“We implemented an APS. A year later we were back in Excel.” It’s the sentence we hear most often from manufacturers — in Europe, North America and Asia alike.
Advanced planning and scheduling isn’t new; there have been dozens of APS products for decades. Yet in many plants, APS projects are expensive, slow and end up lightly used. Meanwhile, “AI planning” and “AI agents” are everywhere. Is that just new branding, or is something fundamentally different?
We think the difference shows up in four places.
1. How the system understands the plant
A traditional APS describes the plant with fixed data structures: item master, BOMs, routings, resource calendars. Those structures decide, at design time, which constraints the system can express. Anything outside them — “these two products can’t run back-to-back without an extra clean,” “this mold can only be on one machine at a time” — needs custom development, or ends up handled by planners outside the system.
Ontology-driven AI planning starts with a model of the plant as a network of entities, relationships, rules and constraints. A new business rule is a new relationship or constraint in that network — not a schema change or a code release. The system’s understanding of the plant evolves with the business.
The most common reason an APS fails isn’t the algorithm. It’s that the system never captured the plant’s real constraints — so its plans can’t be executed on the shop floor.
2. How the plan is solved
A traditional APS leans heavily on heuristics: sort by due date, insert by priority, schedule forward or backward against finite capacity. Heuristics are fast and readable, but they only guarantee a feasible plan, not a good one — and conflicts between rules have to be resolved by someone setting priorities.
AI planning puts mathematical optimization at the core. Decisions — what to make, where, how much, when — become variables; the objective is lowest total cost or highest service; every constraint comes from the ontology. The optimizer searches for the global optimum, or a high-quality plan with a measured gap to it, and can tell you exactly what the cost-versus-service trade-off looks like under your current constraints.
Just as important, optimization can consider production, materials, inventory and transport together, instead of planning production first, then materials, then transport. At Yili, that’s how master planning went from five days across several departments to under half a day.
3. How change is handled
The hardest part of planning was never the first plan. It’s every change after it’s published: rush orders, breakdowns, late materials, demand swings.
With a traditional APS, a planner notices the change, adjusts parameters, re-runs the whole plan and manually compares old and new. It’s slow, and knock-on effects get missed.
With AI planning agents, that work is automated:
- Sense: continuously watch orders, inventory and equipment status;
- Reason: trace the impact of a change through the ontology;
- Decide: call the optimizer for an incremental re-solve and generate options;
- Act: send the recommendation and its reasoning to the planner for one-click approval — or execute within pre-agreed limits;
- Learn: record the decision to handle similar cases better next time.
The planner’s role shifts from firefighter to strategist. Agents handle routine exceptions; people step in for the decisions that matter.
4. How long it takes to get value
Traditional APS programs often start with months of requirements gathering, master-data cleanup and custom development. Gartner puts the average supply chain planning implementation at around 13 months.
An ontology-driven approach compresses that substantially. Industry ontology templates can be reused; the optimization engine is general-purpose, so algorithms aren’t rewritten for each customer; and data can start from Excel before moving to ERP and MES integrations. In our experience, teams can see a first optimized plan in as little as six weeks.
The differences at a glance
| Traditional APS | Ontology-driven AI planning with agents | |
|---|---|---|
| Business model | Fixed data structures; custom code for new rules | Ontology: entities + relationships + rules + constraints |
| Solving | Heuristics; feasible | Optimization; best feasible, with measured gap |
| Scope | Production-centric, sequential | Production, materials, inventory and transport together |
| Handling change | Human notices, full re-run | Agents detect, re-solve incrementally, recommend |
| Explainability | Readable rules, unknown quality | Binding constraints traced; “why this plan” explained |
| Time to value | Months to over a year | As little as 6 weeks to first results |
Which one do you need?
If your plant has a simple product range, few constraints and little change, a rules-based scheduling tool may be enough. But if two or more of these sound familiar, it’s worth evaluating ontology-driven AI planning seriously:
- Planners spend more than a day a week building and adjusting schedules in spreadsheets.
- Rush orders, changeovers and shortages are frequent, and each one costs half a day or more.
- You’ve implemented an APS before, but its plans weren’t executable and the team went back to manual.
- Production, materials and purchasing are planned by different people and rarely line up.
- You need to allocate orders across several plants or lines.
See how this works for production planning and scheduling, or book a demo on a scenario from your plant.
