AI exposes what manufacturers do not know about themselves
Artificial intelligence is often presented as a technology challenge. Organizations debate which models to use, how much computing power they need, and which use cases will generate the fastest return on investment. Yet as manufacturers move from experimentation to deployment, many are discovering that AI is exposing weaknesses that have little to do with the technology itself.
Many enterprises are discovering that AI has an uncomfortable habit of exposing problems that existed long before the technology arrived. Questions around ownership, accountability and fragmented information often surface quickly once businesses begin trying to deploy AI at scale. Levent Ergin, Chief Strategist for Agentic AI, Regulatory Compliance and Sustainability at Informatica from salesforce, believes the organizations making the fastest progress are usually those with the clearest understanding of how their business operates.
“Companies still have a lot of data silos, and it’s not just a technology thing, it’s actually an operating model thing,” he says. “When you think about an operating model, it’s people, process and technology. People naturally feel comfortable working within their own silos and controlling information within their own teams. The companies getting this right have top-down executive sponsorship and treat data as something owned by the business, not something owned by technology.”
The observation helps explain why AI projects can produce very different outcomes despite using similar technologies. While much of the discussion around AI focuses on algorithms and infrastructure, many of the barriers to success remain organizational.
Data silos are organizational silos
Manufacturers have spent years trying to break down the barriers between systems, departments and functions. The ambition of creating a single view of operations is hardly new. Yet despite significant investment in digital transformation programs, fragmented information remains one of the industry’s most persistent challenges.
Ergin believes the reason is straightforward. Most data silos are not created by technology. They are created by the way companies operate. “The capabilities and maturity within an organization start at the top of the house,” he says. “It means putting money where your mouth is, having the right leadership and creating the capabilities needed to treat data as a business asset. If you don’t do that, the technology alone won’t solve the problem.”
The challenge becomes particularly visible when enterprises begin deploying AI. Traditional reporting systems can often tolerate duplicated records, inconsistent definitions and fragmented ownership. AI systems are far less forgiving because they depend on consistent information and clear context. “If we think about data as fuel, then AI is the engine,” Ergin explains. “Data and context about the data, which is really metadata, are what power that engine. If the context isn’t there, the AI doesn’t understand what it’s looking at.”
That issue becomes increasingly important as manufacturers look to deploy agentic AI and autonomous workflows. An AI system may be able to access large volumes of information, but unless it understands the relationships between products, suppliers, assets and processes, the quality of its decisions will inevitably suffer.
AI only knows what you teach it
One of the most useful examples Ergin provides comes from the supply chain.
A manufacturer may use the same microchip across multiple products. That relationship might be obvious to the people managing procurement, product development or production planning. For an AI system, however, that context must be explicitly available. “If you have a microchip that is used in multiple products, the AI needs to understand that relationship,” he says. “Without that context, an agent or AI workflow might get things wrong because it doesn’t fully understand how the business operates.”
The same principle applies throughout manufacturing. Information that appears self-evident to experienced employees often exists only as tribal knowledge held within departments, plants or business units. Humans compensate for missing context through experience and intuition. AI cannot.
This is one reason why Ergin believes discussions around AI readiness increasingly need to move beyond data quality alone. “We talk a lot about clean data, but I don’t think that’s enough,” he says. “The challenge is making sure data is trusted, traceable and contextualized. People consuming that information need confidence that it will lead to the right outcomes.”
That confidence becomes particularly important when AI is used to support operational decisions. Whether the use case involves procurement, customer service, quality management or production planning, organizations need to understand where information originates, how it has been transformed and whether it can be trusted. Otherwise, the old principle of garbage in, garbage out remains just as relevant in the age of AI as it was in the era of business intelligence.
Why a single source of truth remains elusive
The challenge becomes even more complicated as companies grow. Mergers and acquisitions remain one of the biggest obstacles to creating trusted data foundations. Every acquisition brings new applications, processes, naming conventions and operating practices. Over time, manufacturers can find themselves operating multiple ERP systems, procurement platforms and product databases simultaneously.
Ergin illustrates the problem with a simple customer example. “Let’s say Mark is a customer of Company A, and Company A acquires Company B,” he says. “Mark is also a customer of Company B, but the address is different in each system. Which address is correct? The process of deduplicating, matching and merging that information to create the most accurate version of the truth is what we call master data management.”
The same challenge exists across products, materials, suppliers, assets and employees. It becomes even more difficult when organizations attempt to consolidate information from multiple factories operating in different countries.
A plant manager focused on keeping production running smoothly may see little value in changing local systems that already work effectively. At a corporate level, however, leaders need a consistent view of performance across dozens of facilities. “The solution has to act as a translation layer,” Ergin explains. “One factory may call an asset one thing, while another factory calls the same asset something else. You need a common taxonomy that allows you to understand those differences and create a single version of the truth at a group level.”
The objective is not necessarily to force every site to operate in exactly the same way. Instead, it is about creating enough consistency that information can be understood, analyzed and acted upon across the wider business.
Start small and scale deliberately
Despite the complexity of these challenges, Ergin is not an advocate of large-scale AI programs launched across the entire enterprise from day one. The companies making the most progress are often taking a more measured approach. They begin with clearly defined internal use cases, establish measurable objectives and focus on proving value before expanding further. “A lot of AI still comes down to people,” he says. “Companies that are doing well are experimenting with internal use cases first. They’re asking how they can pick a low-risk use case that staff can try before exposing customers to it.”
This approach allows organizations to develop governance frameworks, establish confidence in the data and refine the operating model before introducing additional complexity. It also creates a clearer understanding of what success looks like. “One of the biggest mistakes is not defining the right KPI from the beginning,” Ergin says. “If the use case works well, what impact does it have on the business? Does it improve efficiency? Does it reduce costs? Does it affect profit and loss? Those questions need to be answered early.”
The same philosophy applies to scaling. Rather than attempting to transform every process simultaneously, successful enterprises build momentum gradually, expanding from proven use cases into adjacent areas. “We support millions of transactions and large-scale deployments, but scalability isn’t really the challenge,” Ergin adds. “The challenge is proving the use case, taking on the next problem and making sure you have the change management capability to bring people with you.”
That final point returns to the theme running throughout the discussion. While AI may be one of the most significant technological developments facing manufacturers, technology alone rarely determines success.
“If you don’t understand your business today, it doesn’t matter what investments you throw at it,” Ergin concludes. “A lot of AI comes down to people and process. The organizations succeeding are the ones that understand how they operate, invest in the right foundations and then apply the technology to solve real business problems.”
As manufacturers continue their race to deploy AI, that may prove to be the most important lesson of all. Technology is often capable of far more than organizations realize. The real question is whether the business understands itself well enough to use it effectively.
