An operations VP at a mid-size manufacturer recently described her AI roadmap this way: "We have four vendor proposals on the table — predictive maintenance, computer vision inspection, an energy-optimization platform, and an agentic scheduling tool. Every one of them has a case study showing 30-plus percent improvement. I don't know which one to fund first, and I don't think any of the vendors pitching me know either, because none of them have seen my plant."
That's the actual problem. Not a shortage of AI use cases with proven track records — there are several, and the track records are real. The problem is that "proven track record" is an industry-average number, and the industry average tells you almost nothing about which use case will pay back fastest inside one specific plant, with its specific loss profile, its specific data maturity, and its specific appetite for the workflow change each use case requires. This piece lays out a diagnostic for making that call deliberately, rather than by whichever vendor's deck landed on the desk most recently.
Part 1 — The Wrong Question Is "Which Use Case Has the Best ROI"
Ask five different AI vendors which use case delivers the best return and you'll get five confident, well-cited, mutually incompatible answers. That's not because the vendors are lying. It's because "best ROI" is not a property of a use case in the abstract — it's a property of a use case matched against a specific plant's cost structure, data infrastructure, and organizational readiness. A use case with a spectacular published ROI at a heavy-process chemical plant running three shifts can be a mediocre investment at a mixed-model discrete manufacturer running one shift, and the reverse is just as true.
McKinsey's research on the Global Lighthouse Network — the manufacturers furthest along on AI-driven transformation — makes this variance visible rather than theoretical. Within a single cohort of front-runner sites, AI-driven quality inspection cut defect rates by 68% at one ceramics manufacturer and by 49% at a life-sciences equipment maker — genuinely different outcomes from the same category of use case, driven by how each site's existing process and data already fit the technology. If the highest performers in the world get a spread that wide on the same use case, "which category has the best average ROI" was never the right diagnostic question for any single plant to be asking.
RAND's interviews with data scientists and engineers across industry point at the same root cause from a different angle: the single most common reason AI projects fail isn't the model, it's business leadership misunderstanding or miscommunicating the actual problem the project was meant to solve. A vendor's published ROI answers a question about their case study's problem. It doesn't answer the question about yours until someone has actually defined what your problem is, in your numbers.
Part 2 — The Three Variables That Actually Decide It
Once "average ROI" is off the table, three variables specific to your operation are what actually determine which use case should get the first dollar.
Financial leverage. What does the failure mode this use case addresses actually cost you, in your own numbers — not the industry benchmark? A plant running high-margin, low-volume production has a very different scrap-cost profile than one running thin-margin, high-volume commodity parts. A plant with a $40,000-per-hour bottleneck has a different downtime-cost profile than one with quick, cheap changeovers. This number has to come from your own P&L and your own maintenance logs, not from a vendor's case study.
Data readiness. Every AI use case has a data prerequisite, and it is rarely uniform across use cases at the same plant. Gartner's research on industrial AI data readiness found that over 50% of AI projects fail to reach production, with data issues as the primary blocker for 40% of them — and the specific gap is almost always use-case-dependent. A plant might have excellent vibration-sensor coverage on its critical rotating equipment (ready for predictive maintenance) and almost no structured historical defect-image data (not ready for computer vision). Scoring "data readiness" as one number for the whole plant obscures the fact that it's really five different numbers, one per candidate use case.
Organizational capacity to absorb the change. Every use case implies a different change to how people work, and plants have a finite budget of organizational change they can absorb at once. Predictive maintenance changes what triggers a work order. Computer vision inspection changes what an inspector's job actually is. Agentic scheduling changes who has authority to re-sequence the line. Deloitte's 2026 manufacturing outlook found that 81% of task hours in manufacturing are expected to remain human-driven even as AI investment accelerates — meaning the success of nearly every use case still runs through a human workflow, and the plant's capacity to retrain and re-earn trust in that workflow is a hard constraint, not a footnote.
Part 3 — Why the Highest-Profile Use Case Isn't Automatically the Right First Move
Predictive maintenance gets pitched first at most plants — it has the most mature vendor market, and unplanned downtime is a visceral, easy-to-picture loss. But the three variables don't automatically point to the same use case. A plant that scores high on PdM's data readiness (good sensor coverage, clean maintenance history) but low on financial leverage (low-criticality equipment, cheap repairs) is looking at a use case that will work technically and underwhelm financially.
The same logic runs in the other direction. A plant with a severe quality problem — high scrap, high warranty cost, a customer threatening to dual-source — has enormous financial leverage sitting in computer vision inspection, even if that plant's image-labeling data isn't yet clean enough to hit a vendor's advertised accuracy numbers out of the gate. In that case, the right first move might be a shorter data-preparation project that makes the eventual quality use case viable, rather than defaulting to whichever use case happens to have the cleanest data today. The diagnostic isn't "pick the use case that's easiest to start." It's "identify where the leverage actually is, then figure out what has to be true before that use case can be started well."
