Operations leaders today face a consistent mandate: do more with less — less labor, less margin, and less tolerance for downtime. Most respond by investing in connectivity and analytics, and most of those investments fail to translate into operating performance.
The reason is rarely the technology itself. It is the order of operations. Most digital manufacturing programs begin with technology — assets to connect, models to build, dashboards to deploy — and work outward to the decisions those investments are meant to improve.
The programs that succeed invert that sequence. They begin with the decision, define what the operator needs to know to make it differently, and build only the data, analytics, and integration required to put the recommendation in front of the right person at the right moment.
Stated directly: begin with decisions, not technology. That principle changes what to invest in, in what order, and how progress is measured. It also predicts where most programs fail. The dominant failure mode in the field is not insufficient data or weak models; it is insight that never reaches the operator at the moment a decision needs to be made.
A framework built around the inversion — Connect, Analyze, Act — is the operating model for a decision-first program. Three layers are applied in sequence, each with a defined test. The evidence comes from greenfield and brownfield facilities operating today and is examined below.
Why Now: Three Structural Forces
The pressures on manufacturing today are not cyclical; they are structural. That distinction matters because it determines whether the right response is patience or fundamental change.
Labor scarcity
The average skilled trades worker in the U.S. is over 44. Institutional knowledge is retiring faster than hiring can replace it.
Supply chain fragility
COVID exposed it; geopolitics, reshoring mandates, and single-source dependencies have kept it exposed. Carrying more risk demands sharper visibility to manage that risk.
Margin compression
Energy, materials, and logistics costs have stepped up structurally. Selling prices have limited elasticity. The remaining lever is operational performance.
Taken together, these forces lead to a single conclusion: “wait and see” is more expensive than “start and learn.”
The Mandate Is Quantified
The performance differential is well-documented. Benchmarks for plants with integrated data and analytics report:
- 15-25% improvement in OEE in comparable operations
- Approximately 40% reduction in unplanned downtime
- One-third lower scrap rates
These outcomes come from plants that executed the work with discipline; there is no silver bullet. The implication, however, is clear: when competitors invest along these lines and others do not, the gap compounds.
Investment Is Flowing, Yet ROI Remains Elusive
Budgets are funded. Projects are running. Yet, industry research continues to place the failure or stall rate on digital manufacturing initiatives at 70% or higher. A 2024 BCG survey of nearly 1,800 manufacturing executives found that 89% plan AI implementation in production and 68% have started, but only 16% have hit their AI-related targets. The cause is rarely the technology itself. It is almost always one of three patterns:
- Point solutions without integration: Capable tools are acquired independently. They don’t communicate with one another, creating data islands rather than data infrastructure. Plant leaders are left navigating 8 to 10 different applications simply to operate the floor.
- Pilots that never scale: A 90-day proof-of-concept succeeds and momentum builds, but 18 months later the solution still runs on the same line. No scalability plan was ever defined or resourced.
- Data without decisions: Dashboards and alerts are deployed, but the underlying workflow remains unchanged. The data exists, but it does not drive action.
The remedy is not more investment. It is a different approach.
Why Isolated Automation Falls Short
The default instinct when asked to do more with less is to identify the largest source of waste and automate it. That instinct is not wrong; it is incomplete.
Automating Line 1 by 5%, Line 2 by 8%, and a utility by 3% produces real but bounded localized gains. The same operations connected — sharing data, surfacing leading indicators, supporting decisions across the system — produce materially larger network-effect gains. The difference is not better technology at each node. It is the connections between them, and what those connections enable downstream.
Scrap is illustrative. The temptation is to treat scrap as a downstream quality problem. In reality, scrap is the output of a system: raw material variability, process parameters, operator behavior, and maintenance timing all contribute. Manufacturers that detect signals in incoming material quality, dynamically adjust process parameters, and trace impact through to yield move beyond local optimization. They begin operating with system-level intelligence — anticipating issues before they manifest.
Reaching that level requires avoiding three specific failure modes:
- Technology without architecture: Disconnected tools accumulate over time — silos masquerading as a digital strategy.
- Insights without workflow: Dashboards surface data without changing decisions — dashboard theater.
- Projects without capability: External teams build and hand off the project, leaving the internal team without context. Six months later, the system is running in degraded mode.
The discipline that separates the winners from the 70% is beginning with decisions, not technology.
The Framework: Connect, Analyze, Act
Three layers, each anchored to the decision the program is built to improve, are applied in sequence. Each builds on the one before it; the order cannot be shortcut.
