In the report, The Great Acceleration, Manufacturers Alliance found that 63% of manufacturers needed to clean up and map their data before they could begin AI pilots at all. More than 20% required major audits and corrections. Only 3% found their data ready as it stood.
This obstacle ranking is worth further investigation. Thirty percent named inaccessible or questionable data quality as their single biggest barrier, and 21% named inconsistent data formatting across plants. In the Alliance’s 2024 study of AI, poor data quality ranked sixth among obstacles. Two years later it ranks first — not because the data got worse, but because organizations started to build AI pilots with that data and hit a roadblock.
This Is Governance, Not Hygiene
Inconsistent formatting across plants is not a cleanup task. It is the absence of a common standard and of a named leader empowered to enforce one. The same pattern shows up in two further studies: a recent report from Manufacturers Alliance and one from Cisco.
The Infrastructure Bottleneck asked what obstacle prevents upgrading connectivity for AI. The answer was not capital. Alignment friction between IT and OT over ownership, budget, and requirements led at 59%, more than double capital expenditure at 28%. Ninety-two percent had hit a network capacity problem in the previous year that delayed data or stopped lines, and only 7% were very confident their networks could carry three years of planned AI growth.
Cisco’s State of Industrial AI 2026 found the same mechanism independently: 43% report limited or no IT/OT collaboration, and among those, 47% cite network instability as a top obstacle to scaling AI. Unresolved ownership produces an unreliable data path, and an unreliable data path stops AI before the model is ever the question.
You Cannot Govern What You Have Not Inventoried
Governance also presumes you know what you have. Honeywell’s 2026 OT Cybersecurity Benchmark found 88% of security leaders rated their security programs mature or design-led. Two further questions in the same study asked about practice rather than self-assessment. Only 21% hold a complete asset inventory, and only 20% continuously monitor more than three-quarters of their connected devices.
That gap between confidence and verification recurs at larger scale. The IFS Industrial Index 2026 found over 70% confident in their AI safety and data policies, yet only 11% had actually validated AI outputs for safety- or mission-critical processes.
The Security Half of the Same Problem
Weak data governance not only slows AI down; it is where the losses come from. A Manufacturers Alliance survey of cybersecurity and IT leaders at enterprise manufacturers shows how that risk sits inside the security function. Seventy percent of respondents now utilize AI for cybersecurity. Thirteen percent describe their enterprise AI governance framework as fully implemented.
But what is the AI being used to watch? Respondents reported AI running in security operations at 83% and in phishing and email security at 67%. However, in data security, only 13% of respondents report using AI. AI is being used to watch the network and the inbox, but it is rarely being utilized on the data those systems are supposed to be protecting.
The control requirements follow the same priority. Approved use case intake is required before an AI solution goes live by 87% of respondents. Data classification and handling review is required by 57%, but logging and monitoring, once the system is running, is only required by 30%.
Controls Required for Enterprise AI Solutions
Source: Manufacturers Alliance Cybersecurity KPI, 2026.
Nearly nine in ten manufacturers ask whether an AI use case is approved, but fewer than six in ten ask what data it will touch and how that data is classified. Fewer than a third will be able to reconstruct what it did.
Asked what most obstructs AI adoption, respondents named skills first at 53%. But 20% named data quality outright, and another 20% named privacy or regulatory concerns. Those are not model problems. They are data governance problems surfacing in a cybersecurity survey.
How to Catch Up on Data
The uncomfortable implication is that manufacturers are further along with AI than with the data it runs on.
Inconsistent formatting across sites was the single largest obstacle for 21% of manufacturers, and alignment friction outranks capital three to one as a connectivity barrier. Both are decisions about who sets the standard and who enforces it. The manufacturers making progress named an owner for data standards across plants before they scaled anything.
As you move forward, make data classification a requirement rather than a nice-to-have. At 57%, it trails use-case approval by thirty points, which means a meaningful share of approved AI is already running on data nobody has classified.
Inventory comes before automation, and the order is not optional: 21% of security leaders hold a complete asset inventory while 88% rate their programs mature. Automated monitoring cannot baseline what was never catalogued, and neither can an AI model.
For members of Manufacturers Alliance, full results from this survey are available in the exclusive member-only report: Cybersecurity in Manufacturing 2026: Governance, Investment, and AI.
AI Transparency: Content for this article was analyzed and written with assistance from an AI tool and reviewed by the Manufacturers Alliance research team.