This is also why broad adoption statistics are a poor proxy for whether a use case will actually work at your plant. BCG's 2025 research on the AI adoption gap found that 88% of organizations now regularly use AI in at least one business function, but only 6% qualify as high performers capturing significant enterprise-wide value. Widespread adoption of a use case category tells you it's technically achievable somewhere. It tells you nothing about whether it's the highest-leverage move for your plant specifically.
Even inside the Lighthouse cohort, more than 80% of AI use cases are still deployed at the individual process-step level rather than across an entire operation — narrow, specific interventions matched to a narrow, specific problem, not a category-wide bet. The most advanced manufacturers in the world are running the same discipline this diagnostic asks for: start where the fit is real, not where the category is popular.
Part 4 — Running the Diagnostic Against Your Own Plant
In practice, this is a scoring exercise, not a gut call. List every candidate use case on the table — typically predictive maintenance, quality inspection, energy optimization, and scheduling cover most of what's realistically available today. For each one, pull three numbers: the annualized cost of the failure mode it addresses (financial leverage), a plain assessment of whether that specific data already exists in usable form (data readiness), and an honest read on how much disruption the affected team can absorb this year (change capacity).
The use case that should get the first dollar isn't necessarily the one that scores highest on all three. It's the one where financial leverage is high enough to matter and the other two variables aren't actively working against it. A moderate-leverage use case with excellent data readiness and low organizational friction can be the smarter first move than a high-leverage use case that would require a parallel data-cleanup project and a change-management fight — not because the second use case is a bad long-term investment, but because sequencing it second, after the first project builds internal credibility and cleans up adjacent data, makes it easier to execute well. This is a portfolio decision, not a single bet: the diagnostic should produce an ordered list, not just a winner.
The manufacturers running this well tend to fund the foundation alongside the first use case rather than treating the use case as the entire budget line. Deloitte's 2025 Smart Manufacturing and Operations Survey of 600 executives found that 78% of surveyed leaders are allocating more than 20% of their overall improvement budget to the underlying data, sensor, and cloud infrastructure that smart-manufacturing initiatives depend on — not to the AI use case itself. That split matters for scoring: a use case that looks marginal on data readiness today can become the obvious next pick once that infrastructure share of the budget has done its work.
What counts as "ready" isn't a single checkbox either. Gartner's framework for industrial AI data readiness breaks the prerequisite into three separate layers that all have to hold at once — structured data curation, real-time delivery to the model, and active, continuously updated metadata — before a use case can scale past a pilot. Scoring "data readiness" honestly means checking a candidate use case against all three, not just confirming the sensors exist.
Part 5 — The First Use Case Sets the Baseline for Every One After It
The plants that get this right treat the first AI dollar as a credibility investment as much as a financial one. A well-scoped, well-executed first use case — even a moderate-leverage one — builds the internal muscle (data governance habits, cross-functional trust, a working model of what "AI project" actually means here) that makes every subsequent use case faster and cheaper to deploy. A poorly sequenced first use case, chosen for its case study rather than its fit, tends to do the opposite: it burns goodwill and makes the next proposal a harder sell, regardless of how good it is on paper.
BCG's research on closing the AI impact gap identifies strategic clarity — a well-communicated, deliberately sequenced plan — as the factor that most reliably separates AI leaders from laggards, and notes that clarity improves outcomes even at organizations with limited AI tooling. A diagnostic that produces an ordered list, not just a first winner, is what that clarity looks like in practice. It also compounds: the same McKinsey Lighthouse research found that the time it takes front-runner manufacturers to implement a new AI use case has fallen by nearly 25% compared with earlier cohorts, precisely because the skills, data habits, and organizational trust built on the first use case carry forward into the next one.
None of this is a one-time exercise. As infrastructure matures and the organization absorbs its first use case, the scores on the next candidates shift — a use case that scored poorly on data readiness eighteen months ago may now be the obvious next move, precisely because the data foundation work done for the first project improved the ground truth the second one needs. The diagnostic isn't a gate passed once. It's the same three-variable filter, re-run every time a new candidate enters the conversation.
The question worth bringing to the table isn't "which of these four proposals has the best ROI." It's "given our actual loss profile, our actual data maturity, and our actual capacity to change how a team works this year — which one can we execute well enough to earn the second one." That's harder to answer in a single meeting. It's also the only version that produces a defensible answer.
Alpha Technical Solutions helps industrial operations leaders run this diagnostic against their own plant's numbers — scoring candidate AI use cases on financial leverage, data readiness, and organizational capacity before a dollar is committed. If you're weighing multiple AI proposals and need a structured way to sequence them, reach out.
