The 100x problem: Why agents redefine your factory data infrastructure

During a recent conversation on the Industrial AI Podcast, I shared a way of thinking about data consumers that I believe is important for every manufacturing leader to understand.

In Industry 3.0, 1 to 5 systems consumed shop floor data, such as SCADA, MES, and historians. Industry 4.0 and the introduction of the Cloud pushed that to 10 to 50 consuming systems with data lakes, analytics platforms, ERP integrations, IoT services, and more.

With the advent of agentic AI, I expect to see a 100x increase in edge-based consumers. Thousands of autonomous AI agents will be implemented, each requiring specific, contextualized data delivered in real-time.

This isn’t a theoretical exercise. It’s a scaling problem, and one that the existing data architectures most manufacturers have in place were never designed to solve.

Why Agents Impact Architecture Conversations

To understand why so many agents are required, consider how AI agents work in practice. Unlike a dashboard that pulls a pre-defined dataset on a schedule, agents are goal-oriented and highly specialized.

For a single work cell, you might deploy a quality agent, maintenance agent, scheduling agent, and supply chain agent. Each agent needs different data, from different sources, with different context. For example, the maintenance agent needs asset service history from the CMMS combined with real-time vibration data from the OPC server. The quality agent needs batch data from the MES combined with inspection results. The supply chain agent needs order data from the ERP paired with production rates.

Now multiply that across every work cell, every line, every site in an enterprise. The volume quickly becomes staggering.

The traditional ISA-95 approach of moving data layer by layer was never designed for this. Instead, agentic AI needs a hub-and-spoke model for data integration that can connect to any system and deliver curated data wherever it’s needed.

The Context Problem Only Gets Harder

If the volume problem is challenging, the context problem is even greater. A raw PLC tag is meaningless to an AI agent. Each agent requires context that is entirely dependent on what it’s trying to accomplish.

For example, a maintenance agent analyzing a pump needs pressure data, service history, vendor info, and batch context coming from several systems like the OPC server, MES, CMMS, and possibly the ERP. A quality agent for the same pump needs an entirely different contextual package: batch identification, product specifications, regulatory thresholds, and alarm history.

This is where AI initiatives stall. According to a recent IDC report, 56.6% of industrial organizations are either planning, piloting, or in the early stages of using AI agents (#US52869825). Their success will depend on the presence of a data strategy that supports data contextualization from multiple OT and IT systems.

MCP: A New Protocol for a New Consumer

This is where Model Context Protocol (MCP) enters the picture. MCP is an open protocol designed for LLM-based agents, a fundamentally different data consumer than what OPC UA, MQTT, or REST APIs were built to serve.

Therefore, MCP does not replace existing protocols. OPC UA, SQL, and MQTT still have their jobs, but MCP is the protocol that aggregates and contextualizes data from these sources and exposes it for agent discovery and use.

Importantly, there’s a practical constraint here that many organizations are only now learning: Agents work best when they have a small, focused set of MCP tools (perhaps 5 to 10) rather than hundreds. When agents are overloaded with tool options, they make poor decisions about which tool to use, leading to the hallucinations and errors that undermine trust.

This means that another layer of governance needs to be planned for, where agents are authorized for specific tasks to ensure they are safe and effective. The number of tools they have access to is limited, but the quantity of agents needed goes up exponentially.

The Bidirectional Opportunity

So how do manufacturers solve the 100x problem? The symbiotic relationship between AI and data operations (DataOps) offers a compelling solution.

• DataOps for AI provides the foundation agents need: Curated, contextualized, and governed data delivered through managed tools and pipelines.

• AI for DataOps accelerates data operations deployments by leveraging AI to speed configuration and management of the DataOps solution.

This bidirectional relationship is where the real compounding value emerges, where every investment in one accelerates the return on the other.

Are you researching or planning to invest in a DataOps approach this year? Then I recommend refining your search to solutions purpose-built for industrial data. HighByte Intelligence Hub is a DataOps software solution designed specifically for industrial data modeling, orchestration, and governance. It provides the data infrastructure for industrial AI, including robust support for MCP services. Take advantage of the free trial and download the software today to get started.

Wrap Up

If your organization is evaluating AI for manufacturing, the data infrastructure conversation needs to happen now. Keep these three action items in mind:

1. Treat data as a first-class citizen by investing in data strategy, building standardized data products, and adopting a governance model independent of the consuming applications and agents.
2. Adopt a hub-and-spoke data architecture that connects to required systems and delivers contextualized, productized data to any consumer. Contextualize data at the Edge close to where it originates to reduce latency, lower cloud costs, and ensure that every downstream consumer receives data that’s ready to use.
3. Plan for MCP alongside existing protocols and establish governance before you scale. Define data access policies, curate agent tooling, and implement a data control plane from day one.

The 100x data consumer explosion is already beginning. The organizations that invest in data infrastructure now will lead. Those who wait will find it increasingly difficult to scale AI beyond initial pilots.