Key Takeaways from IIOT World: The GenAI Divide in Manufacturing - Adoption Is High, Transformation Is Low

Manufacturers have raced to explore GenAI but Deloitte's 2025 Smart Manufacturing Survey and MIT's 2025 State of Ai in Business tell a tougher story. In our keynote session at IIOT World's AI Frontiers 2025 in Smart Manufacturing, Chris Huff (CEO, Adlib) framed it plainly: “Adoption is high; transformational impact is low.” The data from MIT and Deloitte explain why, and, more importantly, where leaders should intervene to build data accuracy and AI trust on the factory floor.

‍

Inside the Numbers: Shadow AI, Stalled Pilots, and the Data Gap

1) The GenAI Divide

‍MIT finds that 95% of enterprise GenAI initiatives show no measurable P&L impact, and only ~5% of task-specific tools reach production. External partnerships are deployed 2x more often than internal builds, evidence that collaboration accelerates value. As Chris summarized: “External partnerships are successfully deployed twice as often as internal builds.”

‍

2) Shadow AI

‍Employees adopt faster than enterprises: nearly 90% of workers use LLMs personally, while only ~40% of companies have official subscriptions. Translation: your workforce may already be ahead of your program. “That creates both risks and opportunities,” Chris noted.

‍

3) Why pilots fail

‍Top barriers aren’t surprises: change management, lack of executive sponsorship, poor UX, and output quality concerns. Chris’ takeaway: “The days of handing employees static, clunky tools isn’t going to cut it. To scale AI, we have to deliver both accuracy and usability.”

‍

4) Where manufacturing really is (2025)

‍Deloitte shows commitment is real: 78% of manufacturers allocate >20% of improvement spend to smart manufacturing; 0% expect a decrease next cycle. Investments cluster around automation and data, the foundations for AI.

‍

5) Scaling remains the choke point

‍Only a quarter have AI/ML scaled across multiple facilities. “A single pilot can run on local heroics… scaling to 10 or 20 plants requires a repeatable playbook, standardizing how data is structured, reusable templates, and lowering the barrier to onboarding,” says Chris. And measure success in business terms (downtime, changeover speed), not model counts.

‍

Why trust breaks (and how to fix it)

Chris’ core diagnosis is consistent across the slides: “To scale from experimental to scaled AI requires accurate, contextual, and trustworthy data.” In manufacturing, most of that data starts as documents, CAD, SOPs, inspection logs, batch records, specs, drawings. If these inputs are messy or inconsistent, downstream AI will stall.

His 90-day, document-first plan focuses upstream:

1. Standardize capture & formats. “Harmonize formats across CAD drawings, inspection logs, SOPs… so you’re not feeding AI a patchwork.”

2. Validate quality at the source. “Use automation to flag missing fields and inconsistencies before they flow downstream.”

3. Enforce governance. “Access, lineage, accountability so every piece of data has an owner and provenance.”

How you’ll know it’s working?

‍“Exceptions trend down. Straight-through processing trends up. Most importantly, your engineers stop second-guessing the data because they trust it.” say Chris.

‍

Why document automation & data preprocessing is the unlock

Two realities show up in the Q&A and live poll:

Factory files aren’t “Office docs.” Large language models struggle with non-standard formats (multi-layer CAD, scans, images, handwriting). You need a preprocessing layer that converts, chunks, and annotates content so AI can reason over it.  

Human-in-the-loop is still heavy. A large portion of document traffic still needs manual review. Chris’ guidance: push quality upstream so HITL can be throttled down over time, the same trust curve we rode with RPA.

‍

The ROI case (from real manufacturing programs)

When you standardize formats, validate at the source, and turn documents into AI-ready, searchable, contextual data, the payback shows up fast:

• $1.1M in engineers’ time saved by automating CAD→PDF creation; $186,500 in licensing avoided; reduced non-compliance risk.
• ‍10% PLM efficiency increase, saving millions via fewer delays and errors.
• Up to 98% faster document processing and 12-year accessible, audit-ready records (PDF/A).
• 60% licensing cost reduction and automated, compliant records across 700k+ production/quality docs annually.

This is why Chris keeps steering leaders to leading indicators: fewer exceptions, higher straight-through rates, and rising user trust. Those are early proof points that precede the P&L impact skeptics are scanning for in filings.

‍

Chris Huff’s closing note

“We’re witnessing a paradox: massive investment in AI and smart manufacturing, but limited, measurable transformation to date. The common thread is trustworthy data.”

If you’re serious about closing the GenAI divide, start upstream: transform your documents into validated, governed, AI-ready data, then watch pilots scale.

‍

Watch the full session

See the full discussion for building Data Accuracy & AI Trust in Smart Manufacturing (with Chris Huff, Anthony Vigliotti, and industry experts). It’s a pragmatic blueprint for turning unstructured factory content into reliable AI outcomes at scale.

‍