Where should your factory invest next in digital transformation?

A factory spends heavily on automation and still cannot see, during the shift, why a line stopped. Another connects most of its machines and still builds the week’s plan in a spreadsheet. Another implements an execution system and then retypes quality and maintenance into the ERP. Another starts an AI project and discovers the history underneath it is too thin to trust.

None of those plants picked a foolish technology. They picked it before the operation was ready to use it. Digital transformation in manufacturing is no longer a debate about whether to invest. The argument, on the plants I see, is where that money should go first, and what has to exist before the spend pays back.

I have written before that the process has to be understood before the software. This is the investment version of the same point. The shopping list is long. The sequence is the decision.

The money is already moving

Deloitte’s 2025 Smart Manufacturing and Operations Survey asked 600 manufacturing executives, mostly at large companies, how they were spending and what they expected to spend next. Ninety-two percent said smart manufacturing would be the main driver of competitiveness over the next three years. Seventy-eight percent were already putting more than 20 percent of their improvement budget into it, and 88 percent expected that investment to hold or rise in the following fiscal year.

That money is not going into one box. It is spread across automation hardware, sensors, vision, machine connectivity, networks, cloud and edge, analytics, planning and execution systems, quality, energy, cybersecurity and AI. Deloitte’s own reading of the next 24 months is the useful part. Among digital technologies, 40 percent of respondents ranked data analytics as a first or second priority, 29 percent cloud computing, 29 percent AI and 27 percent industrial IoT. On the physical side, 41 percent prioritised factory automation hardware, 34 percent active sensors and 28 percent vision systems.

Read those two lists together and the order is fairly plain. People are still buying machines and sensors. They are also paying to capture, connect and understand the data those machines produce. They are not, as a group, jumping straight to AI and hoping the foundation appears later. The survey says as much: manufacturers are building the base that makes later decisions possible.

Capital and subscriptions are now the same project

Factories are comfortable with capital projects. A new press, a robot cell, a vision station, a network cabinet, an edge computer, a warehouse crane. For a long time, software tried to look like that too: a licence, some servers, industrial PCs, a network, an implementation, and then a quieter year until the next upgrade.

That shape is less common now. A plant can rent the application, the cloud, the analytics, the connectivity and, increasingly, the AI. Support, training and the people who keep the integration alive were never really a one-time cheque. The result is a project that starts as capital and continues as operating spend. Buy the machine. Fit the sensors and the network. Put in an application, whether that licence can be capitalised or not depends on the cost and the accounting rules, not on the fact that the server sits in the plant. Then pay, every year, for support, cloud analytics, model use and someone to watch the connection.

So “how much capital do we have for digital?” is a narrow question. The better one is what capability the operation is missing, and which way of paying for it matches that capability. A sensor on a machine and a subscription that reads it are not rival philosophies. On a real line they are often the same job, split across two budgets.

Where the investment is actually going

Automation, then the question of whether anyone can see it

Automation is still the largest appetite. PwC’s 2026 Global Industrial Manufacturing Sector Outlook, based on 443 senior executives across 24 territories, found that the median share of industrial manufacturers with highly automated processes is expected to increase from 18% today to 50% by 2030. Labour is part of the reason. In Deloitte’s survey, nearly half of the respondents reported moderate or significant trouble filling production and operations roles, and a similar share said the same for planning and scheduling.

A robot can take a repetitive task off a person. It does not, by itself, tell a supervisor why the cell beside it is waiting on material. Automation without a record of what happened is a faster way to be blind.

Connectivity, which is less glamorous and more necessary

The next layer is dull and decisive. Machines have to report. Counts, stops, reasons and quality events have to leave the HMI. At facility or network level, Deloitte found 57 percent of respondents using cloud computing, the same share using data analytics, and 46 percent using industrial IoT. That is the base. If the history is missing or cannot be combined, the analytics number is a tool looking for a fact.

Software that has to run the day, not only store it

Once some data exists, the plant has to act on it before the shift ends. The survey’s highest system priorities for the next two years were advanced production scheduling (35 percent ranking it first or second), execution systems (33 percent) and quality management (28 percent). Planning, the shop floor and quality. Not a new logo for the same monthly pack.

This is the move from collecting data to running with it. An execution system that the supervisor updates after the fact is a more expensive logbook. I have written about what happens when the plan changes and the factory cannot change with it. Software only helps if the people on the shift can see the change in time to do something.

Analytics, which is where the data starts to answer a question

Data analytics was the top digital technology priority in the survey, at 40 percent. The questions are ordinary and hard. What happened on that order? Why did it happen? What is happening now? What is likely to happen before the truck leaves? What should someone do? A dashboard that cannot answer one of those, for one line, is not intelligence. It is a screen.

AI, which is real and still not the first step for most plants

AI is getting the attention. Rockwell Automation’s 2025 State of Smart Manufacturing research found that 95% of manufacturers had either invested or planned to invest in AI/ML, generative AI or causal AI over the next five years. At the same time, 56% were piloting smart manufacturing and 20% were already using it at scale. Deloitte’s State of Generative AI in the Enterprise describes the same pattern outside the factory: pilots are spreading faster than use at scale.

A plant with a fifth of its machines connected, downtime on paper and planning in Excel will usually get more from fixing those three than from a model. A plant with years of history it can actually combine, and with planning and execution already digital, is in a different conversation. The technology can be the same product. The readiness is not.

