Generative AI agents, on their own, do not have a robust understanding of cause-and-effect for real-world decision-making.

However, when combined with Deep Reinforcement Learning, AI agents gain the ability to reason, learn from interaction, and make decisions that solve operational problems in complex, real-world environments, like the plant floor.

Case in point.

Bryan DeBois and his team at RoviSys developed an Autonomous AI agent to manage a notoriously difficult glass bottle production process, where small disruptions like temperature fluctuations can quickly push the process out of specification.

Here’s how they approached it:

✅ 𝐒𝐭𝐞𝐩 1 - 𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐓𝐞𝐚𝐜𝐡𝐢𝐧𝐠

They captured the knowledge and decision-making strategies of expert human operators and used this to train the AI agent, essentially teaching it how to respond to different operating conditions.

✅ 𝐒𝐭𝐞𝐩 2 - 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧 𝐒𝐮𝐩𝐩𝐨𝐫𝐭 𝐌𝐨𝐝𝐞

Initially, the agent didn’t control the process directly. It simply made recommendations.

Operators reviewed the suggestions and gave feedback using a simple green/red button system. This built trust and allowed the team to validate the AI’s decisions without risk.

✅ 𝐒𝐭𝐞𝐩 3 - 𝐂𝐥𝐨𝐬𝐞𝐝 𝐋𝐨𝐨𝐩 𝐂𝐨𝐧𝐭𝐫𝐨𝐥

Only after months of successful operation in support mode did they enable full automation.

Even then, strict safety measures were in place:

⇨ Limited control authority

⇨ Clearly defined operating boundaries

⇨ Automatic handover to human operators if conditions exceeded the agent’s training

The Results:

⇨ Human operators typically needed 7–20 minutes to bring the process back into spec

⇨ The AI agent consistently did it in under 5 minutes

⇨ And it maintained safety by operating strictly within validated limits

In the latest episode of the AI in Manufacturing podcast, I sat down with Bryan, Director of Industrial AI at RoviSys, to dive deeper into how manufacturers can leverage AI and autonomous agents to optimize manufacturing operations and improve efficiency