Agentic AI · solution
Scheduling agents for manufacturing
A scheduling agent decides which job goes on which machine next, using order priority, machine health and material availability against a schedule policy your operations team owns. It writes a next-slot recommendation into the MES that dispatchers can accept or override in one click.
The decision this agent takes
Every N minutes (typically 5 to 15), the agent recomputes the next-slot recommendation for each eligible machine. The decision uses order due-dates, current machine state, upstream WIP, material availability and the current shift-plan. It writes a ranked suggestion into the MES; dispatchers accept, tweak or override.
Data the agent needs
- Order book with due dates and priorities from the ERP.
- Live machine state (running, idle, breakdown, changeover) from the IIoT layer.
- WIP position and material availability from the MES.
- Setup / changeover matrix per machine, owned by industrial engineering.
- Shift calendar and operator-skill matrix so the schedule is executable.
What it does unsupervised
- Publish next-slot recommendations to the MES scheduler.
- Re-optimise when a machine goes into breakdown, when a hot order arrives, or when material availability changes.
- Notify the shift supervisor when the recommended schedule violates a policy constraint (e.g. overtime cap).
- Log every decision with its inputs so overrides can be studied against outcomes.
Where a human stays in the loop
- Dispatchers can override every recommendation. The agent tracks override rate as a health metric on itself.
- Priority changes to the order book are human-driven; the agent respects them, it does not set them.
- Overtime and shift extensions are approved by the shift supervisor.
- Changeover-matrix updates go through industrial engineering.
Where this agent gets it wrong
Honest failure modes and the data-quality prerequisites we insist on before an agent goes live. If your plant does not meet these, the agent runs in advisory-only mode until it does.
- If the setup / changeover matrix is not maintained, the agent will pick sequences that look optimal on paper but are painful on the floor. Data-quality prereq: changeover matrix reviewed quarterly.
- Operator-skill mismatches are silent unless the skill matrix is loaded; recommended jobs may end up assigned to operators who cannot run them.
- When multiple orders share a bottleneck material, the agent optimises within its objective function; conflicting priorities across product families still need a planner.
- Very high dispatcher-override rate (>20%) is the honest signal that either policy or data is wrong, not that the agent needs more tuning. We escalate rather than compensate.
How Ajinkya ships this today
Ajinkya’s MES service line at reference plants already carries scheduler data and order-book integration; the agentic decision layer on top is scoped per plant. A productionised out-of-the-box scheduling agent is [NOT YET SHIPPED].
FAQ
How does this compare with APS software?
APS is the tool the schedule runs in. The agent decides what the schedule should be given live floor state. Many customers keep their APS and put the agent alongside it.
What is the ideal shop-floor complexity?
Best fit: 8 to 40 machines with meaningful sequence-dependent changeovers. Below that, a rules engine is enough. Above that, the agent still helps but tuning takes longer.
Does the agent decide overtime?
It flags the need. Overtime is a human decision because it is a HR/policy decision.
How is it tested before go-live?
Two weeks of shadow mode where the agent recommends but does not publish. Recommendations are diffed against dispatcher choices and gaps investigated before switching on.
Talk to the engineers who ship this
A 30-minute discovery call gets you a scoped answer on whether this agent fits your plant, what data your team would need to prepare, and what a pilot would look like.
Book a 30-min callReviewed by Amey Kadle, Founder, Ajinkya Technologies. Last reviewed: 2026-08-29.