Agentic AI will bring thousands, potentially tens of thousands, of AI agents to the factory floor. Every one of them will need data, and not all data. Agents do not perform well when exposed to massive amounts. They need focused, usable data scoped to their task. So, how much should manufacturers let them do without a human checking? HighByte has been working through this with customers at Bayer and Alcon, and the answer, for now, is that AI assists and humans decide.
Why Does Factory AI Need a Human in the Loop Today?
The gain from manufacturing AI today is speed: closing the loop between AI suggestion and operator decision faster. The operator still decides.
Ask an LLM the same question a thousand times, and at least one of those times the answer will be different, and most likely wrong. In pharma, where regulations are among the strictest in manufacturing, that matters. HighByte Intelligence Hub version 4.4 addresses this with a three-mode slider for its Pipeline AI agent: fully manual, AI-assisted, and fully autonomous. The mode shipping in production today is AI-assisted: AI proposes changes covering roughly 80 percent of the configuration work, and the operator reviews each one before accepting or rejecting it, similar to source control in software development. A fully autonomous mode exists, but as Aron Semle, HighByte Chief Technology Officer, put it, “I don’t think we’re totally ready for it.”
Unless something fundamentally changes in how these models work, even a couple of years from now, manufacturing will still be in this mode: AI as assistant, human making the ultimate decision.
How Do You Control What Each AI Agent Can Do?
If humans stay in the loop, the next question is how to control what each agent has access to. You do not want to give an AI agent open access to your entire historian and hope for the best.
The answer is scoped tool access. You build custom tools that define exactly which data an agent can request and which parameters it can use. Everything outside that scope stays invisible to the agent. A quality agent and a predictive maintenance agent can run on the same server but see entirely different data. Every manufacturer uses the same underlying LLMs, so the differentiator is the context fed to those models, not the models themselves.
HighByte adopted MCP (Model Context Protocol) early after Anthropic released the standard in late 2024, building its Intelligence Hub as a layer that sits between the AI agents and the industrial data sources. Instead of connecting directly, an agent queries the MCP server, discovers what tools are available, checks authorization, and receives structured data back. The scope of each tool determines how much the agent can do on its own.
What Does Controlled AI Look Like Inside Bayer’s Pharma Plant?
Bayer, a global pharmaceutical and life sciences company, needed AI inside a GXP-certified environment where you control exactly what the LLM can do. The team connected to a PI system that had been running for 15 years, and built custom tools scoped to specific PI event frames and historical queries.
Each tool defined which data the agent could request and which parameters it could use. Everything outside that scope stayed invisible to the agent. The engineer who led the integration had installed the original PI system and knew its data architecture, which shaped how tightly each tool was scoped.
Production reports that previously required physically entering the clean room now come through the AI agent on demand. Early answers improved as the underlying LLMs got better over the past year, and the team is expanding the same approach to other parts of the business. Because the tool layer sits between the agent and the data source, a factory running PI and another running IP21 can present the same interface to the same AI agent. Same questions, same tools, same results, even though the underlying pipelines are built differently at each site.
Does This Controlled Approach Scale?
Alcon, a global eye care company, wanted to deploy predictive maintenance and quality monitoring across its contact lens manufacturing to improve yield and quality. Using HighByte Intelligence Hub, Alcon deployed not on a single machine but across the entire plant, very efficiently, with very low effort. Using traditional methods, the engineering team estimated it would take a year. With HighByte, it took less than a month.
Frequently Asked Questions
1. How is AI being used in manufacturing operations today?
AI agents in manufacturing assist operators by monitoring production variables, analyzing historical data, and suggesting changes for human review. At Bayer, an AI agent, connected to a PI historian via custom MCP tools, pulls production reports on demand from a GXP-certified facility. At Alcon, a data contextualization layer supports predictive maintenance and quality monitoring across an entire contact lens plant, cutting a one-year deployment to under one month. Both use cases keep the operator in the decision-making loop.
2. Should manufacturers use fully autonomous AI in production?
Most manufacturers should not run fully autonomous AI in production today. The recommended approach is AI-assisted mode, where AI proposes changes and operators review each one before accepting or rejecting it. A fully autonomous mode exists but is not considered ready for production. Unless something fundamentally changes in how these models work, manufacturing will remain in AI-as-assistant mode for at least the next couple of years.
3. How do manufacturers control what AI agents can access on the factory floor?
By building custom tool interfaces that define exactly which data each agent can request. Each tool carries a description, parameters, and authorization rules. Agents see only what they are given access to. A factory running one historian and another running a different one can present the same tool interface to the same AI agent, keeping results consistent across sites.
This article is based on video interviews with Aron Semle, Chief Technology Officer at HighByte, and John Harrington, Co-Founder and Chief Product Officer at HighByte, and Lucian Fogoros of IIoT World, recorded at Hannover Messe 2026. AI tools were used to help summarize and organize the content. Reviewed and edited by the IIoT World editorial team.
Editorially independent. Sponsored by HighByte.
About the author
