Everyone Wants the Last Domino
I first heard Walker Reynolds explain industrial digital transformation through a simple progression:
Connect → Collect → Store → Analyze → Visualize → Find Patterns → Predict → Report → Solve
I liked it immediately because it gave industrial data a job to do. The framework begins with the unglamorous work of connecting equipment and collecting information, then follows that information until it contributes to solving an actual problem. It is useful because companies have a habit of buying the most visible technology first and discovering the missing foundation later, usually after the software has been installed, the consultants have left, and someone finally asks why half the machines are still sending data with timestamps from 1970.
One of Walker’s underlying points is especially relevant today: AI does not need to appear as a distinct stage. It can help analyze data, discover patterns, make predictions, produce reports, and support problem-solving. AI is one of the tools used throughout the progression. If the assets are disconnected, the data is unreliable, and nobody understands what the signals represent, giving the problem to a larger model simply creates a more sophisticated way to be wrong.
Walker’s framework was the inspiration for mine, although I am applying the idea to a different question. His progression helps explain how industrial data eventually contributes to solving a business problem. I wanted to describe how an industrial organization builds the capabilities required to move from basic connectivity toward intelligent decisions, automated execution, and eventually autonomous operations. That led me to this:
Connect → Collect → Contextualize → Govern → Understand → Predict → Decide → Automate → Autonomize
I visualize the progression as nine standing dominoes, with every domino larger than the one before it. The size represents the potential value of each capability, although there is a slightly annoying truth hiding inside the picture: the largest domino depends completely on the smallest one. Everyone wants autonomy, adaptive operations, self-optimizing production, and intelligent systems making decisions in real time. Almost nobody gets excited about connecting an old machine tool or fixing an asset hierarchy, or deciding who is actually responsible for the data once it begins moving everywhere.
The Nine Capabilities
These nine capabilities describe how an industrial operation moves from simply accessing data to using it to make and execute increasingly intelligent decisions. Each one builds on the capabilities before it, which is why I visualize them as dominoes that become progressively larger and more valuable. The early stages create visibility and meaning; governance establishes trust, control, and accountability; the later stages turn that understanding into prediction, decision, action, and eventually autonomy. AI can contribute throughout the progression, but it cannot compensate for missing foundations. The sequence matters because the final capability only works when the entire chain works.
Connect
Connect means establishing reliable communication with machines, sensors, controllers, applications, and other sources of operational information. In a real factory, this often involves equipment built across several decades, multiple vendors, inconsistent protocols, and at least one machine that everyone treats like a sleeping bear because it works today and nobody wants to be responsible for waking it.
Collect
Collect means capturing the data consistently enough to create a trustworthy operational history. Frequency, timing, completeness, and quality matter because a few isolated readings may be interesting, while a reliable stream of data can reveal how the operation actually behaves.
Contextualize
Contextualize means adding the relationships and meaning that allow people and systems to interpret the data correctly. A temperature reading becomes useful when it is associated with the correct asset, operating state, product, work order, environmental condition, maintenance history, and expected performance range. I used Contextualize where Walker uses Store because the purpose of this progression is different, and because storage is rarely the biggest limitation anymore. Most industrial companies can store more data than they could ever reasonably use. They have historians, databases, cloud platforms, lakes, warehouses, and lakehouses, which is an alarming amount of water imagery for an industry filled with electrical equipment. The harder issue is determining what all that stored data means.
Govern
Govern means deciding who owns the data, who can use it, what “good data” actually means, and what happens when something is wrong. It includes agreeing on basic things like names, definitions, access, quality expectations, and who has the authority to fix an issue. Those decisions may sound administrative, but without them, every plant, department, and software system slowly creates its own version of the truth.
This matters because the further data travels, the more damage bad assumptions can cause. A questionable sensor reading on a dashboard is annoying. That same reading feeding a prediction, triggering a maintenance workflow, or changing a production setting is a much bigger problem. Before a system can understand, predict, decide, or act, someone has to establish which information can be trusted and how it is allowed to be used.
Governance is one of those things companies often plan to add later, once the platforms are connected and the interesting AI work begins. By then, the data has already spread across systems, teams have created their own definitions, and nobody wants to give up the spreadsheet they have been using for seven years. Governance belongs early in the progression because it shapes everything that follows. You do not need every rule figured out on day one, but you do need enough agreement to prevent the rest of the system from being built on conflicting assumptions.
Understand
Understand means explaining what happened, why it happened, and how different conditions are related. This can involve statistical analysis, process knowledge, engineering models, machine learning, operator experience, or the technician who hears a motor for four seconds and knows exactly which bearing is beginning to fail. That technician is part of the intelligence architecture, whether the architecture diagram admits it or not.
Understanding also depends on governance more than it may initially appear. Two teams can analyze the same operation and reach different conclusions because they use different definitions, time windows, asset hierarchies, or versions of the data. Governance does not produce understanding on its own, but it helps ensure that everyone is at least investigating the same version of reality.
Predict
Predict means estimating what is likely to happen next based on current and historical information. Industrial applications include predicting equipment failures, quality problems, production delays, process drift, energy demand, safety risks, and material shortages. Prediction is where many industrial AI initiatives begin because it is easy to demonstrate and sounds advanced. It also produces wonderfully impressive charts. The problem is that knowing what will happen does not guarantee that anyone will do something useful about it, which brings us to the capability I think is missing from many industrial AI discussions.
Decide
Decide means selecting an appropriate response based on the prediction, available options, operating objectives, costs, constraints, risks, and authority. If a model predicts that a motor is likely to fail within five days, the operational question becomes whether to replace it immediately, reduce its load, move production elsewhere, wait for a scheduled shutdown, or continue running because the replacement part is currently sitting on a cargo ship in a location nobody can identify. This is where intelligence starts becoming operational. A prediction describes a likely future. A decision determines how the organization should respond to that future.
