In 2024 we published an ebook with five predictions about AI in manufacturing: predictive maintenance, quality control, supply chain optimization, training and assistance, and what we called hive mind problem solving.
Not a lot of people grade their own homework in public. We felt it was time for us to do so. So here is our scorecard. Five predictions, how they held up, and what we completely missed. Every figure behind these verdicts is sourced in the 2026 ebook.
- Three 2024 predictions were right, one was half right, and one evolved into a different mechanism.
- According to a widely shared MIT NANDA study, 95% of enterprise generative AI pilots delivered no measurable P&L impact.
- In that same study, AI bought and embedded into existing workflows succeeded about 67% of the time.
- We missed two major themes: agentic AI, and the EU AI Act. Both get new chapters in the 2026 edition.
- Download the free 2026 ebook for evidence, case studies, and all 49 sources.
Where AI in manufacturing stands in 2026
In 2026, there are two sides of the story that might seem to contradict each other:
Failure
A widely shared MIT NANDA study found that 95% of enterprise generative AI pilots delivered no measurable P&L impact. Gartner placed generative AI in manufacturing in its "Trough of Disillusionment". By the end of 2025, at least half of generative AI projects had been abandoned after proof of concept, worse than Gartner's own forecast of 30%.
Adoption
Even though most AI projects still seem to fail, AI got adopted significantly over these past two years. Roughly 88% of organizations use AI in at least one business function, but only about a third have begun scaling it across the enterprise. Among US manufacturers, 29% run AI or machine learning at facility or network scale. In Europe, Eurostat counts 20% of EU enterprises using AI in 2025, up from 13.5% a year earlier, with manufacturing slightly below average at 17.3% and a wide national spread: 42% in Denmark, under 9% in Poland.
So while experimentation with AI is widely adopted, scaling AI into real business purposes is still rather an exception.
The scorecard
| 2024 prediction | Verdict | Quick summary |
|---|---|---|
| Predictive maintenance | We were right | Moved from emerging trend to default practice. The hard part now is scaling it, not proving it works. |
| AI for quality control | We were right | Vision QC delivered on accuracy, then got cheaper and faster to deploy than we expected. |
| Supply chain optimization | Half right | The forecasting capability landed as predicted. What drove adoption, and what came next, did not. |
| Training and assistance | We were right | AI authoring for work instructions shipped across the board, pushed by labor shortages more than hype. |
| Hive mind problem solving | Evolved | Operators asking AI for help is real. It arrived through language models, not the shared database we pictured. |
What we missed on generative AI and agentic AI
The verdicts are the easy part. The 2026 edition walks through what we said, what happened, and what we missed for each prediction: the numbers behind predictive maintenance, why vision QC commoditized so fast, what actually drove supply-chain AI, why digital work instructions became the urgent entry point, and how the "hive mind" idea we sketched in 2024 got rewired.
But we also just plainly missed two major things: the rise of AI agents caught us by surprise, and we completely ignored the EU AI Act. The full ebook adds two new chapters on both of these topics.
Our own AI experiments
Grading industry predictions is one thing. Running the experiments ourselves is another. That is what Azumuta Labs is for: a research division that connects AI and vision technology to real shop-floor problems, then publishes what we learn in public.
One example is already live. Vision-language models that turn a filmed demonstration into structured digital work instructions, cutting creation time by up to 90% in our trials. We share the method, the tradeoffs, and the results on Labs so other manufacturers can judge the work for themselves.
If you have a pilot, a research question, or a shop-floor problem you want to explore with us, we want to hear it. We collaborate with manufacturers and researchers on real experiments, not slideware.
If you think we got something wrong again, tell us. It becomes input for the next round of experiments.
Frequently asked questions
Quick FAQs to get you up to speed
Three were right: predictive maintenance became standard practice, AI for quality control delivered and then got cheaper to deploy, and AI for training and work instructions shipped as predicted. Supply chain optimization was half right: the capability landed, but adoption was driven by trade policy and agentic systems we did not foresee. Hive mind problem solving evolved: the idea was right, but it arrived through large language models rather than a dedicated shared database. Full evidence and case studies are in the 2026 ebook.
Experimentation is widespread; scale is still rare. Roughly 88% of organizations use AI in at least one business function, but only about a third have begun scaling it across the enterprise. Among US manufacturers, 29% run AI or machine learning at facility or network scale. In Europe, Eurostat counts 20% of EU enterprises using AI in 2025, up from 13.5% a year earlier, with manufacturing slightly below average at 17.3%.
According to a widely shared MIT NANDA study, 95% of enterprise generative AI pilots delivered no measurable P&L impact. By the end of 2025, at least half of generative AI projects had been abandoned after proof of concept, worse than Gartner's forecast of 30%. Standalone "ask our factory anything" chatbots fail especially often when they run on outdated or contradictory procedures. The 2026 ebook covers what the successful minority did differently.
In the MIT NANDA study, AI bought from specialist vendors and embedded into existing workflows succeeded about 67% of the time. Internal builds succeeded far less often. That is one of the most actionable findings in the 2026 ebook: embed AI in tools operators already use, rather than launching a standalone pilot.
No. In the World Economic Forum's 2025 showcase factories, 77% of top use cases ran on analytical AI and only 9% on generative AI. Generative AI is the fastest-growing layer on top of that foundation, through copilots and agents, not a replacement for predictive maintenance, quality inspection, or planning analytics.
Agentic AI systems detect a disruption, re-plan, draft actions such as purchase orders or task priorities, and wait for a human to approve. Gartner expects half of all supply chain management solutions to include agentic capabilities by 2030, and forecasts $53 billion of spend on supply chain software with agentic AI by then, up from under $2 billion in 2025. Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The 2026 ebook has a dedicated chapter on copilots and agents.
Often no. Work instruction assistants, quality analytics, predictive maintenance, and scheduling copilots are generally not high-risk under the EU AI Act. Two areas do deserve real attention: AI acting as a safety component of machinery, and AI that monitors or evaluates workers. The 2026 ebook covers what European manufacturers need to know, including the compliance dates for their teams.
Between 2024 and 2026, speech-to-text authoring, video-to-step segmentation, translation, and personalization moved from roadmap slides to standard product features. Vendors routinely claim 60% to 70% less time spent on documentation. In Azumuta Labs trials, vision-language models turned a filmed demonstration into structured digital work instructions with up to 90% less time spent on creation, editing, and maintenance.
The free 32-page edition is available as a free download. It grades all five 2024 predictions with evidence and case studies, adds chapters on copilots, agents and physical AI, covers the EU AI Act, and includes action checklists plus all 49 sources.
