88% of US mid-market manufacturers run AI.

The factory floor has quietly become one of the most AI-dense places in the economy. In RSM's 2026 survey of US mid-market manufacturers, 88% said AI is at least partly integrated into their organisation, 32% fully across core operations, and 56% already use agentic AI. The machines are ready too: the International Federation of Robotics counts 5 million industrial robots in operation worldwide, a record. And yet the same RSM survey names the reason pilots stall: data quality, behind 51% of pilots with limited success. Unclear return comes second, at 30%.

88% of US mid-market manufacturers have AI at least partly integrated; data quality is behind 51% of weak pilots (RSM, Jul 2026)

That is the manufacturing story of 2026. Adoption is no longer the question. Scale is, and scale depends on the data and the decision discipline behind each use.

Robots at record, models on the shop floor

More than 600,000 industrial robots were installed worldwide in 2025, up 11% (IFR, September 2026). Density is highest in Korea, with 1,220 robots per 10,000 manufacturing employees; Germany ranks third with 449, almost twice the EU average of 231, and the US stands at 307 (IFR, April 2026). Beyond the cell, growth is faster still: almost 250,000 professional service robots were sold in 2025, up 24%, and IFR counted about 7,000 full-size humanoids for the first time.

449 industrial robots per 10,000 manufacturing employees in Germany, third worldwide (IFR, Apr 2026)

What changes is the brain. Robot foundation models such as NVIDIA GR00T and Google DeepMind's Gemini Robotics make skills like grasping unfamiliar items available across many machines. Vision control, autonomous motion and natural-language programming move from research into products from Universal Robots, Yaskawa, Micropsi Industries and Wandelbots. The Robotics & Physical AI brief maps these offerings and the safety rules that apply.

Robots and cameras lead: 50 of 96 AI offerings

Manufacturing is one of two areas where the Good AI Prism goes beyond a listing. It places 99 entries in Manufacturing & Industrial, 96 AI offerings and three provider records (NVIDIA, Agile Robots, Physical Intelligence), and the research edition of 20 October 2026 holds 87 of the offerings researched in depth, each with a business application, pricing unit, integrations, deployment prerequisites, data hosting and limitations where the provider documents them. Robots and cameras lead: 25 offerings bring AI into robots and 25 inspect quality by camera. Eighteen predict maintenance needs, 11 optimise processes, nine are engineering copilots and knowledge assistants, and eight cover plant data, planning and edge AI.

Maintenance shows the range. AspenTech's Aspen Mtell, AVEVA Predictive Analytics, C3 AI Reliability, GE Vernova's asset performance management, Honeywell Forge and Festo AX sit next to specialists such as Nanoprecise, Petasense and Asystom. Engineering copilots come from the automation majors: Rockwell Automation's FactoryTalk Design Studio Copilot, Dassault Systèmes' LEO, KUKA's iiQWorks.Copilot and fruitcore robotics' horstOS.

The depth research also shows what buyers cannot see. Only 12 of the 87 records state a pricing unit on public pages; 51 document where data is hosted. For a buyer that means one thing: the decisive facts arrive in the sales conversation, so they must be asked for in writing.

Why pilots stall

Data quality sits behind half of the pilots that disappoint, according to RSM. In practice that means sensor data with gaps, quality records that do not link to process parameters, and maintenance logs written for people rather than models. A use-case portfolio with a data-readiness check in front of each case filters this early. The second reason, unclear return, is a decision problem: a pilot without a baseline and a threshold has no way to end.

The size gap makes it harder for smaller firms. In Germany a majority of firms now use AI (Bitkom survey, 2026), but the depth of use differs by size. In Korea, only 4.2% of small and mid-sized manufacturers use AI, and 80.7% have no AI talent (KCCI, cited in the Good AI Prism Global Market Study). Supplier programmes run by large manufacturers are one way to close that gap.

The rules: machinery law first

In Europe, robot safety rests on machinery law. From 20 January 2027, the Machinery Regulation (EU) 2023/1230 replaces the Machinery Directive. Its essential requirements cover self-evolving behaviour, AI-enabled safety functions and cyber-safety, with third-party assessment for listed high-risk machinery. The AI Act adds high-risk duties for AI in products under EU harmonisation law from 2 August 2028, as amended by Regulation (EU) 2026/1744; Commission guidance on the interplay is due by 1 August 2027.

Standards fill in the detail. ISO 10218-2:2025 for robot cells on robot applications and cells, with Part 1, is the reference for industrial robot safety, alongside ISO/TS 15066 for collaborative operation; the US adopted them as ANSI/A3 R15.06-2025. In Germany, works councils co-determine the introduction of technical systems that can monitor performance (BetrVG § 87(1) no. 6) and must be informed early about new technical installations (§ 90).

Seven questions before a pilot

Before AI touches a line, settle seven questions in writing. Which outcome should move, and what is today's baseline for downtime, scrap rate or OEE? Is the data there: complete sensor history, quality records linked to process parameters, maintenance logs a model can read? Does the system act on a safety function, so that the Machinery Regulation's conformity rules apply? Where is the data hosted, and does the provider train on it? How does it integrate with the control systems, the MES and the ERP, and can we exit? Who on the shift can override it, and how fast? And what result scales the case, and what result stops it? A provider who answers these in writing is ready for a pilot. The why AI does not scale article explains the operating model behind the answers, from ownership to measurement.

People and the plant

The evidence on jobs depends on how adoption is managed. In US labour markets, one more robot per 1,000 workers reduced the employment-to-population ratio by 0.2 percentage points (Acemoglu and Restrepo, 2020). In Germany, robots did not reduce total local employment between 1994 and 2014; manufacturing losses were offset in business services, although fewer young people entered manufacturing (Dauth et al.). The lesson for a plant manager: plan training, involvement and oversight with the technology, not after it.

In practice

Best practice: a data-readiness gate. Before a use case enters the portfolio, a short check asks three questions. Is there continuous sensor history for the asset? Are quality records linked to process parameters? Can a model read the maintenance logs? Cases that fail go to a data backlog, not to a pilot. That one gate addresses the cause RSM found behind half of the disappointing pilots.

Scenario: predictive maintenance on a press line. A mid-sized supplier pilots a maintenance model on one press line. The baseline is unplanned downtime and scrap rate over the past twelve months; the threshold is the reduction that pays for licence, integration and a planner's time. If the model misses it in the agreed window, the case stops and the data lessons move to the next line.

What to do Monday

1. Build a portfolio, not a pilot. Pick three to five uses across maintenance, inspection and copilots, each with a data-readiness check.
2. Fix the stop/go gate before the start. Agree the baseline and threshold for downtime, scrap rate and OEE; an AI Opportunity Case sets both.
3. Ask for the hidden facts in writing: pricing unit, data hosting, training on customer data and integration with your control systems.
4. Prepare the conformity file. Map every AI-enabled safety function to the Machinery Regulation before 20 January 2027.
5. Involve the works council early, wherever a system can monitor performance.

This article describes market and regulation. It is not legal or safety-engineering advice.