Connect – Build a Unified Data Layer
The first task is to make data usable, not merely accessible. That requires an integration layer above existing OT assets — PLCs, SCADA, historians, MES — that normalizes data into a unified model that analytics applications can consume.
Three principles govern this layer.
- The integration layer is additive, not a forklift replacement: Existing systems remain in place.
- Operational context matters as much as the sensor value: A temperature reading is meaningless without knowing which asset, which product, and under which process conditions it was recorded.
- Avoid connecting everything at once: Begin with 3 to 5 high-value use cases and build infrastructure just-in-time to support them.
The sequencing rule is: start with the use case, work backward to the data required, and then implement the architecture. Manufacturers who build infrastructure first and identify use cases later consistently over-engineer and under-utilize.
This is also where the “our data isn’t ready” objection arises. It is almost always true and almost never as disqualifying as it feels. No manufacturer has perfect data. The relevant question is not whether the data is ready in the abstract — it is whether it is sufficient for this specific use case. A data readiness assessment makes that question answerable.
| Low Data Readiness | High Data Readiness | |
|---|---|---|
| High Business Value | Foundation first: The use case justifies the targeted infrastructure investment. | Quick wins: Execute in weeks. Begin here. |
| Low Business Value | Avoid | Nice to have: Defer |
Analyze – Move From Lagging to Leading
Once data is connected, the focus shifts to insight, and specifically to leading indicators over lagging ones.
Lagging indicators report what has already occurred: OEE at end of shift, QC lab results, maintenance work orders. By the time the indicator registers, the value has already been lost. Leading indicators report what will occur while there is still time to act: parameter drift detection, early-warning anomalies, soft-sensor quality inference. The defining question of the Analyze layer is: what is the process indicating now about what will happen in the next four hours?
This is where AI and machine learning belong, but only against the right problems. AI is genuinely transformative for dynamic scheduling (real-time optimization across constraints), predictive quality (soft sensors inferring attributes in real time), anomaly detection (high-frequency patterns beyond human bandwidth), and multivariate optimization (20 or more interacting non-linear variables that exceed human intuition).
It also exhibits predictable failure modes. AI without a data foundation produces garbage in, garbage out — and worse, confidently expressed wrong answers. AI without process understanding produces correlations mistaken for causal levers; trust erodes the moment a recommendation does not make physical sense. AI without a decision to improve produces a well-built model, strong technical metrics, and no operational change.
Three questions warrant clean answers before any AI investment:
- What decision is being improved?
- What data is actually available?
- Who changes their behavior based on the output?
In the absence of clear answers, the use case is not yet worth pursuing.
Act – Insight Has No Value Until It Changes a Decision
The Act layer is where the most value is left unrealized, and it is the least technical, which is precisely why it tends to be overlooked.
The contrast is stark. In a typical implementation, an analytics tool produces an insight, a dashboard displays it, the operator glances at it intermittently, and no action follows. The value delivered is $0.
In a workflow-integrated implementation, the model generates a prediction, a contextual alert appears within the HMI or MES the operator is already using, the operator acts at the decision point, and the outcome is measured and fed back to the model. Value is measured in production KPIs.
The design principle is that the best implementations are nearly invisible, and the recommendation resides within the existing workflow, not within a separate application. The path can extend toward fully autonomous operation, but the typical starting point is human-in-the-loop, which is why integration with existing decision processes is so consequential.
Change management becomes as important as the technology at this layer. Operator trust is built through early wins, transparency about why the model produces a given output, and a means for operators to flag when the model is wrong. A perfectly accurate model that operators do not trust delivers no value.
From the Field: Operationalizing a Soft Sensor
Several years ago, I worked with a pharmaceutical manufacturer on an agitated-pan drying process. The team had built a soft-sensor model and required support in operationalizing it.
The challenge in batch operations is that the quality attribute of greatest interest often cannot be measured in real time. A sample is pulled, sent to QC, and the result returns 4 to 8 hours later. A soft sensor, on the other hand, uses process data already being collected — temperatures, NIR readings, mixing parameters — to infer that quality attribute in the present. It’s not perfect, but it is accurate enough and on a time horizon that permits action.
The model itself represented roughly 20% of the work. The remaining 80% was operationalization: routing the output into the control system in a form operators could act on, validating against lab results, and earning operator trust over time.
Proof Points: Greenfield and Brownfield
The framework holds at both ends of the spectrum.
