AI agent for production planning and scheduling

How it works

1. Gather the day's facts

   The agent pulls open orders and due dates, the bill of materials, routings with standard times, machine calendars and material availability from your ERP or spreadsheets.

2. Build a schedule

   A scheduling method that respects capacity, due dates, material and changeover times arranges the orders. The language model does not make the schedule. It explains it in plain words.

3. Planner reviews

   The planner sees the proposed sequence, which orders will be late and which constraint causes each delay. They approve it or move things around, and the agent shows the effect.

4. Publish and adjust

   The approved sequence goes into the ERP as work orders. When a machine goes down or a rush order arrives, the agent re-plans only the affected area and shows the planner what changed.

New to AI agents? How they work, in plain words

Real examples, setup steps and tool connections for this agent

An example conversation

An example of how a conversation runs. The grey lines show what the agent does in your systems meanwhile.

1. Customer(planner, Teams) Morning. What does the week look like?
2. Read 63 open production orders, machine calendars for 9 work centres, and material stock from the ERP. Built a draft schedule for 5 days.
3. AgentDraft is ready. 59 of 63 orders are on time. Four will be late. Three of them wait for aluminium sheet that arrives Wednesday. The fourth, order 3318, is late because press 2 is full on Tuesday and Wednesday.
4. Customer3318 is for our biggest customer. Can it move earlier?
5. Re-ran the schedule with order 3318 first on press 2.
6. AgentYes, if 3318 goes first on press 2, it finishes Tuesday. Two other orders move back one day and stay inside their due dates. Changeover time rises by 40 minutes. Do you want that version?
7. CustomerYes. And press 4 needs maintenance Thursday morning, is that included?
8. Checked the maintenance calendar: press 4 is blocked Thursday 06:00 to 12:00. The draft already avoids it.
9. AgentYes, press 4 is blocked in the plan. Waiting for your approval before I publish work orders to the ERP and the line leaders.

How it is built

Where the data comes from

Orders, bills of materials, routings, machine calendars and the changeover matrix are read from your ERP or MRP through its API, or from spreadsheets, at least once a day. Machine status can come from your MES or through OPC UA, the standard protocol for shop floor machines, as read-only information.

A scheduler makes the plan

The sequence is built by a scheduling method (a constraint solver or a rule-based heuristic) that treats capacity, due dates, material and changeovers as hard rules. The same data gives the same plan, which makes it checkable. The language model only explains why an order is late and what a change would cost.

Checking your standard times

A plan is only as good as the times behind it. The agent compares routing times with recent actual times and flags routings with large gaps, so a plan built on wrong standards does not look more certain than it is.

The planner has the final say

Nothing is published until the planner approves. The planner can edit any step, and the agent shows the effect on due dates and changeovers straight away.

Firm orders first

Confirmed orders are planned first. Work driven by a forecast is kept as a separate, smaller block, so a forecast error does not spread through the whole week.

Re-planning stays local

When something changes, the agent re-plans only the affected area and keeps the next shift's sequence frozen, so the floor is not surprised by a full re-plan every hour.

Logging, and cloud or your own servers

Each plan, edit and approval is logged with who made it, and the agent reports on-time delivery, utilization and changeover time weekly. It can run in the cloud, or on your own servers, for example on the factory network. Then the agent, its database, its logs and the language model itself stay with you, using an open model such as Llama, Qwen or Mistral, so production data stays inside. Planning from your ERP and machine data needs no internet connection, while email and messaging still pass through those services. Local open models are smaller than the largest cloud models. For explaining a schedule and answering from your data they are usually enough, and we test on your own cases first. The scheduler is not a language model, so it does not depend on model size. The agent does not send commands to machines.

What you can connect to it

SAP S/4HANA (PP)

Reads production orders, work centres, routings and bills of materials, and writes approved sequences back through OData services.

Odoo Manufacturing

Reads manufacturing orders, bills of materials and work centres, and writes the approved order through the external API.

Microsoft Dynamics 365 Business Central

Reads items and production orders and updates the scheduled dates through its API.

Oracle Fusion Cloud

Reads manufacturing work orders and resources through its REST APIs.

MES or machine data via OPC UA

Reads machine status and counters as information. It does not control equipment.

Maintenance log or CMMS

Reads planned maintenance, so downtime is in the plan before the day arrives.

Excel or Google Sheets

For plants that plan in spreadsheets: reads orders and capacity, and shows the planner the schedule in a sheet.

Email, Teams or WhatsApp

Sends the approved schedule and changes to the line leaders and receives breakdown or rush-order reports.

Purchasing and stock data

Reads material availability and expected deliveries, so orders are not scheduled before the material is there.

Related agents

• AI agent for inventory management
• AI agent for procurement and purchase orders
• AI agent for data entry