← All episodesAutomated Gemba / A Monday letter

Episode #09

The number
looks wrong.

AI agents can read every system in the plant. A batch of 30 shows what they need before they change anything.

Illustration of two operators in welding helmets brazing the corners of a stack of galvanized panels on a steel table. Labels read: We used to run 60. ERP / Batch size: 30.
Illustration

Stainless ran one-piece flow. Right next to it, galvanized ran in batches of 30.

Same 9,000 square feet. Same press brake. Both products went through the station where we glued the metal skins to the honeycomb core. If you’d seen the two batch sizes on a production report, you might have wondered why we’d stopped at 30 when one-piece flow was working a few feet away.

We’d already cut the galvanized batch in half. To explain why we kept the other half, I’d have taken you to the brazing station.

Why we kept the batch

At Flush Metal Partition, the crew laid the galvanized panels flat, turning every other one slightly so the corners stayed exposed. Then they brazed those corners from the top of the stack down. By the time the operator started brazing, the stack itself was part of the setup. Asking for one finished panel at a time would have meant asking the crew to develop a different way to do the work.

The old batch was 60. Some operators had worked that way their whole careers. At 30, they could finish a smaller stack using the sequence they knew. We ran 2 batches of 30 and kept the same weekly output.

The exposed corners explain why the crew needed a stack. We settled on 30 because it cut the batch in half and the people doing the brazing could make it work.

A few feet upstream, the press brake operator had a different decision to make. They might bend 3 galvanized panels, switch to 1 stainless panel, then look downstream to see which station needed work. A stainless operator waiting for the next panel was a good reason to switch.

Now imagine an AI agent reading “batch size 30” from the planning system and using it to schedule the press brake. It could tell the operator to finish all 30 galvanized panels before switching to stainless. Every quantity on that schedule could be accurate, and the stainless operator would still be standing there waiting.

What I’d want the agent to know

Picture the plant’s software as a house. Product lifecycle management (PLM) is the foundation, and the rooms hold different parts of the story: drawings, plans, production records, quality checks, and maintenance history.

If I opened the batch setting in ERP a year later, I’d want to see what we changed, from 60 to 30, and which operation that decision applied to. A photo linked to the galvanized brazing operation could show the stack with its turned panels. A short note from the crew could explain how they worked down it and kept the same weekly output after the change. If maintenance later recorded a change to the setup, the person reviewing the batch size would know what to check on the floor.

I’d want a separate connection for what happened before paint. The product drawing could show the bonded honeycomb core. The production record could tell me when brazing ended. After brazing, that core could lift away from the skin. The adhesive needed to cure overnight, and any panel that needed re-gluing had to be dealt with before paint. A quality record linked to that panel could explain what the crew found the next morning.

Those connections would lead an agent to ask two different questions. At the brake, does the brazing batch determine when the operator switches material? Before paint, how much of the recorded wait does the adhesive need? I’d want the agent to take an unanswered question to the people doing the work before it proposed a change.

The experiment I’d run today

I’d give an agent the routing, batch quantity, and production times for one galvanized panel family. I’d ask it to recommend one change, name the operation it would affect, and tell me what someone needed to verify on the floor. Then I’d repeat the exercise with the crew’s photo, their explanation of the batch, and the overnight cure requirement.

Getting the explanation down is practical. At a signage plant, I used an SOP generator to turn my description of how a sign got damaged into a draft procedure in about a minute. I’d try the same approach here. Let the brazing crew explain the setup in their own words, have the agent draft the note, and give it back to the crew to correct.

I’d take both recommendations to the brake operator and the brazing crew. Could they use either one on the next run? If the agent proposed a smaller batch, I’d want to hear how it expected the operator to braze it. If it proposed an earlier release to paint, I’d want to know where it had accounted for the cure.

There’s one wait I’d measure on that next run. When was the stack complete, and when did brazing begin? If a finished stack sat waiting, we could ask the crew what was happening during that interval. That’s a delay we could investigate on the same shift, with the people who saw it.

Which number in your plant is right for a reason only the floor knows?