“Our data isn’t clean enough.” “We don’t have enough data” “We need to digitize everything first.”
These objections come up time and again in our exchanges with industrial clients. Ultimately, they reveal a common misconception, that artificial intelligence requires a massive, clean, and perfectly structured database to work.
As a result, this bias slows down many industrial projects. In reality, industrial data is rarely perfect; it’s often partial, noisy, and heterogeneous. Yet, that doesn’t stop AI and data science from delivering value, quite the opposite.
Generate Data Rather Than Wait for It
Letting go of the assumption that AI requires a large set of pre-existing data is a key mindset shift. While this idea may be true in many common cases, where a model is trained on past data to predict output from new input combinations, it doesn’t apply universally.
In fact, in many industrial cases, the goal isn’t prediction but optimization under constraints (space, labor, capacity, machines, inventory, etc.). Rather than relying on historical data, these models use synthetic inputs, modeled to simulate the environment and its constraints. Within his virtually reconstructed environment, reinforcement learning or constrained optimization models can be applied. These approaches identify optimal strategies in scenarios too complex for human decision-making and cognitive capacity, due to the number of variables involved
Practical Example
•In fact, in many industrial cases, the goal isn’t prediction but optimization under constraints (space, labor, capacity, machines, inventory, etc.). Rather than relying on historical data, these models use synthetic inputs, modeled to simulate the environment and its constraints. Within his virtually reconstructed environment, reinforcement learning or constrained optimization models can be applied. These approaches identify optimal strategies in scenarios too complex for human decision-making and cognitive capacity, due to the number of variables involved
• Our team’s analysis focused on the factory’s bottleneck stations, where scheduling was done empirically based on operator experience, and therefore not optimized.
• The combinatorial complexity of this scheduling, due to the diversity of process flows and cycle times on resource-constrained stations, required the use of data science. It enabled the exploration of millions of scenarios in seconds to identify an optimal solution, far beyond human capabilities.
• By modeling the workstations, product references, and physical constraints, the manufacturer was able to unlock over 20% additional production capacity, without any additional CAPEX.
Imperfect Data Can Still Create Value
Even when data is partial, noisy, or heterogeneous, data science can extract value. It can help reconstruct consistent datasets from incomplete elements. AI models can detect anomalies in past databases, fill in missing data points in time, correct biases, or generate substitute data.
Practical Example
• A defense manufacturer wanted to improve the performance of material flows across its five sites and challenge its existing logistics organization.
• U The first step involved reconstructing all past logistics routes using partial data from the ERP system. The large data volume (~2 million movements), the number of possible movement types (40 tracked logistics actions), and the incomplete data (no differentiation between items with the same reference) made it necessary to automate the analysis. A tailored statistical model was developed to trace the path of each unique part over the past 10 years. .
• A supervised machine learning model, trained on the reconstructed data, was then used to classify all logistics routes as either normal or erratic. .
• The model detected over 15% of logistics routes as unjustified. Once again, the diversity and sheer volume of paths made such analysis beyond the reach of human capability alone. .
In short, data science can be fed incomplete data, process it for usability, and extract valuable insights.
Industrial AI: Applied Mathematics and Pragmatism
Industrial AI is not limited to the generative tools currently dominating the headlines. It’s first and foremost about applied mathematics and solving complex problems tailored to each industrial context.
In this sense, AI isn’t a turnkey solution, but a tailored analytical approach. True value comes from finely tuned algorithmic models that can handle complex, often non-reproducible situations. To succeed, this requires accepting uncertainty and favoring targeted experiments: quick proof of concept, value demonstration, and gradual deployment.
“Our solutions are neither LLMs nor chatbots. Unlike generic SaaS platforms, we develop tailor-made models designed to solve the most complex industrial challenges“
In the end, methodological rigor is key to turning ideas into real value. What matters isn’t the tool itself, but its alignment with real-world complexity
Conclusion: Industrial AI Deserves a New Perspective
The real value of industrial AI lies not in accumulating perfect data, but in its ability to model, enrich, and make smart use of the data available. Whether by building models with synthetic data to simulate and optimize strategies, or by extracting value from incomplete or noisy datasets, the opportunities are real and available now.
Instead of waiting for ideal conditions that may never come, it’s time to put AI to work where it’s most valuable: at the core of industrial complexity. That’s where AI delivers real impact on industrial performance.