Search “AI in manufacturing adoption” and you’ll get a wall of confident percentages. One says 77% of manufacturers have implemented AI. Another says 95%. Nationally representative government data puts the share of manufacturers using AI at all well below either number.
These numbers don’t answer the same question, so they shouldn’t be read as competing measures of the same thing.
The useful question is narrower than “how many manufacturers use AI.” It’s this: how many have AI running in production, versus how many are still piloting, experimenting, or just planning to spend. Those are three different stages, and headline coverage frequently blurs them.
Here’s what holds up when you separate deployment from intent.
What these numbers measure
The word “adoption” carries at least four definitions across the surveys people quote, and they sit at very different bars.
The loosest is planned investment: “we’ve invested in AI/ML or plan to within five years.” Rockwell Automation’s number lives here. The next rung is any use: “we used AI somewhere in the business.” The US Census Bureau and McKinsey measure versions of this. Higher still is production deployment: “we’ve deployed AI at the facility or network level.” Deloitte provides the closest facility/network-level comparison. The strictest is measurable financial return: “the pilot moved P&L.” MIT’s widely shared figure sits there.
A manufacturer can be a “yes” on the first definition and a “no” on the last one at the same time. That’s not a contradiction. It’s the distance between buying a tool and running it in production.
Sample frames matter just as much. Nationally representative government data covers smaller firms that are often underrepresented in consulting surveys. Vendor and consulting panels skew toward large manufacturers with the budget to deploy, so their numbers run higher.
The Reliable Confidence Score
| Source or Claim | Figure | Reliable Confidence | What It Really Means |
|---|---|---|---|
| US Census Bureau, BTOS (2026 AI supplement) | In late 2025, ~18% of US firms reported using AI in any business function (32% employment-weighted); manufacturing reported below that national average. The older "in producing goods or services" wording ran near 10% nationally in fall 2025. | HighNationally representative | Drawn from a frame of ~1.2M businesses and weighted to the national firm population. No vendor stake. Measures whether a firm uses AI at all, not depth or production scale. By firm count, most manufacturers don't yet report using AI. The conservative anchor. |
| Deloitte 2025 Smart Manufacturing and Operations Survey | 29% are using AI/ML at the facility or network level; 24% have deployed generative AI at that scale. 23% are piloting AI/ML; 38% are piloting generative AI. | MediumLarge-firm sample | 600 executives, self-reported, fielded Aug–Sept 2024. Sample is very large manufacturers only ($500M+ revenue, 1,000+ employees), so it runs well ahead of the broad population. The clearest facility/network-use-versus-pilot comparison we found. Deloitte labels the 29% as "using," not "deployed"; only the 24% generative-AI figure is called deployed. |
| McKinsey Global Survey, State of AI 2025 | ~88% of organizations use AI in at least one function; about one-third have begun scaling, nearly two-thirds have not. | MediumCross-industry | Strong multi-year survey, but cross-industry, not manufacturing-specific. Captures the gap between using AI somewhere and scaling it across the enterprise. |
| Rockwell Automation, State of Smart Manufacturing 2025 | 95% of manufacturers have invested in, or plan to invest in, AI/ML over the next five years. | LowInvestment intent | 1,560 respondents, run by a company that sells the technology. A legitimate read on investment appetite, weak evidence for production: it bundles current spend with five-year intent and doesn't establish what's running on the floor. |
| MIT Project NANDA, The GenAI Divide 2025 | Only ~5% of integrated generative-AI pilots reach measurable P&L impact ("95% fail"). | LowP&L, not deployment | Self-described as preliminary and directional. Built on interviews with 52 organizations, 153 senior leaders surveyed at four conferences, and 300+ public initiatives. Cross-industry, not manufacturing. Measures financial return, not whether a pilot deployed. Routinely misread as a pilot death rate. |
| Rootstock 2025 State of AI in Manufacturing Survey | "77% of manufacturers have implemented AI to some extent" (up from 70% in 2023). | LowVendor-commissioned | Rootstock is an ERP vendor; survey of 369 manufacturers with 100+ employees across the US, UK, and Canada, fielded via Researchscape in late 2024. Traceable and competently run, but vendor-commissioned, broadly worded ("to some extent"), and not population-representative. A snapshot of a vendor's survey sample, not a deployment rate. |
The Big Takeaway
Put the rows side by side and a consistent picture shows up.
Nationally representative US data puts manufacturing AI use, in any business function, below the all-industry average, and the all-industry average itself sits around 18%. Among the very large manufacturers Deloitte surveyed, 29% reported using AI/ML at the facility or network level and 23% were piloting it; for generative AI, 24% reported deployment and 38% were piloting. Higher figures like 77% and 95% use broader definitions or fold in future investment, and they shouldn’t be read as plant-floor deployment rates.
So the honest framing of “deploying vs piloting” is this. Production deployment in manufacturing is a minority position even among the largest firms. Across the full population, government data put AI use lower still, though that measure captures any use, not production deployment. Piloting accounts for a substantial share, particularly for generative AI, where piloting (38%) runs well ahead of deployment (24%). For conventional AI/ML the two are closer, with facility-level use (29%) slightly above piloting (23%).
Broad definitions produce impressive adoption numbers. Measures tied to facility-level use, deployment, or financial impact paint a far more restrained picture.
Why the numbers disagree so much
Five things drive the spread, and definition is the biggest.
First, what counts as adoption. Invested-or-plan-to, used-it-somewhere, deployed-at-scale, and moved-P&L are four different questions. Surveys rarely lead with which one they asked.
Second, who got surveyed. A nationally representative frame includes small shops; a consulting or vendor panel leans toward big manufacturers that deploy more. Same question, different population, different answer.
