Everyone’s talking about AI. Almost nobody’s talking about their data.

I’ve spent more than 20 years in manufacturing, and I’ve never seen a technology wave hit the trade-show floor as hard as AI has. Every booth has a copilot. Every vendor has an agent. Every keynote promises the lights-out factory is two quarters away.

Then I walk back onto a real shop floor. The quote is still built in a spreadsheet. The router lives in someone’s head. The ERP was installed in 2009, customized beyond recognition, and half the team still doesn’t trust what it says.

Here’s my thesis, and I’ll back it with data below: AI is not the constraint in most small and mid-sized manufacturers. Data is. And the ERP system is where that constraint lives. If you’ve studied Theory of Constraints, you know what that means. Investing anywhere other than the constraint is just building inventory in front of the bottleneck.

Where adoption actually stands: it depends on who you ask.

If you only read vendor surveys, you’d think AI adoption in manufacturing is basically done. Look closer and two very different pictures appear.

Source	Who They Asked	Headline Finding
Rootstock 2026 State of Manufacturing Technology	Manufacturers surveyed by a cloud ERP vendor	94% use some form of AI
RSM Middle Market AI Survey 2026	129 middle-market manufacturers	88% at least partially integrated; 32% fully integrated
Parsec 2026 State of the Manufacturing Industry	1,200 global manufacturing leaders	72% adopted AI; only 10% at scale
U.S. Census Bureau BTOS, analyzed by Digit Software	Broad sample of U.S. manufacturers	13.9% use AI in any business function (Feb 2026)

That’s the gap I want every owner to sit with. The surveys that make headlines skew toward larger, more digitally engaged companies. The Census data, which reaches the shops that never answer a vendor survey, says roughly nine out of ten U.S. manufacturers still aren’t using AI at all. It also shows large manufacturers are 2.3 times more likely to use AI than small ones.

Even among the adopters, scale is the problem. Parsec found just 10% have deployed AI across their operations; the rest are piloting, experimenting, or “actively implementing.” Deloitte’s 2026 AI in Manufacturing survey, as reported by Manufacturing Digital, found only about one in five AI use cases had been scaled consistently across sites.

So the honest summary is this: big companies are piloting everywhere and scaling a little. Small and mid-sized shops, which is most of us, are mostly still on the sidelines.

Where AI is actually earning its keep.

When AI works in manufacturing right now, it’s narrow, practical, and pointed at a known pain. Parsec’s respondents ranked quality control (50%), IT operations (46%), and supply chain management (45%) as their top use cases. RSM found 79% of middle-market manufacturers using generative AI and, surprisingly, 56% already experimenting with agentic AI.

The use case I’m watching most closely is the one closest to my world: quoting and estimating. In a job shop, the quote is the front door. Price too slow and nothing else matters.

The tool market here moved fast in 2026:

• Uptool came out of stealth in February 2026 targeting small, high-mix machine and fab shops. It reads RFQs straight from the quoting inbox, pulls materials and finishing steps off the drawing, and applies the shop’s own rates. In an IMTS feature, Bullray MFG’s owner said quotes that took him two hours now take about ten minutes. (Worth noting: that’s a vendor-sponsored story, so treat it as a ceiling, not an average.)
• CloudNC raised $20 million and demoed a Quote Agent at IMTS 2026, aimed squarely at the hours shops spend interpreting drawings and estimating cost.
• Established quoting platforms such as Paperless Parts have layered AI assistants onto existing estimating workflows.

We’re building in the same direction at Big Rocks. Our AI estimating agent concept reads 2D drawings and 3D CAD, extracts the manufacturing features, drafts the router and BOM, and closes the loop with win/loss data. But here’s what I learned designing it: the AI layers were the easy part to imagine. The hard part was the layer labeled ERP integration.

Why AI pilots stall: it’s rarely the AI.

Read the challenge data across these surveys and a pattern jumps out. The model is almost never the problem. The plumbing is.

RSM asked manufacturers whose AI pilots had only moderate or limited success what went wrong. Data quality issues topped the list at 51%, followed by integration challenges (43%), security or compliance concerns (34%), unclear ROI (30%), and adoption resistance (30%). Parsec’s barrier list rhymes: high implementation cost (40%), data privacy and security (39%), and difficulty integrating with existing systems (38%).

Then look at the environment those pilots are landing in:

• Mixed generations of technology. 69% of manufacturers in Parsec’s study run a hybrid of legacy and modern equipment, and only 37% have a unified, data-driven strategy in place.
• A people gap. 84% of RSM respondents agreed leadership is more excited about AI than employees are, yet only 22% are focusing AI spend on upskilling their own people. NAM’s Q2 2026 outlook survey found 41% of manufacturers provide no AI training to frontline workers.
• Talent we can’t hire. Parsec found IT and tech specialists are the hardest roles to fill (60%), and 45% cite internal skill gaps when integrating new systems.

One more number from RSM worries me. Manufacturers increasing AI spend said they’d likely pull funding from legacy application modernization (40%) and business intelligence (50%). In other words, some companies plan to fund the AI by starving the very data foundation the AI depends on. That’s the TOC equivalent of speeding up a non-bottleneck machine while the constraint starves.

