AI Demand Forecasting: Smarter Planning For Companies
Interview with Dr. Sven Flake
“AI forecasting does not realise its value in
the dashboard, but rather where it enables better day-to-day planning decisions.”
Six practical questions for Sven Flake
Ask almost anyone about AI in demand forecasting, and the conversation quickly boils down to a single question: Will AI replace forecasting? That is the wrong question. The more useful one and the one that practitioners are actually grappling with is: How do we make forecasts more adaptable, more granular and more relevant to the decisions they are actually meant to support?
After all, a forecast is not an end in itself. It exists to help someone decide how much to produce, when to reorder, and what to promise a customer. And this is precisely where the real tension lies: it’s easy to get excited about better algorithms, but they only create added value when they are integrated into the way a company actually plans.
In this ‘Ask an Expert’ conversation with Dr Sven Flake, we look beyond the hype and delve into practical reality. When is machine learning truly superior to traditional methods, and when are the traditional methods still good enough? Which data actually improves a forecast, and which merely generates noise? Where do AI forecasts deliver measurable business value, when should a planner override the model, and what mistakes do companies most commonly make on the journey from Excel to AI? The answers revolve less around technology than most people expect, and more around data, processes and trust.
Where does AI actually outperform traditional forecasting methods and where are classical approaches still good enough?
In many companies, demand forecasting still starts with a very pragmatic setup: historical sales data is exported into Excel, planners compare it with previous periods, adjust for seasonality, add market knowledge and manually update the forecast.
That is not necessarily wrong. In fact, for many stable products and markets, this approach can work reasonably well. The challenge is that it often depends heavily on individual experience, manual effort and local knowledge. As soon as the number of products, customers, regions or planning cycles grows, it becomes harder to keep the process consistent and scalable.
It is also important to distinguish this manual planning reality from classical statistical forecasting methods. Classical methods include approaches such as moving averages, exponential smoothing, ARIMA models, regression-based models or seasonal time-series models. These methods can be very powerful, especially when demand patterns are relatively stable, historical data is reliable and the main drivers of demand are well understood.
So the difference is not that classical methods are “simple” and AI is “smart”. That would be too simplistic.
The difference is more about the type of complexity each approach handles well. Classical forecasting methods are often strong when the relationships in the data are fairly stable and can be modelled with clear assumptions. AI and machine learning approaches become useful when demand is influenced by many interacting factors, when patterns change over time, or when forecasts need to be produced automatically across many detailed combinations — for example product, customer, region, channel and location.
In practice, this can include methods such as gradient boosting models, random forests, neural networks or probabilistic time-series models. Often, several models are tested or combined in ensemble approaches to produce a more robust forecast. These methods can learn from historical sales as well as additional demand drivers such as prices, promotions, stock availability, product hierarchies, customer groups, locations or seasonality.
The advantage is not that the model “knows the future”. The advantage is that it can process many influencing factors consistently across thousands of product-location or product-customer combinations and update those forecasts much faster than a manual planning process.
Some of these factors can also be included in classical models. The difference is that machine learning is often better suited when the relationships are non-linear, difficult to predefine or different across product groups, markets and planning levels.
That said, AI should not replace classical forecasting by default. In many real projects, the best solution is a hybrid one: use robust statistical methods where they work well, apply machine learning where the additional complexity creates measurable value, and keep human expertise in the loop where business context is needed.
From a data science perspective, the key question is not: “Can we use AI for forecasting?” The better question is: “Which forecasting approach fits the data, the business process and the level of decision-making we need to support?”
Which external data sources can actually improve demand forecasts, and how do you decide whether they add value or just create noise?
Not every additional data source improves a forecast. That is one of the first things to understand. More data can help, but it can also make a model more complex without making the result better.
In demand forecasting, we usually start by looking at the data that is closest to the actual demand: historical sales, orders, product information, locations, customer groups, prices, promotions, marketing activities, stock availability and seasonality. Depending on the business, external data sources such as market indicators, events or public holidays can also be relevant.
The important part is to test whether a data source adds value in a consistent way. We first check whether the data is available at the right level of detail and early enough to support the planning process. Promotion data, for example, is only useful if it can be linked clearly to products, time periods, regions or customers. Stock availability is also important, because sales history does not always show true demand. Sometimes it only shows what could be sold because enough stock was available.
Then we compare different model versions. We use a baseline model without the additional data and compare it with models that include the new data source. Through backtesting, we simulate past forecasting situations and check whether the model would have predicted demand more accurately on data it had not seen before.
We also look at how the model uses the data. Feature importance or similar explainability methods help us understand whether a factor is really relevant or whether it only improved one test result by chance. If a data source only helps in one specific period, but not across product groups, locations or planning cycles, we need to be careful.
This is also where solutions like prognotix are valuable. Much of this model comparison, evaluation and forecast generation can be automated. Instead of manually building and maintaining different forecasting approaches in Excel, teams can use automated forecasting workflows that evaluate relevant demand drivers, generate forecasts and help planners focus on exceptions and business context.
The goal is not to include every possible data source. The goal is to identify the data that reliably improves the forecast and supports better planning decisions.
Where does AI-powered forecasting create measurable business value in practice, for example in inventory, service levels, production planning or planner productivity?
The biggest business value usually does not come from a slightly better forecast number but comes from what companies are able to do differently because of that forecast.
