88% of AI pilots never make it to production.
For every 33 proof-of-concept manufacturers build, only 4 reach the factory floor. The rest die in "pilot purgatory" i.e, demos that never impact the P&L.
I started this newsletter because I kept seeing the same pattern: plant managers excited about AI, IT teams running pilots, boards asking about ROI, and nothing shipping to production.
A cement plant hit 57x ROI in six months with predictive maintenance, and without capex.
It’s the Implementation that determines who captures these gains.
👋🏻 I'm Leonardo Ubbiali, founder of Visum Labs. Each week, I'll share what's working on factory floors: wins, failures, numbers, and implementation details.
Two-thirds said their AI implementations remain at "exploration or targeted-implementation stage."
Meanwhile, 77% of manufacturers with $10B+ revenue have scaled AI successfully. Mid-market adoption sits at 2-4%.
That gap will define who leads manufacturing in 2030.
Why pilots fail:
RAND Corporation studied AI project failures across industries. The top causes weren't technical:
Misunderstanding the problem before building the model
Poor data quality (this alone kills 70%+ of projects)
AI applied to problems too difficult for current technology
Unclear business objectives with vague success metrics
Most manufacturers flip that ratio. They overspend on technology and underinvest in organizational change.
What separates companies that scale:
BCG found that AI leaders prioritize an average of 3.5 use cases. Laggards spread resources across 6.1. Going deeper on fewer initiatives generates 2.1x greater ROI than spreading thin.
McKinsey's COO survey revealed that companies with clear KPI targets meet or exceed AI goals 65% of the time.
Yet 60% of manufacturers lack AI targets entirely.
Accountability mechanisms force the investment discipline pilots need to scale.
Five things you can do this quarter:
The problem: You know your scrap rate is too high, but you don't know why. The data exists somewhere in your MES, ERP, or even just Excel files. But who has time to analyze six months of production logs?
What you need:
Your defect/scrap data exported to CSV (date, shift, machine, operator, defect type, quantity, any other variables you track)
Claude, ChatGPT, or any LLM with file upload
10 minutes
The prompt (copy this):
I'm a plant manager at a [YOUR INDUSTRY] manufacturing facility. I've attached our scrap/defect data from the past [TIME PERIOD].
Analyze this data and answer:
⦁ Which machine, line, or work center has the highest defect rate? Is it consistent or did it change at some point?
⦁ Is there a shift pattern? (First shift vs. second shift vs. third shift)
⦁ Is there an operator pattern? (Be careful here - I want to know if training is needed, not to blame individuals)
⦁ Is there a time-based pattern? (Day of week, time of month, seasonal)
⦁ Is there a correlation with any other variable in the data? (Material batch, supplier, ambient conditions if tracked)
⦁ What's the single highest-impact root cause I should investigate first?
Show me the data behind each conclusion. I need to present this to my team with specific numbers.
What you'll get back:
Claude will break down your data by each variable and show you where the patterns are. Here's a sample of what the Ohio plant manager saw:
"Extrusion_E2 accounts for 16% of total production but 31% of defects. However, this wasn't always the case. The machine's defect rate was stable at 6.2% January through March, then suddenly jumped to 10.9% starting in April i.e, a 76% increase.”
“Looking at the time series, the deterioration accelerated each month: April (7.3%), May (12.4%), June (13.2%). This points to a mechanical failure that's progressively worsening.”
“Additionally, we found that Operator OP_105 has a 6.2% defect rate compared to the facility average of 4.3%, but only works second shift. When we isolated second shift performance excluding OP_105, the rate dropped to 4.4%, suggesting a training gap rather than a shift-wide problem.”
Why this works:
You already have this data. It's sitting in your system right now. Most plants have months of production logs that nobody has time to dig through. AI does in minutes what would take your quality engineer days.
One caveat: The output is only as good as your data. If you're tracking defects as just "scrap" without defect type, machine, shift, and date, you'll get generic answers. The more variables in your export, the more specific the root cause.
McKinsey: "From Pilots to Performance"
A detailed breakdown of what differentiates manufacturers that scale AI from those stuck in pilots. Includes a framework for redesigning production processes, building scalable tech architecture, and driving adoption. Skip to the "Three building blocks" section for the actionable framework.
Time to value: 20 minutes
That's it for Edition #1.
If 88% of pilots fail because of people and process problems (not technology), why do we keep treating AI implementation like a technology project?
Hit reply. I read every email.
Leo