Where AI pays off on the shop floor, and what predicts it

The AI value gap

The benefit is real. Few companies collect it.

BCG: value captured McKinsey: AI at scale

Share of companies, 1,250 executives surveyed

• 5%capture outsized value from AI
• 35%are scaling toward it
• 60%report minimal gains in revenue or cost

Select a group to see what it means.

The 5% settles whether AI can work on the floor. Their gains are real, repeatable, and bankable. The more useful question is what those companies did that the other 60% didn't, and whether you can tell, before you spend, which group your plant will land in. The answer is less glamorous than the vendor brochures suggest, and you can learn it.

Success with AI is attainable, but unevenly distributed

Credible numbers exist. Deloitte's 2025 Smart Manufacturing and Operations Survey (600 US executives at companies above $500M in revenue, fielded in 2024) reports 10% to 20% gains in production output, 7% to 20% gains in employee productivity, and 10% to 15% more capacity recovered from smart-manufacturing and AI initiatives.² These are actuals reported by practitioners at large manufacturers, not vendor case studies or Lighthouse highlights, and they are the most credible benefit figures in the field. A plant that adds 15% output without new capex has improved its P&L.

The catch is where those gains show up. McKinsey's work on scaling AI in manufacturing finds that only about 2% of manufacturers have AI fully embedded across global operations, and nearly 60% "struggle to distinguish between a success and an expensive experiment."³ Put that next to BCG's 5/35/60 split and the picture is clear. A small group collects double-digit gains. A much larger group runs pilots it can't grade, on floors that were never instrumented to prove a benefit happened.

The timeline points the same way. A separate Deloitte study (1,854 senior executives across Europe and the Middle East, surveyed in 2025) found that companies expected payback in 7 to 12 months, but most reported satisfactory returns only after two to four years. Only 6% achieved payback in under a year, and even among the most successful projects, only 13% did.⁴ Spending kept climbing anyway: 85% raised AI investment over the prior year, and 91% plan to raise it again. Money moves fast and returns arrive slowly, and an unmeasured pilot can survive in that gap long enough to look like a success.

The two clocks

The money moves fast. The returns arrive slow.

Payback, in months. Drag the marker along the timeline.

Payback the business case expected

7–12 months

When most reported satisfactory returns

2–4 years

Month 18

012243648 mo

Months since the AI investment

Month 18: past the expected window, and before most companies reported satisfactory returns.

Paid back in under a year

All companies Most successful projects

6%of companies achieved payback in under a year

85%raised AI investment over the prior year, and 91% plan to raise it again

The concentration is structural, and the rest of this piece is about what drives it.

Most of the evidence answers a question that's already settled

Vendor decks and transformation galleries mostly use their evidence to answer one question: does AI help the floor? The World Economic Forum's Global Lighthouse Network is the polished version. Its audited sites posted roughly 40% gains in labor productivity and 48% reductions in lead time in the latest cohort, and the network's best all-time results include defect reductions above 80% and productivity gains above 50% at individual sites.⁵ Those numbers are independently audited. The standard use cases (predictive maintenance, vision-based quality inspection, demand forecasting) are proven and now table stakes.

A gallery of leaders proves AI can work. It can't tell you whether it will work in your plant, on your line, against your problem, because the median plant isn't in the gallery. Galleries exist to show wins, and the audits are sound. The cost is that they draw attention to the trophy instead of the conditions that earned it, and they never show the denominator: everyone who tried the same thing and got a fraction of the result. You can study a hundred Lighthouse sites and still not know whether your bet will pay off.

The better questions are where the benefit concentrates, and what predicts whether a given plant lands in the 5% or the 60%.

What predicts the 5%

WEF's Lumina analysis models the predictors instead of celebrating the winners. It covers 1,085 transformation solutions across 32 countries and 35 industries from 2017 to 2025, enough data to ask what separates wins from the rest. Sites that addressed talent priorities (skills, safety, worker experience) were far more likely to perform above the median, and sites with structured skills programs scaled their initiatives meaningfully faster. About half of successful transformations started from a specific cost, quality, or safety problem rather than a technology mandate, and those problem-first efforts produced more measurable impact than top-down programs.⁶

For an ops reader, that runs against instinct. The strongest predictors of payoff are starting from a concrete floor problem and giving the tool to people ready to use it. The algorithm and the budget rank below both. McKinsey's enterprise data agrees: high performers are about three times more likely to redesign workflows than to insert AI into existing processes.⁷ AI amplifies an operation you already run well. It can't supply one you lack. A plant with a clear scrap problem, a team that trusts its data, and a process worth fixing is where the model lands. A plant that bought AI because the board wanted an AI story ends up running an expensive experiment it can't grade.

The uncomfortable part is that the best predictor, operator readiness and skills, is where manufacturers are weakest. Deloitte's survey ranks human capital as the lowest-maturity category it measured.² The gains are blocked less by technology or price than by the capability most plants have invested in least.

The skills paradox

Same factor. Top of one ranking, bottom of the other.