I once saw a predictive-maintenance project correctly identify that a piece of equipment was going to fail. The model worked, the alert was generated, and the information entered the company’s normal maintenance-review process. Unfortunately, the equipment failed before the next meeting. The organization had invested in prediction without developing a faster way to evaluate the recommendation, assign authority, and act.
The clever domino was standing by itself.
Automate
Automate means executing predefined tasks, responses, or workflows without requiring a person to perform each step manually. Manufacturing has been doing this extraordinarily well for decades through control systems, PLCs, robots, machine logic, recipes, sequencing, and workflow tools.
Automation works best when the situation and the appropriate response can be defined in advance. When a sensor crosses a threshold, the system performs a programmed action. When a product reaches a station, the machine runs the correct sequence. When a condition occurs, a known workflow begins. The complexity can be enormous, but the decision logic and boundaries have already been established. Automation is also how intelligence finally reaches the physical operation. A model can identify an issue, recommend an action, generate an explanation, and produce a very attractive dashboard, yet none of that changes throughput, quality, safety, or downtime until something in the operation responds. Industrial value is created when insight becomes action.
Autonomize
Autonomize means enabling a system to pursue a defined objective, evaluate changing conditions, choose among available actions, and operate within established constraints with limited human intervention. Humans still define the objectives, acceptable risk, authority, safety boundaries, escalation rules, and accountability. Those responsibilities become more important as the system gains greater freedom to act.
An autonomous production system might detect quality drift, determine the likely causes, adjust process parameters, rebalance production across equipment, revise the schedule, and monitor whether those changes improved the result. It may continue adapting as conditions change, or it may decide that the situation exceeds its authority and involve a person. Knowing when to stop and ask for help is part of autonomy; a system that acts confidently in every situation is less intelligent than it is terrifying.
Factories rarely behave according to the neat little diagrams we make for them. Materials show up slightly different. A machine starts wearing down. Someone changes an order. A shift runs short-staffed. Energy costs spike. One line is having its best day in months while the one beside it begins making noises that suggest it has other plans. Traditional automation works well when we already know what should happen next. Autonomy matters when we do not. The system has to read the situation, decide what matters, choose a reasonable response, act, watch what happens, and possibly change its mind. And it has to do all of that without turning a manageable production issue into a very efficient, fully automated disaster.
Why the Dominoes Get Larger
The progression becomes valuable when the capabilities are considered together rather than treated as eight unrelated technology projects. Each stage gives the next stage more to work with, and the value compounds as the operation becomes capable of moving from sensing a condition to understanding it, choosing a response, and changing the physical outcome. I think of the nine stages in three broad groups:
Connect, Collect, and Contextualize create a usable foundation.
Govern, Understand, and Predict create trusted intelligence.
Decide, Automate, and Autonomize turn intelligence into operational action.
That grouping looks tidy, which is slightly misleading because the work itself is rarely tidy. Companies usually have different levels of maturity across plants, assets, applications, and use cases. A facility may have excellent automation but terrible contextualization. Another may have a modern data platform connected to only a small portion of its equipment. A company might make advanced decisions in one process while emailing spreadsheets around in another. Industrial maturity is usually less like climbing a staircase and more like searching several rooms for pieces of the same staircase.
Still, the sequence matters. Reliable predictions require meaningful information. Meaningful information must also be trusted, controlled, and understood consistently. Good decisions require understanding, objectives, and constraints. Automated actions require dependable systems. Autonomy requires all of it, plus governance, confidence limits, and a clear method for involving humans when the system reaches the edge of what it should be allowed to do.
The progression also creates a feedback loop. Actions generate new operational data. The organization can evaluate whether the decision worked, improve its understanding, update its models, and make better decisions next time. Governance ensures that this learning is traceable, controlled, and incorporated without quietly breaking something that was working yesterday. The dominoes fall forward, while the learning circles back.
AI can contribute across nearly every step. It can help classify and contextualize data, identify quality problems, apply governance policies, track lineage, monitor models, explain unusual behavior, identify patterns, predict outcomes, compare alternatives, support decisions, coordinate workflows, and eventually help systems operate with greater autonomy. That is exactly why I do not include AI as its own domino. Industrial AI success comes from the capability it enables throughout the chain, rather than from the mere presence of an AI model.
This is where I become fairly opinionated. Companies spend too much time asking which model, platform, copilot, or agent they should purchase, while spending far less time on questions that will determine whether any of those technologies can produce a result:
Can we reliably access the operational data?
Do we understand what the data represents?
Do we know who owns it, whether it can be trusted, and what rules govern its use?
Can we explain what happened before trying to predict what happens next?
Who can make the decision, and how quickly?
Can the operation execute the decision safely?
What happens when the system is uncertain, wrong, or facing competing objectives?
These questions are not as exciting as an autonomous-factory demonstration. They are, however, the questions that determine whether the demonstration can survive contact with an actual factory. Everyone wants the last domino because it promises the greatest value. It represents operations that can sense, understand, decide, act, and adapt without waiting for a weekly meeting or a person to move information manually between disconnected systems. I want that future too, and I think parts of it will arrive faster than many industrial organizations expect. The companies that achieve it will not be the ones that simply add AI. They will be the ones that build the capabilities AI requires, connect those capabilities into an operating system, govern how intelligence and authority are used, and gradually earn the right to give technology greater authority.
Everyone wants the last domino. The real work is standing up the first eight.
