Manufacturing companies across the globe are investing in AI as they look to futureproof their operations and adapt to the inevitable changes coming in the industry. However, this brave new world of AI comes with some challenges from the old including predictive maintenance, fault detection on production lines, digital twins of plants and processes. McKinsey's 2025 State of AI survey found that 88% of organizations now use AI in at least one business function, up from 78% the year before, but only 39% report any measurable enterprise-level EBIT impact from it, and just 6% describe themselves as capturing significant value. The gap between adoption and outcome is not a technology problem. It is a data problem.

In manufacturing, where data is fragmented across machines, operators, and disconnected legacy systems, the risk of poor data readiness is amplified. Most manufacturers are not failing at AI because they chose the wrong model. They are failing because they are trying to build intelligence on a foundation of noise.

Where AI has worked, and where it has not

Most of the AI value captured in manufacturing to date has happened at the machine level. Things such as computer vision systems inspecting body panels and welds on automotive assembly lines, acoustic and vibration models flagging bearing wear before equipment failure, and vision-guided pick-and-place in electronics assembly. These are bounded problems, with rich sensor data, fast feedback loops, and a clear definition of success. The technology has matured, the ROI is well-established, and most large manufacturers have at least pilot deployment

The harder problem, and increasingly the binding constraint on plant performance, is coordination at the human level. A defect caught by a vision system still needs a maintenance technician dispatched, a quality engineer briefed, a production planner re-sequencing the line, and a shift handover that does not lose the context. In high-performance environments running lean methodologies, the obeya, the daily stand-up, the action log and the cross-functional escalation are where minutes and margin are won or lost. This is where AI adoption is now heading. AI is not replacing the operator, but compressing the loop between detection, decision, and action across teams.

That shift changes what AI needs from the data layer. A vision model on a single line needs clean images and labelled defects. An AI assistant helping a shift leader prioritize their morning needs production data, quality events, maintenance status, supply constraints and the previous shift's open actions, all reconciled, all current, and all in the same vocabulary. The first is a perception problem. The second is an integration problem that most manufacturers are not ready for.

The data reality on the shop floor

Production lines generate enormous volumes of real-time sensor data, yet frontline teams running them still capture critical process knowledge in spreadsheets, paper logs, and tribal memory. Meanwhile, the systems meant to manage all this run in parallel with different definitions, different formats, and different owners. Quality, maintenance, supply chain systems, manufacturing execution systems (MES), quality management systems (QMS), computerized maintenance management systems (CMMS), enterprise resource planning (ERP) and a long tail of point tools, each operate on their own silo, rarely speaking the same language.

A “defect” in one system is an “incident” in another. A downtime event is logged automatically at one site and manually at another. When these islands of information cannot speak to each other, the humans coordinating across them spend their time reconciling reports rather than acting on them, and the AI models built on top inherit the same confusion.

According to KPMG's Intelligent Manufacturing 2025 report, 56% of manufacturers identify data challenges as their primary barrier to AI adoption, while 52% report insufficient integration between systems. Most manufacturers know this. Few have a clear path to fixing it.

The real question has shifted

The industry has started asking a harder question. Not “how do we deploy AI?” but “how do we make our data usable for AI in the first place?”

That shift matters because manufacturers are beginning to understand that AI is a multiplier, not a repair tool. Feed it clean, connected, contextual data and it delivers. Feed it chaos and it amplifies the chaos, and once that chaos is encoded into automated decisions, it becomes considerably harder to unwind.

What AI-ready data actually looks like

There are three levels of data maturity in manufacturing operations:

1. Chaos. Data lives in spreadsheets, local servers, and handwritten notes. No single source of truth. AI in this environment is counterproductive.
2. Consistency. Each department has its own digital system, well-organized locally but invisible to the rest of the business. This is where most manufacturers sit today.
3. Intelligence. Data flows across systems in real time, with shared definitions, clear ownership, and automated capture. This is the foundation from which AI can actually learn and act.

Moving from the first stage to the third is not about buying more software. It is about connecting and standardizing what already exists, typically through a unified data layer that sits over the operational systems and provides a shared semantic model. That way an event in the MES, a measurement in the QMS, and a work order in the CMMS can be reasoned about as part of the same production reality.

That unified layer is itself evolving as until recently, integration meant hand-built pipelines between each pair of systems, expensive, brittle, and a source of ongoing maintenance debt. Emerging standards like the Model Context Protocol (MCP) are changing this. They give AI agents a common way to read from and act on the operational systems around them, without custom integration code for each. An agent reasoning about tomorrow’s production schedule can pull current work-order status from the CMMS, open quality events from the QMS, and live throughput from the MES through the same interface. The importance is not the protocol itself, standards evolve, but the pattern. Coordination across systems is moving from a humans-in-spreadsheets problem to an agents-with-shared-interfaces problem, and the manufacturers who get their underlying data in order will be the ones able to exploit it.

Four practical steps start that journey. First, establish a shared taxonomy: consistent labels and data definitions across systems and sites. Second, assign data ownership: someone accountable for accuracy and consistency in each domain. Third, automate capture where possible to reduce manual input and the errors that come with it. Fourth, attach operational context, batch, equipment, shift, product variant, to every data point at the moment of capture. Without this context, sensor streams are telemetry; with it, they become training data.

Getting the foundation right before scaling

Manufacturers who are seeing real AI returns share a common starting point. They integrated production and quality data streams before training models, cleaned historical datasets before deploying AI, and built small, high-trust use cases before scaling across sites.

Manufacturers in heavy industry and pharma have reported double-digit reductions in unplanned downtime not by deploying sophisticated models, but by first connecting maintenance and production logs so events in one system can be reasoned about against state in the other. Throughput gains have come from similarly unglamorous work, using cross-system analytics to identify redundant inspection steps that no individual function had visibility on.

These are not futuristic projects. They are the result of getting the basics right.

How this changes the work

The first wave of manufacturing AI changed what machines could see. The next wave will change what teams can coordinate. Shift leaders who today spend the first 90 minutes of a shift reconciling reports from four systems will increasingly start their day with a synthesized brief to find out what changed overnight, which open actions are at risk, and where the biggest constraint is likely to be by 11 a.m. Quality engineers will spend less time chasing data and more time investigating the root cause. Maintenance planners will work from a prioritized list informed by both equipment condition and production demand, rather than a queue.

None of this removes the operator, the engineer or the planner from the loop. It removes the coordination overhead that currently sits between them and their actual job. That is a meaningful change to the working day, and it is one that becomes possible only when the underlying data is integrated enough to support it.

The opportunity

Manufacturers must make a choice to proceed with fragmented data and accept fragmented results, or take the time to structure, connect and standardize the data foundation first and give AI something worth working with.

Most manufacturers are sitting on enormous amounts of data. The ones who will win with AI are not the ones with the most data. They are the ones who can actually use it.

That window to get ahead is open right now, but it will not stay open. The technology is ready. The data, in most cases, is not. That is the gap worth closing.