AI and production planning: real opportunities and pitfalls to avoid

Artificial intelligence has worked its way into every conversation about production planning. The promises are tempting, the projects are multiplying — and so are the disappointments.

A client reached out to us some time ago with an ambitious project: to develop an in-house AI solution to automate their planning. The technical team was solid. The commitment, real. The starting idea seemed logical: capture what the planner did manually, then automate it.

That's exactly what they did. They automated the planner's limitations.

By faithfully reproducing the existing process, they encoded into their system all the constraints a human imposes on themselves because they can't hold more than a dozen variables in their head at once. The result wasn't optimized planning. It was manual planning, only faster, with all its flaws intact.

In the end, they realized they were recreating optimization algorithms that had already existed for decades in proven APS solutions.

The real risk of an AI project in production planning

There's a lot of talk about technical risk in AI projects. In production planning, the real danger is elsewhere. It isn't the absence of technical skill — it's the absence of an understanding of what planning is.

A developer who can't tell the difference between a work center and a machine center can write perfectly functional code. They simply won't be able to correctly model a production shop. And an AI model built on a poor representation of the constraints will produce plans no one can execute on the shop floor.

Production planning is a discipline. It has its concepts, its rules, its subtleties. It isn't a generic optimization problem you solve with a good algorithm and good intentions.

The 4 concrete roles of AI in planning

That said, AI does have real applications in planning. Four, mainly.

Forecasting: predicting demand

Machine learning models excel at analyzing historical data to produce demand forecasts more accurate than traditional statistical methods. Seasonality, market trends, external variables: AI detects patterns a human wouldn't see. A better forecast means less excess safety stock and fewer stockouts.

Prediction: how likely is your plan to actually hold?

Predictive AI goes beyond a plan: it assesses the likelihood that the plan will materialize. Which order risks being late given historical breakdowns? Which supplier poses the most risk this month? You move from a static plan to a plan with an associated confidence level — and the planner can focus their interventions where the risk is real.

The AI agent: talking to your data

Conversational agents let planners query their data in natural language. "Which orders risk being late this week?" "What's the impact if I push order 4521 back by three days?" No SQL query, no pivot table. The agent doesn't replace the planner — it gives them immediate access to the information.

Metaheuristics: simulating thousands of plans

Planning a plant with dozens of resources, hundreds of orders, and thousands of constraints is a problem mathematically unsolvable by brute force. Metaheuristic algorithms — genetic, simulated annealing, particle swarms — intelligently explore the solution space to find a near-optimal plan in a few seconds. That's where AI meets operations research.

The prerequisite no one wants to hear

These four applications have one thing in common: they all need a solid data structure that faithfully represents the full set of your operational constraints. If your production times are wrong, if your machine capacities don't reflect reality, if your bills of materials are incomplete, AI will produce unusable results.

Garbage in, garbage out. The rule has never been truer than with AI.

Why program when you can configure?

The client we mentioned at the start eventually drew the conclusion themselves: they were investing considerable resources to recreate something that already existed, developed by specialized teams and refined over decades of deployments in real plants.

No one would think of developing their own ERP. So why want to develop your own APS?

Solutions like DELMIA Ortems now integrate AI natively: plan-adherence prediction, optimization through advanced algorithms, scenario simulation. This isn't AI bolted on after the fact. It's AI built around a planning model that understands the difference between a work center and a machine center.

Configuring a tool that already does all of that is infinitely faster and more reliable than programming it from scratch. And it lets your teams focus on what really matters: faithfully modeling your operational constraints. That's where your competitive advantage lies — not in the algorithms.