Two studies, two rankings. Select either end to read what each study measured.

WEF Lumina: what predicts payoff

Strongest predictor Talent and skills readiness Other predictors not shown Weaker

Deloitte: maturity by category

Most mature Other categories not shown Human capital Lowest maturity measured

Sites that addressed talent priorities (skills, safety, worker experience) were far more likely to perform above the median, across 1,085 solutions.

Payoff also scales with company size. McKinsey reports that about 50% of companies above $5B in revenue have reached the scaling phase, versus 29% of those under $100M.³ Gartner adds a maturity gate: organizations with successful AI initiatives invest up to four times more in data and analytics foundations.⁸ A small plant that failed to reproduce a $5B firm's enterprise rollout may have made a wrong-sized bet rather than lagged: enterprise scope on a single-site budget, or a global data platform for a problem that lived on one line. Size the bet to the operation you run.

One question the data doesn't answer is who captures the benefit once it's realized. The numbers suggest the enterprise books the scaled value while the plant carries the ramp-up, oversight, and retraining costs. Nobody has quantified that split. A plant leader should still ask before signing the business case, because the answer determines who has to keep the win alive after launch.

The benefit comes from the redesign

BCG's manufacturing work reports AI-driven cost reductions of 60% or more, with a condition most decks bury: "only if processes are redesigned end-to-end."⁹ The 60% is the headline; the condition is the finding. AI is the forcing function, and the end-to-end redesign it forces is where the gain comes from.

Deloitte describes the same issue as a measurement problem: AI is "typically rolled out alongside broader digital, operational or structural changes," which makes the benefit hard to attribute to AI at all.⁴ When the AI and the process change ship together, you can't fully separate which one paid off. The direction is still clear. Bolt AI onto an unchanged workflow and expect a fraction of the headline number. That's an error in the business case: you bought a tool and budgeted for a result only a redesign delivers.

Read every 60% figure that way. It's probably real for the site that earned it and rarely transferable, because what produced it was the rebuilt process around the model, and that process doesn't ship in the box.

The durable, measured value also sits in unglamorous analytical machine learning, while the generative and agentic layers get the attention. WEF's Lumina found that analytical AI powered 41% of the value-generating solutions in predictive operations and quality.⁶ BCG put agentic AI at about 17% of AI value in 2025, projected to reach 29% by 2028.¹ Deloitte found 15% of respondents reporting significant, measurable ROI from generative AI, versus 10% from agentic AI.⁴ Budgets that chase the spotlight tend to fund the expensive experiments.

Where measured value sits today

The bankable benefit is real. It isn't where the spotlight is.

Three measures from three sources. They are not comparable and do not sum.

WEF Lumina

41%

of the value-generating solutions in predictive operations and quality were powered by analytical AI

BCG

17%

of AI value came from agentic AI in 2025

2025 2028, projected

Deloitte

15%

of respondents reported significant, measurable ROI from generative AI

Generative Agentic

The benefit is perishable

The business case usually leaves out one more thing, and it decides whether the win survives a second line and a third quarter: models degrade.

A peer-reviewed study by Vela et al. ("Temporal quality degradation in AI models," Scientific Reports, 2022) tested 32 datasets across four industries and four model families and found temporal degradation in 91% of cases.¹⁰ In production, model quality doesn't hold without retraining. The datasets came from healthcare, transport, finance, and weather, so the decay rate on a shop floor is unmeasured. The mechanism is general, though, and easy to picture on the floor. A vision model tuned on this quarter's parts, suppliers, and lighting loses accuracy as those change. A demand model trained on last year's mix drifts as the mix shifts. Launch-day "99% detection" has a half-life no deck mentions, and the decline is quiet: the dashboard stays green while real performance slides.

Model drift

Launch-day accuracy has a half-life.

Illustrative shape, not shop-floor data

No retraining Retraining funded

The number on the dashboardThe number measured in production

Time since deployment

Time since deployment

Launch day: both numbers agree.

The implication, reasoned rather than measured for manufacturing, is that a real business case needs a maintenance line it almost never includes: an owner, a retraining cadence, and a way to catch drift before a customer does. Without it, the win slides back toward baseline while the original number stays in the deck.

The same logic exposes another gap. Deloitte's top-ranked barrier to realizing AI ROI is "intangible benefits difficult to monetize."⁴ Faster operator ramp-up, captured tribal knowledge, and fewer middle-of-the-night escalations all belong here. The field admits the category exists, but every magnitude attached to it comes from vendors, and no independent study prices it. So it gets cut from the deck, and the deck understates what a good deployment is worth.

Three things separate the 5% from the 60%, and the algorithm isn't one of them. The winners start from a concrete cost, quality, or safety problem rather than a technology mandate, because problem-first measurably beats top-down. They fix operator skills first, since readiness is the predictor that scales the win and the one most plants are weakest on. And they fund retraining, because the gains decay, and an unfunded model fades while the dashboard says it hasn't.

AI's benefit is real, earned, and perishable. Almost no business case is written to account for all three, so check whether yours does before you sign it.