The ERP connection: AI is only as smart as your system of record.

Here’s the part most AI conversations skip. Every useful manufacturing AI application needs three things: clean part and process data, accurate actual costs, and a feedback loop that compares what you estimated to what really happened. In a manufacturing business, all three live in, or should live in, the ERP.

And that’s exactly where many established manufacturers are weakest. Industry estimates have long put a large share of privately held U.S. manufacturers in the hands of Baby Boomer owners, many without a succession plan. Those companies tend to run on one of two things: an ERP installed 15 to 25 years ago and bent around decades of workarounds, or no real ERP at all. The data on the “no ERP” side is sobering. In MRPeasy’s July 2026 analysis of 70 small-manufacturer implementations, 64% were running on spreadsheets before switching, half relied on accounting software like QuickBooks or Sage, and 74% said they lacked visibility into their own operations. The top implementation hurdle wasn’t software. It was data cleanup and migration (25%), followed by employee resistance (19%). That’s a vendor’s customer sample, so it skews toward success stories, but the pattern matches what I see in every plant I walk.

Now connect the dots back to AI:

• The Census data showed large manufacturers are 2.3 times more likely to use AI than small ones. Size is a proxy for something else: data infrastructure.
• RSM’s top reason AI pilots fail is data quality (51%), and integration with legacy systems is a top-three barrier.
• Parsec found nearly 7 in 10 manufacturers straddle generations of equipment and systems.

Some shops are skipping the ERP question entirely. One machine shop owner quoted in that IMTS piece said she doesn’t need a big, expensive ERP; a focused AI quoting tool does the job. For a one-person shop, that may be exactly right. But for a growing company, a bolt-on AI tool that never talks to your system of record just builds a smarter silo. Your quote engine learns from your quotes, not from your actual run times, scrap, and margins. The win/loss loop never closes.

That’s the trap. AI doesn’t replace the need for a solid ERP foundation. It raises the stakes on it. Bad data used to cost you a bad report. With AI, bad data gets automated, scaled, and delivered with confidence.

A Theory of Constraints playbook for AI in a mid-sized shop

You don’t need a seven-figure AI budget or a rip-and-replace ERP project to get started. You need to treat AI adoption like any other throughput problem. Goldratt’s five focusing steps work just as well here as they do on the shop floor:

1. Identify the constraint. For most shops it’s not AI capability. It’s the quote desk, the scheduler, or the fact that nobody trusts the ERP’s cost data. Find the one place where faster, better decisions would release the most throughput.
2. Exploit it with what you have. Before buying anything, get your existing data out where you can use it. Clean up routers, standard times, and item masters for your top revenue part families first, not all 40,000 parts.
3. Subordinate everything else. Don’t fund AI by cutting the data modernization it depends on. If the RSM trade-off data describes your budget, rethink the budget.
4. Elevate the constraint. Now bring in AI, pointed directly at the bottleneck, with a human in the loop and a clear success metric like quote turnaround, hit rate, or margin accuracy.
5. Don’t let inertia become the constraint. Once quoting is fixed, the bottleneck moves. Go find it.

Here’s how that’s playing out at Big Rocks across our five businesses. We run ECI’s E2 ERP, hosted with a managed service provider. Rather than wait for a perfect ERP, we set up a read-only extract that copies E2 data every 15 minutes into an Azure environment we control. That gives us a clean place to build without touching production.

The first project wasn’t AI at all. Our new information systems integrator is building KPI dashboards for each company, starting from the spreadsheets each one already uses. Only once that data foundation is trustworthy does the AI estimating agent make sense, because its most valuable layer is the one that compares what we quoted to what actually happened on the floor.

That’s the whole lesson in one sentence: fix the data flow first, then let AI accelerate it. The manufacturers who win the next five years won’t be the ones with the most AI pilots. They’ll be the ones whose data was ready when the tools arrived.

Sources and Further Reading:

• How manufacturers are using AI in 2026, RSM US, July 2026 (129 manufacturers from the RSM Middle Market AI Survey)
• Parsec Survey: 72% of Manufacturers Have Adopted AI, but Only 10% Have Done So at Scale, Parsec Automation, July 2026 (1,200 manufacturing leaders)
• Report: AI Adoption in U.S. Manufacturing Has Quadrupled, But 87% of Facilities Still Haven’t Made the Switch, Automation.com, May 2026 (Digit Software analysis of Census BTOS data)
• Survey: AI adoption continues apace, despite uncertainties and failure risk, Smart Industry, Feb 2026 (Rootstock 2026 State of Manufacturing Technology)
• Is Your Factory Ready for AI?, Manufacturing Digital (summarizes Deloitte, Rockwell, and NAM 2026 findings)
• ERP Statistics: Insights From 70 Manufacturing Case Studies, MRPeasy, July 2026
• Quote Us: AI for Faster, Better Quotes, IMTS 2026
• CloudNC raises $20M to expand AI tools for precision machining, Robotics 24/7, Sept 2026
• The Goal by Eliyahu M. Goldratt: still the best book on why you fix the constraint first