One source of truth is the wrong target

The phrase one source of truth promises order. In a factory, however, it can encourage the wrong ambition. A plant may need a machine identifier that makes sense to maintenance, a product definition that satisfies quality, and an enterprise code used by finance. Forcing every function and site into one system can destroy useful local context without creating better decisions.

The real objective is not one database or one universal version of every record. It is a governed way to recognize when records describe the same asset, material, customer or event, determine which is reliable, and trace how it was created. Manufacturers need agreement on meaning more than uniformity of storage.

The factory does not need one database

Industrial companies rarely begin with a blank architecture. Plants have acquired equipment, historians, MES platforms, ERP systems and specialist applications over decades. Mergers add further duplication, while individual sites protect systems that continue to run production effectively. Replacing everything may satisfy an architectural diagram, but it can introduce cost and risk without improving the operation.

At Ataccama, Chief Product Officer Jay Limburn sees data modernization as selective rather than absolute. Moving information into cloud platforms helps only when what arrives there can be trusted. “There is no magic button that you can press,” he says. “There is a lot of legacy data and complexity across manufacturing, but the companies that think about that data estate and its modernization are the ones that are going to be successful with AI. If you still have disjointed, not well understood, not well cataloged and not well profiled data, there is a huge amount of risk.”

One Ataccama manufacturing customer discovered that 20 percent of millions of CRM records shared the surname Customer. A default value had been used whenever the field was incomplete. A catalog could describe the table accurately while missing that much of it was unusable.

“Having a context layer that explains what data is used for, how it can be used and how to describe it is important,” Limburn continues. “But it also needs trust signals. An agent must be able to see that one customer table is trustworthy because its lineage and quality are understood, while another is not trustworthy because names are incorrect or values are missing. Without that quality dimension, you have a problem.”

Translation matters more than standardization

The tension becomes most visible between corporate functions and individual plants. Headquarters wants comparable performance across a global network, while a site manager sees little reason to replace an MES or naming convention that supports reliable production. Both positions can be legitimate.

A British factory may use one term for an asset that a German plant describes differently. The enterprise needs a taxonomy that recognizes the equivalence while allowing each operation to retain the local language and system that people understand. This is where Levent Ergin, Chief Strategist for Agentic AI, Regulatory Compliance and Sustainability at Informatica, believes technology should act as a translation layer.

“Someone on the factory floor is worried about running that factory like fine clockwork, and they do not want to make changes to the way the production line runs,” Ergin says. “At group level, the company still has to consolidate information from 20 or 30 factories. You need a taxonomy or translation table that says one factory calls an asset this, while another factory calls the same asset something else. Then you can make sense of the operation at group level.”

This is a practical interpretation of master data management. The manufacturer identifies domains such as assets, materials, products, suppliers and customers, then reconciles records within each. It does not have to standardize the entire enterprise at once.

Integration makes data reachable, cataloging shows what exists and lineage reveals where it came from. Governance identifies the critical elements and quality rules test whether they are valid before mastered records become trusted data products.

“Companies have finite resources, finite time and finite costs, so they have to focus on the outcome they are trying to achieve,” Ergin explains. “They need to identify the critical data elements, define the quality requirements and test whether the source systems are adhering to those rules. Once that has been done, they can create the most accurate version of the truth and make it available to the business.”

Governance must become an enabling service

Data governance has often failed because it was experienced as a central function saying no. Definitions were debated, access moved slowly and the business found workarounds. AI is changing that relationship because operational teams need trusted information quickly enough to move beyond contained pilots.

Limburn has seen governance teams reposition themselves as data-enablement organizations. The work still covers classification, quality, ownership and control, but success is how quickly an engineer, analyst or application can locate the right data and understand its limitations.

The same shift is visible from the infrastructure side. Sunitha Rao, General Manager of Hybrid Cloud at Hitachi Vantara, describes governance as a continuous operational foundation rather than a compliance exercise. “Manufacturers are gaining value not simply by collecting data, but by making it usable,” she says. “They must focus on how trusted the data is, how connected it is and how actionable it is. When those three things come together, data from factory systems, applications and the supply chain can support operational decisions rather than ending in another monitoring dashboard.”

That foundation must span hybrid environments. Low-latency or sensitive data may remain on premises, model training may use scalable cloud resources, and edge systems may continue making local decisions when connectivity is interrupted. Governance cannot depend on moving everything to one place.

“The goal is to place data and workloads for the right reasons, whether those reasons are performance, cost, security or compliance,” Rao says. “Manufacturers succeed when they can manage data consistently across edge, on-premises and cloud environments. That flexibility allows AI to scale while the business retains the operational control required for each workflow.”

AI raises the cost of ambiguity

Poor data has always created waste, but AI can execute an error at a speed and scale that human-led processes rarely achieve. An analyst may notice that an order for 10,000 units looks unusual when the normal quantity is 10. An automated workflow can accept the value, create the purchase order and leave the warehouse to discover the mistake when the material arrives.

Traditional applications follow explicit rules; an AI system infers what to do from the information and context it receives. Greater authority makes lineage, quality and access controls more important.

Ergin gives the example of a component used in several finished products. Without a mastered relationship between the raw material, bill of materials and product records, an agent may treat the component as relevant to only one line. The result could affect procurement, traceability or regulatory disclosure even though each source system appears internally consistent.

The physical consequence is most apparent in machining. Stephen Graham, Vice President of Product and Technology for Hexagon’s Production Software division, works with manufacturers whose data must ultimately guide machines cutting real material. “Making an impressive AI demonstration is one thing, but when it meets the reality of a shop floor, which is unpredictable and highly variable, there is a big difference between what you can show and what you can actually do,” he says. “Bringing the data together consistently is a prerequisite, but the workflow and the domain expertise around that data are just as important.”

The operational workflow is the final test

A trusted data foundation proves its value when it improves work without asking people to abandon the knowledge and systems that keep production safe. Strong deployments introduce recommendations where engineers can verify them and gradually delegate lower-risk decisions.

Hexagon has followed that pattern in computer-aided manufacturing. Rather than taking a CAD model and automatically producing final machine code with no intervention, its AI capability supports programmers within the existing CAM workflow. The engineer can invoke assistance for a defined task, inspect the result and continue to the next stage.

“You could put a CAD model in one end and have G-code come out at the other, but nobody is interested in that,” Graham says. “What works is supporting the CAM engineer as they go through the workflow that exists today. At each step, the engineer can invoke the AI to complete a specific task, see the output, decide that they are happy with it and move on. They want to stay in the loop and use the tools to accelerate the work incrementally.”

That principle applies beyond CAM. Different functions and sites do not need identical systems, but they need a dependable way to recognize the same business object, assess the quality of the information and understand which decisions it can support. A data product should carry its meaning, lineage, quality and usage rules wherever it is consumed.

The manufacturer that wins will not be the one that forces every plant into a single repository. It will be the one that preserves local operational context while creating enterprise-level trust. One source of truth is an appealing slogan, but the more useful target is a network of governed, translated and traceable truths that people and machines can use with confidence.