Factories do not start from the same place

There is no trustworthy global count that sorts every plant into “just starting”, “in the middle” and “largely there”. I will not invent one. Publishing “half are starting, a tenth are mature” would be a guess wearing a percentage.

PwC India’s research shows that manufacturers are at different stages of digital transformation. Its digital-factory research found that 38% of firms surveyed were yet to create a digital-transformation roadmap, while 54% showed an upward implementation trend towards analytics and AI. The research also highlights the importance of building the digital backbone and aligning digital transformation with business strategy. Deloitte’s 2025 operations survey still finds comparatively low maturity in maintenance, material management and the workforce that has to live with the new tools. Starting is common. Being finished is not.

For thinking about the next investment, three pictures are enough. They are a way of looking, not a benchmark.

A plant that is still mostly manual

Paper and Excel still carry production, quality or the plan. Few machines report on their own. Systems, if they exist, do not talk. Traceability is a file. Maintenance is a breakdown or a calendar. KPIs arrive after the shift, sometimes after the month. The useful order here is dull: digitise the records people already keep, connect the machines that matter, capture the events, and make one process the same on every line. AI is usually a distraction at this point. There is nothing solid for it to learn from.

A plant that has bought the systems and not yet joined them

ERP is in. Maybe an MES, a warehouse system, a quality system, some SCADA, a dashboard nobody opens during the shift. Machines are partly connected. The definitions of a good day differ by department. Someone still exports a file to make the plan match the floor. Pilots worked on one line and stopped there. This plant has already spent the money. The next investment is often the unglamorous one: make two systems share one fact, cover the machines that still need a person with a clipboard, and stop treating a pilot as a strategy. The value is in what is already owned.

A plant that can see the operation while it is happening

Connectivity is high. Business and manufacturing systems exchange data. Execution is digital. Traceability runs from material to shipment. Planning can see a change. Maintenance and quality are not only after-the-fact. History can be found. Here the question changes. It is no longer “how do we get this off paper?” It is “which decision should get smarter?” Prediction, a tighter plan, energy against the part, the same pattern on the next plant. AI can belong here, aimed at one decision, not at the idea of AI.

The factory next door is not your sequence

This is where the copied project goes wrong. One plant hears about AI. Another about an MES. Another about a digital twin. All of them might be right. All of them might be early.

Take a plant where about a fifth of the machines are connected, production is typed in at the end of the shift, the plan lives in Excel and downtime reasons are on a sheet by the machine. Connectivity, a trustworthy capture of output and stops, and a view of production during the shift will change more than a model will.

Take another where most machines already report, and ERP and MES are both live, but planning, quality and maintenance still meet in a spreadsheet. The missing piece is not another system. It is the joint between the ones they have, and a definition of the numbers that more than one department will use.

Take a third where machines, systems and history already line up, and planning and execution are digital. The next money can go to prediction and to decisions the plant already makes badly or late. Same catalogue of products. Different place in the queue.

A level on a chart does not tell you what to do on Monday

“Level 3 out of 5” is a comfort and a dead end. It does not say whether the bottleneck is the machines, the plan, the material, or the fact that three systems each hold a piece of the same order. The question worth paying for is where the next investment should go.

That means looking at the plant as it runs, not as a technology stack. How many machines are connected, and does anyone trust the data. How the systems are joined. How production is released and recorded. How far ahead material can be seen, and how often the plan breaks. Traceability, quality, maintenance, energy. Whether KPIs are used while the shift can still act. How much history exists, and whether it can be combined. How ready the organisation is to spend, and when it expects to start.

The order of those investments should follow the gap, the size of the business effect, whether the plant is actually ready, what has to exist first, and the timing. Technology comes after that, not before. I have looked at the costs that sit inside the process and never quite make the accounts. Those are often the reason a connectivity or planning project is worth more than it looks on a software slide.

And AI

AI should be in the conversation. It should not be the default recommendation. It gets more useful as the plant moves from a recorded fact, to a fact with context, to an analysis someone acts on, to a prediction, to a decision. A machine with no usable history is not a predictive-maintenance candidate. A machine with a clean stop history might be. A plant that can put production, quality and maintenance on one timeline can support a harder decision. A plant that is already digital in planning and execution can ask a model to help. The question is not whether you want AI. It is whether this factory can create value from it yet.

Sometimes the right recommendation is to buy nothing new. Fix the foundation. That is an uncomfortable thing for a proposal to say, and it is often the correct one.

Where this factory should invest

There is no universal next project. It might be connectivity. It might be getting production off paper. It might be planning, traceability, quality, maintenance, energy, or joining two systems that already exist. It might, later, be analytics or AI. The only general rule I trust is that the next investment should match the hole in the operation, not the hole in a trend report.

I built the Factory Digital Investment Assessment for that question. It does not sell a product. It asks how the plant works today: how large the operation is, which systems are in use, how connected the factory is, how quality, materials and traceability are managed, how far material can be seen, how production is planned, how execution, maintenance and delivery are run, how energy and KPIs are used, how far the data can be analysed, how much of the factory is visible in real time, how ready the next investment is, and which problems should be solved first.

At the end you get a score and three short lists: what to do now, what belongs in the medium term, and what can wait. Work that is already in place stays off the page. You can download it.

Take the assessment

Digital transformation is not one project. Connect the machines that matter. Capture the record. Join the systems that force people to type the same fact twice. Digitise execution. Make the plan see the material. Let the shift see the operation. Then, if the history is real, predict and recommend. The plants that get value are not always the ones that buy the most. They are the ones that fund the next capability, not the next fashion.