Greenfield: Hitachi Rail, Hagerstown, Maryland
Hitachi Rail’s Hagerstown facility opened in September 2025 with a $100 million total investment. It produces one rail car every day and a half, and includes more than $30 million dedicated to digital infrastructure.
The decision the Hagerstown team faced is the decision every manufacturer faces: build production capability first and layer digital on top later, or treat data infrastructure as a day-one requirement. They chose the latter.
The result was AI quality inspection, digital twins, real-time supply chain visibility, and predictive maintenance all live on day one. It wasn’t a roadmap; it was operational at launch. With the backbone in place, the team is now planning the move from assistive analytics to agentic and autonomous operations.
The principle is: the data layer is infrastructure, not an afterthought.
Brownfield: Omika Works
Omika Works — the first factory owned by a Japanese company to receive World Economic Forum Lighthouse distinction — is the brownfield mirror image. Connect began with visualization, making factory information accessible and reliable, with RFID tracking applied to work-in-process. Analyze layered intelligence incrementally — statistical process monitoring first, predictive models thereafter — within a high-mix, low-volume environment. Act added decision support as each foundation element proved reliable; each layer is earned before advancing to the next.
The result was a 50% reduction in production lead time, without compromising quality.
Regardless of the starting point, the Connect, Analyze, Act framework is effective. Only the scope and pace of each layer change.
Where the ROI Comes From
When the framework is applied with discipline, returns concentrate in three operational levers:
Quality
In-process quality prediction catches drift before defects occur. SPC paired with real-time model output reduces false alarms.
Throughput
Dynamic scheduling, real-time bottleneck identification, and fewer unplanned stops are driven by predictive alerts.
Maintenance
Condition-based monitoring replaces calendar-based preventive maintenance. Spend shifts from reactive to planned.
These results reflect published benchmarks. There is no silver bullet, but the results are documented and reproducible.
A Phased Roadmap: Prove Before Scaling
The most common error is attempting to build everything simultaneously across the full plant footprint. That approach is the predictable origin of a $5 million project that stalls.
| Phase | Window | Focus | Outcome |
|---|---|---|---|
| 1. Prove | First 90 days | Single high-value, data-ready use case, tied to a KPI already in use. Not a pilot, a deployment. | ROI evidence and organizational credibility |
| 2. Foundationalize | 90-180 days | Use early wins to justify expanding the data foundation. Build internal capability. External partners transition knowledge to a sustainment team. | Scalable infrastructure owned by an internal team |
| 3. Scale | 6-18 months | Expand across lines and plants. Internal champions drive adoption. Each new use case compounds value on the same data foundation. | Embedded program; compounding returns |
Phase 1 exists to neutralize the two objections that derail most digital programs. The first — “we can’t justify the ROI” — is a sequencing problem, not an investment problem. A properly scoped first deployment should pay for itself and be tied to demonstrable KPIs before any commitment to Phases 2 and 3. The second — “we don’t have the internal expertise” — is real, and is why the engagement model matters. External partners should deploy with the internal team, not instead of it. The test is straightforward: if the partner leaves and the system fails, it was never truly deployed.
The trap to avoid is chasing high-value use cases regardless of data readiness. Manufacturers that do so contribute directly to the 70% failure rate.
Three Behaviors of the Manufacturers Who Are Winning
The patterns that distinguish leaders are organizational, not technological:
1
Define success in business terms.
Don’t use “digitize the plant.” Instead use: “15% scrap reduction on Line 3 by Q3.” Include a target, a KPI, and an end date. Technology is the means, not the objective.
2
Treat data as infrastructure.
Make a deliberate decision that data is an asset, not a by-product. Someone owns the architecture and governs its quality.
3
Build internal champions.
Identify a single person inside the organization — typically in operations or engineering — who owns the transformation and possesses both floor credibility and leadership authority. Without both, the change does not endure.
The Path Forward
For organizations prepared to apply this framework, three concrete steps follow. First, apply the readiness matrix: identify where high business value intersects with reasonable data readiness; that intersection is the starting point. Second, conduct a structured assessment — 2 to 4 weeks covering current state, data landscape, and highest-priority opportunities. Third, design Phase 1 as proof-to-production, not a pilot, tied to a KPI already in use, scoped to deliver measurable business value.
The manufacturers that succeed at “do more with less” will not be those with the most technology. They will be the ones that begin with the decision. Then, connect, analyze, and act in service of it.
The Act layer is where decision-first discipline pays off. It is also where most programs leave value on the table. Closing that gap is the next decade of competitive advantage in industrial operations.
