AI in Production Planning & Scheduling: Between aspiration and reality

Anyone talking about AI in production planning today quickly arrives at bold promises: precise forecasts, self-optimizing processes, automated decision-making. It sounds coherent, but in day-to-day operations, it often does not perform as expected. This is not due to inadequate algorithms. Rather, the limiting factor lies in the underlying conditions in which many approaches fail. This is exactly where a closer look is worthwhile: what does it actually take for AI to be meaningfully applied in production planning?

Many of these challenges stand in direct contradiction to common expectations of AI. Public discourse often paints an overly simplified picture, one that neglects key prerequisites. A classification of these widely held assumptions can be found in the article AI in manufacturing: Separating fact from fiction.

The true foundation for AI in Production Planning & Scheduling: Data that withstands reality

In theory, the starting point is clear. Planning systems access existing data, analyze it, and derive decisions from it. In practice, the picture is significantly more complex.

The data basis in production planning consists of a wide range of sources. There are classical master data such as routing plans, bills of materials, setup times. Additionally, there is feedback from production itself: actual processing times, disruptions, scrap rates. Finally, there is contextual information that is often only indirectly captured such as qualifications, seasonal influences, or logistical constraints.

The problem rarely emerges abruptly. It builds up gradually. An operation is consistently evaluated too optimistically, a machine “provides” higher availability in the system than it does in reality. At this point, a simple rule becomes decisive, one that is particularly evident in production planning: garbage in, garbage out.

What may sound trivial has far-reaching consequences here. Planning systems, whether traditional or AI-driven, always operate based on available data. If this data is incomplete or biased, the results are not merely “inaccurate,” but systematically incorrect.

A misjudged process step triggers further incorrect decisions. An unrecorded disruption affects delivery reliability, capacity planning, and prioritization. And when these effects persist over multiple planning cycles, a distorted picture of the entire production system emerges.

This is the critical point: AI does not automatically correct such errors. It identifies patterns – but only those present in the data. If those patterns are flawed, they are reinforced rather than exposed.

The blind spot: What happens between planning and reality

A second aspect is often underestimated. Many companies have sophisticated production planning, but little transparency regarding how far it deviates from reality.

Plans exist; results exist. What is missing is the clean link between them.

Only through versioned historization of planned and actual data do patterns such as recurring delays, typical bottlenecks or systematic deviations emerge.

Planning thus becomes measurable and therefore finally capable of learning.

The critical question: What happens with this data?

Up to this point, much remains understandable. Data must be accurate, and planning must be reconciled with reality. The real challenge begins afterward: what do companies do with this data? This is where data analytics becomes the central instance – not as a dashboard, not as a report, but as the link between data and decisions.

There are five stages of data analytics that can be distinguished:

• Descriptive: What happened? KPIs on availability, performance, or delivery reliability create transparency.
• Diagnostic: Why did it happen? Patterns and root causes behind deviations become visible.
• Predictive: What will happen? Forecasts regarding capacities, risks, or lead times become possible.
• Prescriptive: What should be done? Systems derive concrete measures or alternatives.
• Cognitive: What do we not yet know? Systems identify blind spots or missing data.

With each stage, the benefit increases – but so do the requirements regarding data, structure, and integration.

The real challenges behind AI-driven Production Planning & Scheduling

The theory is clear. Practice looks different.

Many companies operate solidly at the first two stages: they know what has happened—and often why. Transparency emerges, and causes can at least partially be explained.

Beyond this point, it becomes difficult.

The next step – forecasting or even concrete recommendations – is not a technical upgrade but a structural shift. Data must be consistent and context-rich, patterns must remain stable over time, and results must not exist in isolation but must be integrated into production planning.

This sounds obvious but is challenging in implementation. In reality, analytics often remains superficial. Evaluations, reports, dashboards are created, but they do not influence production planning. The result: the past is better understood, but the future is rarely changed.

Why AI in Production Planning & Scheduling often fails at this point

This is where it becomes clear why many AI initiatives do not succeed. If analytics is not established as a central instance, there is no consistent data foundation, patterns remain fragmented, and no reliable decision logic emerges.

AI then becomes an experiment – not a component of the production planning process. Only when analytics assumes this central role does something emerge that AI can use meaningfully: a consistent, data-driven planning process

Where AI delivers value today – and where it does not

Once this foundation is in place, AI becomes more concrete and above all, tangible.

It provides value where data reveals patterns:

• Operation durations can be estimated more realistically.
• Risks become visible earlier.
• Planning becomes more stable and less susceptible to disruptions.

These are not spectacular effects – but they have a direct impact on daily operations. This is where expectations diverge from reality. AI supports production planning; it does not replace it.

The commonly discussed visions – fully autonomous planning or self-optimizing systems – currently have little to do with operational reality. They require a level of data depth, integration, and stability that is not yet present in most environments.

A Look ahead: Agents and networked planning

Things become interesting when one thinks one step further. In research and development, agent-based approaches are increasingly emerging – systems that no longer plan centrally but distribute decisions.

At their core, two directions can be distinguished.

One approach follows classical logic: a central instance creates an overall plan with a global perspective. It considers all constraints, prioritizes orders, and ensures system stability. The advantage is clear: centralized control, high controllability, and transparent results.

In contrast stands a decentralized model. Here, individual resources – machines, workstations, or entire areas – act as autonomous units. They identify local bottlenecks, respond to changes, and coordinate with each other. Planning is no longer created in a single location but distributed across the system. The appeal lies in its dynamics. Disruptions, rush orders, or short-term changes could be handled much faster because decisions are made closer to the actual events.

At the same time, the challenge shifts. The more decentralized decisions become, the more critical the quality of the underlying information, and thus, once again, data analytics. Only if all participating “agents” operate on a consistent data foundation and correctly interpret deviations can stable results emerge.

This is where the connection to reality becomes evident: agent-based models require functioning data, history, and analytics. Without this foundation, complexity is merely shifted – not resolved.

Therefore, such approaches are currently primarily forward-looking. They indicate where planning may evolve, but the prerequisite is established much earlier: in the disciplined handling of data and in an analytics logic that makes decision-making robust in the first place.

Practice instead of buzzwording: Why implementation experience matters

Especially in the context of AI in production planning & scheduling, the gap between concept and reality quickly becomes apparent. Many approaches appear coherent, as long as they remain abstract.

In real production environments, conditions change:

• Data is never perfect.
• Processes are constantly evolving.
• Systems must be integrated.

This is why progress currently occurs where research and practice intersect.

DUALIS has been engaged in this field for years through corresponding research projects – with a focus on more realistic planning, improved forecasts, and the integration of these approaches into existing planning systems.

A target vision that intentionally remains grounded

Ultimately, this is not about an autopilot. It is about a planning approach that continuously learns from reality, systematically evaluates deviations, and derives forecasts as well as robust courses of action – not as a replacement, but as a basis for decision-making. The planner remains at the center, evaluating these options in the context of their specific production environment.

This is the key point: AI in production planning rarely fails because of algorithms. The decisive challenges lie upstream – in data quality, in the consistent linkage between planning and reality, and above all, in the role of data analytics as the connecting instance. Only when these elements interact does a system emerge that truly learns. And that is exactly where the boundary lies between a convincing concept and a solution that performs in practice.