Manufacturers report live AI—but production agents are still mostly an intention

Snowflake’s latest manufacturing analysis contains two numbers that look similar but describe very different levels of maturity: 52% of manufacturing organizations say they have live AI in core supply-chain and operations workflows, while 51% expect to have active AI agents in production within the next 12 months.

That gap matters. The first figure is a claim about systems running now; the second is an expectation about future deployment. Teams evaluating agentic analytics should not use them interchangeably as evidence that autonomous operational agents are already common.

What the survey actually establishes

Snowflake says its Global AI & Data Trends Survey covered 2,050 enterprise leaders, including nearly 300 manufacturing decision-makers. Among that manufacturing group, 41% described their organizations as being in an “initial use cases” phase. At the same time, 52% reported live AI in core supply chain and operations, versus 36% across other industries.

The results suggest adoption is concentrated in high-value operational domains even when organizations do not describe themselves as broadly mature. Manufacturers prioritized operational efficiency (57%), operational research and development acceleration (59%), and product innovation (48%).

But the agent figures are still forward-looking. One quarter of manufacturing respondents characterized their posture toward agentic AI as “cautious exploration,” while 51% expected production agents within a year. The source does not report how many already operate agents in production, nor does it provide independently measured reliability, financial return or incident data for those deployments.

The data boundary is the practical finding

The clearest current constraint is not model capability. Snowflake reports that 43% of manufacturing respondents named fragmented data architecture as their largest obstacle to scaling AI. It also says 67% of manufacturing data remains in disconnected databases, spreadsheets and CSV files, and 54% of manufacturers rely heavily on unstructured visual data such as inspection-camera feeds.

For data teams, that turns an “agent roadmap” into a systems-integration problem. A production agent that must connect equipment telemetry, ERP records, supplier data and visual evidence needs governed joins, time alignment and explicit action permissions before autonomy is meaningful. A natural-language interface over one curated warehouse is not equivalent to an agent that can safely reroute supply, trigger maintenance or alter a production workflow.

The procurement takeaway is straightforward: ask vendors and internal teams to separate live analytical AI, human-approved agent actions, and autonomous production actions in deployment counts and ROI claims. Snowflake’s survey shows substantial operational AI use and strong intent to deploy agents. It does not yet show that operational autonomy is the norm.