In our projects, we see this very clearly. AI-powered demand forecasting creates value when it helps teams reduce manual planning effort, react earlier to demand changes, improve availability, reduce waste or excess stock, and make planning processes more reliable.
A good example is our work with Zumtobel Group. Their planning environment was highly complex, with more than 10,000 raw materials, around 500 suppliers and more than 250,000 sellable product variants. In that kind of setup, forecasting is not just about predicting demand. It is about making the supply chain more flexible and helping teams make better decisions with less manual effort. By implementing an AI-based demand forecasting solution, planning processes could be automated and optimized, forecasts became more precise, and the company was able to respond more flexibly to market changes.
Another example is SPAR. In food retail, demand forecasting has a very direct impact: if you order too much, you increase waste. If you order too little, customers do not find the products they want. With an AI-based forecasting solution, SPAR was able to forecast demand for supermarket branches and articles on a weekly level, using data such as sales volumes, weather, promotions, marketing activities and seasonality. The result was a forecast accuracy of more than 90 percent, reduced food waste and significant cost savings.
For me, that is the important point: AI forecasting should not be measured only by forecast accuracy. Accuracy matters, of course. But the business value appears when the forecast is connected to real operational decisions replenishment, production planning, procurement, S&OP or customer commitments.
An AI forecast sitting in a dashboard is interesting. An AI forecast that supports day-to-day planning decisions is valuable.
When should planners override an AI-generated forecast, and when do manual adjustments make the forecast worse?
Planner overrides are not bad by default. In fact, they can be very valuable when planners know something the model does not know yet.
For example, a planner may know that a customer will place an unusual order next month, that a product will be discontinued, that a competitor has changed prices, or that a one-time event will impact demand. If this information is not yet in the data, a manual override can improve the forecast.
The problem starts when overrides are based on gut feeling only, or when they are made repeatedly without tracking whether they actually improved the result. Then the process becomes subjective, and the company loses the benefit of having a consistent forecasting model.
A good AI forecasting process should not eliminate human expertise. It should structure it. That means planners should be able to override forecasts, but the override should be documented: Why was it changed? What information was used? Did the override improve the forecast later?
This creates a learning loop. Over time, companies can see which overrides add value and which ones introduce bias. That is very important, because humans often react strongly to recent events, customer pressure or internal politics. AI can bring consistency, but humans bring context. The best setup uses both.
What do companies most often get wrong when moving from Excel-based or traditional forecasting processes to AI-powered demand forecasting?
One of the most common mistakes is treating AI forecasting mainly as a tool or technology project.
Many companies start with questions like: Which software should we use? Can we automate the forecast? Can AI make our planning more accurate? These are valid questions, but they are not the whole story. In practice, the harder questions often come earlier: Is the historical sales data reliable? Are product hierarchies consistent? Do we know where stockouts distorted the sales history? Are promotions, price changes or market effects documented in a way the model can actually use? And And the key question is [SF1] how will the forecast be used in daily planning decisions?
Excel-based forecasting often hides these issues. Planners manually correct data, adjust assumptions, copy information from different sources and add context from experience. That can work for a while, but it also means that a lot of business logic lives in individual files, comments or people’s heads. When companies move to AI-powered forecasting, these hidden dependencies become visible.
Another common mistake is expecting a perfect forecast from day one. AI forecasting should be introduced step by step: start with a clear use case, define the business outcome, compare against a meaningful baseline, involve planners early and improve from there.
This is where we support customers very closely at paiqo. We do not just implement a forecasting model and leave the rest to the business. We help clarify the use case, assess the data foundation, identify relevant internal and external data sources, connect the solution to existing planning processes and make sure the results are understandable and usable for the people working with them.
Because in the end, successful AI forecasting is not only about building a good model. It is about creating a forecasting process that the business can trust, operate and improve over time.
What is the biggest misconception business teams have about AI in demand forecasting?
The biggest misconception is that AI forecasting is mainly about prediction accuracy.
Of course, accuracy is important. But in business planning, the goal is not to win a forecasting competition. The goal is to make better decisions under uncertainty.
A forecast is only useful if people can act on it. That means it needs to be available at the right time, at the right level of detail, and in the systems where planning actually happens. It also needs to be explainable enough for business users to trust it. They do not need to understand every technical detail of the model, but they need to understand the main drivers: Why is demand expected to go up? Why is the model warning us about a drop? What changed compared to last month?
That is why I think successful AI forecasting is not just a data science topic. It is a combination of data engineering, machine learning, business understanding and process design.
Conclusion
If there is one recurring theme running through this discussion, it is this: successful AI forecasting is not a modelling problem, but a decision-making problem. The accuracy of the figure is important, but rarely where the real value lies. The added value becomes apparent when the forecast is available at the right time, with the right level of detail, and within the systems where planning takes place, and when the people using it understand and trust what it tells them.
That is why the best solutions are usually hybrid rather than dogmatic. Robust statistical methods where they work, machine learning where the added complexity pays off, and human expertise firmly embedded in the process to provide context that the data does not yet contain. That is why clean data, consistent hierarchies and honest baselines matter more than the choice of algorithm. And that is why the shift from Excel to AI is just as much about making hidden business logic visible as it is about automation.
The aim was never to predict the future perfectly. It’s about helping the company respond sooner, with greater certainty and less manual effort, and establishing a forecasting process that the organisation can trust, operate and continuously improve. That’s the difference between a forecast that sits in a dashboard and one that actually changes what a company does. At paiqo, and with solutions such as prognotix, we bridge precisely this gap.