Third, who’s asking. A government statistician has no stake in the result. A technology vendor surveying “will you invest in AI” has an obvious one.
Fourth, AI/ML versus generative AI. These get merged constantly. Generative AI piloting runs well ahead of generative AI deployment, so a survey that folds both together can land anywhere depending on the mix.
Fifth, recency and question wording. The Census Bureau rewrote its core AI question in November 2025, moving from “in producing goods or services” to “in any business function.” That single change roughly doubled the national headline rate, from about 10% to about 18%, and broke the time series. A 2024 number and a 2026 number from the “same” survey may not be comparable. The 10% was a national figure tied to the old wording, not a measure of plant-floor deployment.
How to use these numbers safely
Cite the definition and the date with every figure. “29% using AI/ML at the facility level (Deloitte, fielded 2024, very large manufacturers)” is defensible. “AI adoption in manufacturing is 95%” is not, because it hides the intent-versus-deployment gap and the vendor sample behind it.
Keep the three stages separate in your own reporting: deployed in production, piloting, and planning to invest. If a source won’t tell you which one it measured, downgrade it.
For population-level claims about how many manufacturers use AI, lean on nationally representative data like the Census BTOS. For the deploy-versus-pilot split, a large consulting survey like Deloitte’s is reasonable, as long as you carry the large-firm skew and the fielding date as caveats.
Treat “implemented to some extent” vendor numbers as broad any-use indicators, not deployment measures. They’re the least useful figures in the category for answering the production question.
Where teams go wrong
The most common error is quoting Rockwell’s 95% as if manufacturers have AI in production. That figure measures invested-or-plan-to over five years, and it doesn’t establish production deployment either way.
The second is reading MIT’s 95% as a pilot failure rate, or as a manufacturing statistic. It’s cross-industry, it measures P&L impact rather than deployment, the sample is small, and the authors call it directional.
The third is treating “uses AI somewhere” as “AI in production.” A finance team running a chatbot is a real AI user and tells you nothing about the shop floor.
The fourth is quoting the Rootstock 77% as a deployment rate. “Implemented to some extent” among a vendor’s survey sample of large firms is a different measure entirely.
The fifth is stitching numbers from different surveys into a fake trend line. A 2023 production-only figure and a 2026 any-function figure aren’t two points on the same curve.
Methodology
We pulled the specific deploy-versus-pilot splits where surveys publish them, anchored the population floor to nationally representative government data, and rated each source on a single question: how much should you trust this as evidence of AI running in production in manufacturing.
That question combines two things on purpose. A survey can be competently run and accurately quoted yet still rate Low, because it answers a different question. Rockwell’s 95% is a legitimate investment-intent result and a poor deployment proxy, so it rates Low here without any implication that the survey was done badly. The same logic puts Rootstock’s traceable, vendor-sponsored 77% in the Low tier.
Vendor marketing blogs were not used as load-bearing citations. Where a widely repeated figure came from a vendor-commissioned survey, we named the sponsor, the sample, and the wording rather than launder it into a clean statistic.
The Reliable Confidence Score rates confidence in the claim, not always in a number. A government estimate that manufacturing trails the national average is a High-confidence claim even though AI adoption itself keeps moving. These are survey-driven numbers in a fast-moving category, so treat any single figure as a snapshot with a definition and a date attached, not a settled benchmark.
The Short Version
Manufacturing AI adoption depends entirely on what you’re counting. Nationally representative government data (Census BTOS) puts manufacturing AI use, in any business function, below the ~18% all-industry average. Among the very large manufacturers Deloitte surveyed, 29% reported using AI/ML at the facility or network level and 23% were piloting it; for generative AI, 24% reported deployment and 38% were piloting. Higher figures like 77% (Rootstock) and 95% (Rockwell) use broader definitions or include future investment and shouldn’t be read as plant-floor deployment rates. Cite the definition and the date, or the number means nothing.
Sources
- US Census Bureau, AI Use at U.S. Businesses (BTOS), May 2026: https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
- US Census Bureau, The Microstructure of AI Diffusion (working paper, 2026 BTOS AI supplement): https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
- Federal Reserve, Monitoring AI Adoption in the U.S. Economy (FEDS Notes, Apr 2026): https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html
- Deloitte, 2025 Smart Manufacturing and Operations Survey (fielded Aug–Sept 2024): https://www.deloitte.com/us/en/insights/industry/manufacturing-industrial-products/2025-smart-manufacturing-survey.html
- McKinsey, The State of AI in 2025: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Rockwell Automation, 10th Annual State of Smart Manufacturing Report (press release, fielded Mar 2025): https://www.rockwellautomation.com/en-us/company/news/press-releases/Ninety-Five-Percent-of-Manufacturers-Are-Investing-in-AI-to-Navigate-Uncertainty-and-Accelerate-Smart-Manufacturing.html
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (Challapally, Pease, Raskar, Chari; July 2025; preliminary findings), published by MIT’s Project NANDA, whose official GitHub points to its research repository (https://github.com/projnanda). Third-party mirror of the report PDF: https://www.artificialintelligence-news.com/wp-content/uploads/2025/08/ai_report_2025.pdf. Accessible coverage with the methodology: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Rootstock Software, 2nd Annual State of AI in Manufacturing Survey (announced Jan 2025; conducted by Researchscape, 369 manufacturers): https://www.rootstock.com/press-releases/rootstocks-ai-survey-shows-82-of-manufacturers-increasing-ai-budgets-for-2025/
- Peterson Institute (PIIE), The adoption of AI by industrial sectors, May 2026: https://www.piie.com/blogs/realtime-economics/2026/adoption-ai-industrial-sectors
