There is hardly a day on which the topic of artificial intelligence is not mentioned in publications. The topic is "hot" and a guarantee for good reader or click numbers. This article aims at expectation management, it is aimed at readers from controlling who want to operationalize the topic and understand what the current status is and what to think about it.
In view of the current changes in controlling, some trends can be identified. The classic controlling (reporting, top - down - planning) loses relevance and is technically well supported. Complexity and volatility of the companies and the environment are increasing. The early recognition of market trends is becoming more important and results must be commented or new ideas developed. Projects are also becoming more important and the controller has to take over the "horizontal leadership" to design different models with project managers to enable evaluation and control.
The subject area of artificial intelligence (AI) is a sub-area of computer science and includes elements of statistics and mathematics. It tries to solve existing problems with exactly these competences. To do so, it automates a (supposedly intelligent) behaviour that has been learned by machine.
A distinction must be made between the "strong AI" and the "weak AI". A strong AI - for the sake of completeness - has all the abilities of a human being and can really think independently. This may sound tempting, but in practice it simply does not exist. The weak AI finds its home in many everyday operations. It solves a concrete problem by learning a behaviour that was not modelled by hand, it simulates this behaviour and "only pretends" to think.
The best known and most relevant fields of application are image recognition, regression models and classification methods. These are the areas that a controller is most likely to encounter. In the foreseeable future, replacing the colleague at the desk opposite with a shiny chrome companion will remain science fiction.
Neural networks enable the recognition of images, words and signs. In practice, this is used, for example, to read handwritten or printed invoices and to transfer the information to a computer system (OCR). (https://candis.io)
Regression models help to forecast sales volumes in the future. They enable better production planning in less time.
Classification procedures support the risk assessment of individual transactions or customers.
Especially the regression models and the classifying procedures are certainly not new. The logistic regression or the decision tree as classifying procedures are classics. The special feature of the current version is that the models are provided with a large number of variables and a defined target variable in order to be able to learn on their own. With a training data set the algorithm can identify the relevant variables and optimize its results. This is called machine learning, a sub-discipline of artificial intelligence. The quality can then be checked using another test data set. It is important that the algorithm can only perform this one task afterwards. It is neither intelligent and can interpret, nor can it respond to special effects that have not been explained to it beforehand. Simplified it can be said that an AI is only as good as the trainer describes the basic conditions
Let us take a closer look at a machine learning algorithm for forecasting sales figures (which CoPlanner and Hadoco have implemented together) as an example close to controlling:
Before the system was ready for use, it had to be trained. For this purpose, a training data record was created, which consisted of the relevant information. Even very tolerant and intelligent models like ARIMA or the Prophet developed by Facebook are not able to choose their data sources themselves. What are possible influencing variables (weather, days of the week, holiday periods, etc.) and what is the target variable (sales volumes)? The definition has to be done by humans. So does the respective degree of aggregation - should months be predicted on days, weeks, rather on customer level, customer groups or products? For special outliers, as in this case, a generalized outlier correction can significantly increase quality. For any analysis, data quality is the be-all and end-all, and only the truly intelligent person can judge it. The credo for every data scientist: Shit in - Shit out!
The algorithm checks a variety of models and selects the best one. In this explicit case, the results were astonishingly accurate: in this specific case, the machine prediction was used as a guideline for the colleagues from regional sales who had to revise it. On balance, the automatically calculated planned turnover for the following year only deviated by 3% from the man-made turnover in the following year. The planners did not simply adopt the automatic target - in some markets there were considerable, mutually offsetting shifts. After the project has been a few months ago - even the actual course of events is close to the forecast.
These algorithms can be used everywhere for autocorelating time series and Facebook uses e.g. Prophet to predict their own key figures. So who do you turn to when you want to introduce AI? A combination of technical and methodological expertise is required. The natural contacts - at least in medium-sized businesses - are ERP and even more so the providers of business intelligence tools.
In the documents and price lists you can already find corresponding items. The general suitability of such models should be checked - they represent one or more application cases, but it is always necessary to check whether they are useful in a specific case. This clarification in itself can be as demanding as the modelling of an individual new solution. It's a pity if it turns out too late that the acquired time series analysis is not suitable because the mechanical engineering company cannot determine its turnover from quantity * price.
A standardized forecasting, which determines dependencies from the given financial key figures of the company and determines a financial forecast with an AI driver model from AI sales forecast, sounds great. However, it must be taken into account that the pattern recognition used can very well determine correlations, but not causalities. Nor can such systems per se take external influences into account, let alone disruptive events (-> Brexit, to name one example).
Nevertheless: the provider of the business intelligence solution is a good starting point for artificial intelligence in controlling, because the BI system already has all or at least a lot of the necessary data and it enables the appropriate modeling. Together with experts from the AI area (internal or external), it is possible to find out which is the best approach and also what effort can be expected. For some use cases, web services are available that make AI models usable without software procurement.
Human input will continue to be necessary for AI systems to create value. In controlling, they will facilitate many routine tasks, in some cases taking over completely, and creating freedom for the actual task: recognizing trends, moderating, modeling and simulating measures, and being a coach for management and colleagues from the line. In the end, AI is nothing more than another tool in the management cycle for corporate performance management. Of course you can carry your goods to the customer with many strong men, but would anyone think of that today?