Manufacturers are exploring artificial intelligence, but most are still struggling to scale it across operations. According to the latest Sikich Manufacturing Industry Pulse survey, 92% of manufacturers are actively exploring AI. Yet roughly three-quarters of respondents remain in research or pilot phases rather than deploying it in operations.

The National Association of Manufacturers reports that 51% of manufacturers already use AI somewhere in operations, and McKinsey estimates that only about a third of organizations have begun scaling AI at all. The funnel is remarkably consistent throughout industry: broad exploration, far narrower production.

Why AI efforts stall

Most manufacturers aren’t starting from scratch. Many already use AI tools such as machine learning for demand planning, predictive maintenance, machine vision, and automation.

Although AI adoption remains in its early stages for many manufacturers, investment intentions continue to move in the right direction. The percentage of executives planning major investments in AI increased from 28% in the Sikich 2025 Volume 2 survey to 31% in the latest research, reflecting a growing desire to turn exploration into action. The starting line is behind most companies; the finish line is what’s in question.

What slows progress is the transition from isolated pilots to scaled deployment. Here are the most common challenges.

Unclear ROI and trust gaps

Organizations hesitate to scale without results they can measure or trust—or, worse, they jump straight into production without understanding those results or how they intend to measure them.

Part of the problem is what gets measured. Pilots are often judged as technology experiments (i.e., answering the question, “Did the model work?”) rather than measuring them against business outcomes the organization already cares about. This leads to situations where pilots technically succeed while still failing to earn their way into production because no one can say what it was worth. In fact, earlier last year Sikich survey data indicated that 23% of manufacturers who started implementing AI stated that they hadn’t seen benefits.

Data readiness challenges

AI is only as effective as the data behind it. In this era of AI, context is king. Poor-quality data starve a model of the context it needs to be useful. Sikich found that only 56% of manufacturing executives have a data strategy in place that supports their business initiatives. A majority of executives (57%) report that lacking an internal team or resources to manage the data prevents them from implementing these strategies.

But readiness is widely misread as perfection. The goal is not a pristine, enterprise data estate before anything begins; it’s the right data you need to make the decisions or scope of decisions. When humans analyze a problem, we carry an innate sense of what information is missing, and we go looking for it. Building an AI solution means making that instinct explicit, defining the decision you’re trying to improve and the data those dimensions require. That is a data-strategy question as much as an AI question.

McKinsey is blunt about the alternative, warning that treating data quality as a “big-build” project that holds back everything downstream until the data look perfect leaves significant value unrealized. Quality is better fixed in targeted iterations tied to the use case in front of you.

Siloed initiatives

Pilots often live outside day-to-day operations, which limits their effectiveness. This issue is further compounded when the underlying data are similarly siloed.

There’s a subtler version of this problem, too. Many organizations now have capable teams building genuinely valuable AI solutions, but only in pockets, in ways that never scale for the rest of the business. Capturing that value without letting it fragment is a managerial discipline in its own right—we call it AI innovation management—and it’s one most manufacturers haven’t yet developed.

Talent constraints

Manufacturing teams are already juggling AI experimentation with keeping core systems running. The skills gap is the most-cited workforce barrier: 76% of Sikich survey respondents acknowledged that fewer than 10% of employees are training on or using AI.

The harder truth is that the specialized skills required to exploit AI are scarce, and it’s unrealistic to expect a procurement specialist, a manufacturing engineer, or a marketing lead to acquire them overnight. This barrier shows up in uneven engagement. Executives are typically far more engaged with AI than the factory-floor supervisors are. Progress stalls where the work actually happens.

Left unaddressed, these barriers keep AI stuck in testing environments rather than delivering measurable business value.

The shift from AI pilots to scaled deployment

Many organizations still approach AI as a “check the box” implementation rather than defining what it can and should achieve.

Initiatives that prioritize experimentation without a clear objective generate insights but rarely drive measurable outcomes. Scaling requires an outcome-first mindset: Start with a business problem worth solving, a specific decision, cost, or constraint. Then decide what solving it is worth and align the AI effort to that, rather than the reverse.

Putting the outcome first also changes what you measure; this is where many programs sell themselves short. The instinct is to measure efficiency, time saved on work the organization already does. That number is real, but it’s usually a smaller portion of the value. The larger return often sits in work that wasn’t happening at all: the analysis no one had time to run, the inspection that wasn’t economical at full coverage, the backlog quietly deprioritized because the effort never justified the payoff. AI changes that math.

