AI Agents in Manufacturing: What a Good Answer Actually Looks Like

Furnace 2 is down until 6:00 tomorrow morning. I asked an AI assistant a simple question: which orders are at risk? The answer came back in two seconds, well written and confident. It was also wrong.

AI agents in manufacturing are software assistants, usually built on a large language model (LLM), that can answer questions and take steps by calling other systems: reading orders, querying stock, running a planning calculation. The word that matters is calling. An AI that answers from general knowledge sounds smart. An AI that queries your MRP and your schedule can actually be right.

The situation

It's Tuesday, 6:00. Furnace 2 needs a repair and will be back Wednesday at 6:00. Three manufacturing orders were scheduled on it today:

• MO-4471, customer Alpha: 10 hours in the furnace, then 6 hours of finishing. Ships Wednesday 18:00.
• MO-4475, customer Beta: 8 hours in the furnace. Ships Friday.
• MO-4480, customer Gamma: 6 hours in the furnace. Ships Thursday 12:00.

Furnace 1 has a free window today from 14:00 to 21:00, which is 7 hours. It is approved for the MO-4480 and MO-4471 products, but not for MO-4475.

The bad answer: fluent and invented

"Furnace downtime may affect orders due this week. MO-4471 and MO-4475 appear to be at risk. I recommend moving both to Furnace 1 and informing customers of possible delays."

It reads well. Now check it. MO-4471 needs 10 hours, and Furnace 1 has 7. MO-4475 isn't approved for Furnace 1 at all. And the answer missed MO-4480, the one order that can actually be saved by moving. The model produced something that sounds like a planner. It didn't look at the schedule, the routing or the calendar.

The good answer: sourced and calculated

"Checked against the current schedule and the Furnace 1 calendar: - MO-4480 (Gamma): move to Furnace 1, today 14:00–20:00. Ships Thursday 12:00 as planned. - MO-4471 (Alpha): too long for the 7-hour window. First slot on Furnace 2 is Wednesday 06:00–16:00, finishing 16:00–22:00. That's 4 hours after the 18:00 ship time. At risk. Options: overtime in finishing, or ask Alpha to accept Thursday morning. - MO-4475 (Beta): Furnace 2 only. Wednesday 16:00–24:00. Ships Friday, no risk."

Every line names an order, a resource and a time. The arithmetic can be checked: 06:00 + 10 hours = 16:00, plus 6 hours of finishing = 22:00, which is 4 hours past 18:00. And the answer gives the planner a decision, not a lecture.

The good answer isn't smarter writing. It exists because the AI queried real planning data: the production schedule, the routing approvals, the resource calendar and the ship dates. Then it let the scheduling calculation do the math instead of guessing.

How to tell the two apart

Think of two new hires. One speaks confidently in every meeting and has never opened the ERP. The other says less, but comes back with a printout. You trust the second one. Use the same test on AI:

• Does it name specific orders, resources and times? Vague answers ("some orders may be affected") are a warning sign.
• Can you trace each claim to a record? Order number, calendar slot, routing approval.
• Does it respect constraints? Capacity, approved resources and material availability, not just "move it somewhere".
• Does it say what it doesn't know? "Repair time is an estimate" is a good sign.

Myth: "AI will replace the planner"

Look at the good answer again. It stops at a choice: pay for overtime or call Alpha. That choice depends on things no system holds: how important Alpha is this quarter, whether the finishing crew already worked last weekend, what the sales rep promised on the phone. The AI saved the planner 30 minutes of digging. The planner still made the call.

What AI actually replaces is the searching: opening five screens, exporting to Excel, cross-checking dates. That's valuable. It just isn't planning.

What AI agents in manufacturing need from planning data

An AI assistant is only as good as what it can query. If your ERP holds fixed 10-day lead times and no machine calendar, the AI can't tell you what Furnace 2 does to Wednesday. That's the gap between ERP and real planning. The data an AI needs is:

• A finite-capacity schedule with real start and end times per operation
• MRP results that link each shortage to the order it hurts
• Resource calendars and approved alternatives
• Current confirmations from the shop floor

Get those right and the AI has something true to read. It also makes delivery performance measurable, since OTIF depends on exactly these dates.

How this looks in factory.online

factory.online doesn't offer an autonomous AI agent that reschedules the plant. What it holds is the data a good answer depends on: finite-capacity schedules on a Gantt, MRP with shortages linked to orders, resource calendars, and output confirmations. In the furnace case, a planner sees MO-4471 turn late the moment Furnace 2 is blocked, and can test the Furnace 1 move in a planning scenario before touching the live plan. That kind of data is what any AI assistant would need to query. For the bigger picture of how planning tools evolve, see manufacturing planning software.

For optimization tasks beyond answering a question, kafaa.ai focuses on jobshop schedules, product mix and cutting scenarios. A useful evaluation checks the proposed result against actual constraints and the business objective before the team acts on it.

To see your own late orders in real schedule data, book a demo.