The most useful questions to ask of a pilot are not only, “How much time did this save?” but also, “What are we now able to do that we simply weren’t doing before?” That’s the number worth taking to leadership, and the one that earns a pilot its place in production.

Although difficult to justify to the CFO, opportunity cost is very real. The AI transformation era is going to change the leaderboard in just about every industry, and today’s front-runners no longer have the clear advantage.

Discipline matters as much as ambition. BCG’s 2025 research found that the companies generating the most value from AI concentrate on fewer use cases than their peers and win by going deep rather than spreading thin. A few outcomes carried all the way to production beat a dozen pilots that never leave the lab or are limited to departmental use.

To move from pilots to production, manufacturers should prioritize five actions.

1. Define the outcome and what it’s worth. Start with a single business problem and put a number on it: the cost of unplanned downtime, the margin lost to scrap, the working capital tied up in the wrong inventory. Focus on an objective you can quantify; choose one of which you can later prove the value. If you can’t say what success is worth (usually in dollars and cents) and how you’re going to measure it, you’re not ready to scale it. Keep experimenting.

2. Pick a few high-value use cases, and right-size the data to each. Prioritize on three axes: the value of the outcome, the availability of usable data, and the feasibility of acting on the result. Then scope the data to that one use case rather than waiting for a clean enterprise estate. Most first wins don’t require a fully instrumented plant—predictive maintenance can start on the handful of machines causing the most downtime; demand forecasting can run on the order history already in the ERP; and knowledge assistants can work directly over the SOPs, manuals, and work orders a plant already has, with no data cleanup required.

3. Build data readiness and governance in parallel, not in sequence. Fix data quality iteratively, tied to the use case in front of you, and stand up governance at the same time. Size your governance efforts to the decision at stake. A tool that drafts a maintenance summary needs lighter oversight than one that releases a lot to a customer. The NIST AI Risk Management Framework and its Generative AI Profile give manufacturers a practical, widely adopted structure for this without inventing controls from scratch. Clear governance is also what closes the trust gap that keeps pilots from scaling.

4. Embed the pilot in the operation. Run it where the work happens, with the people who do the work, against the real workflow and the metric from the first step. A result that lives in a side system proves nothing about production.

We’re often tempted to experiment in the shadows to shield ourselves from failure. Well, if there’s no risk, then the pilot is likely not very valuable. Start with something manageable but important and, just as important, visible. In other words, start with something that matters. When it’s successful, it proves you can do something with AI that’s meaningful and either saves money, reduces real risk, has a positive effect on revenue, or all three.

5. Measure the result, then scale what worked. When the pilot ends, measure what it delivered and compare that to the target you defined in the first step. Count both kinds of value the pilot produced: the efficiency gained (i.e., time or cost saved on work you were already doing) and the value of new work the pilot made possible that you weren’t doing before. Add the two together and compare the total to the target. Scale the use cases that met or exceeded it, and pause or rework the ones that fell short. This keeps the decision to scale, grounded in measured value rather than in how promising the pilot looked.

None of this requires a manufacturer to build a data-science organization from scratch. One of the more useful patterns to emerge from the frontier AI labs is the “forward-deployed” model—embedding specialists directly alongside the people doing the work, rather than handing over tools and hoping the skills follow. Applied to manufacturing, it means teams that can refactor aging code, untangle data estates that quietly tax operations, sharpen the insights that decisions depend on, and keep governance in place so the work stays safe while it moves fast—building capability from the inside out. It’s an accelerator, and increasingly, it’s how the gap between a promising pilot and a production system gets closed.

From pilots to competitive advantage

Manufacturers recognize the need for AI; the challenge is how to scale it. Many have launched pilot programs, but far fewer have translated those efforts into consistent operational impact.

Organizations pulling ahead aren’t the ones running the most pilots. They are the ones that tied AI to outcomes worth measuring, scoped the data to those outcomes, and scaled only what proved its value.

The spread is measurable. BCG’s 2025 analysis found that the companies it identifies as AI leaders are increasing revenue at roughly 1.7 times the rate of laggards and delivering more than three times the total shareholder return. Research from MIT Sloan points in the same direction in manufacturing specifically: Adopters often dip before they climb, but those who push through pull ahead in productivity and market share.

The competitive divide will be defined by execution, not experimentation. Manufacturers that scale AI into core operations—measured by the value it creates, including the work that was never getting done before—will outpace those still asking whether their data are ready. The data rarely get perfect. The advantage goes to the companies that start with a problem that matters. The data must be better, and they must be suitable for the work. But they don’t need to be perfect.