The use of artificial intelligence can generate growth, but AI projects all too often resemble the famous hammer, the use of which suddenly turns everything into a nail: This is because it is often inadequately clarified in advance whether its use makes sense at all in this case - sometimes not even whether there is a sufficient data basis. In order to identify the AI use cases that lead to profitable use cases,
- you need to understand your own data,
- the involvement of as many areas of the company as possible and
- an iterative working method with small sprints.
It is therefore important to go through the following steps in the shortest possible iterations with the earliest possible feedback.
1. Identify data sources.
The first step is to identify the right data for the intended AI application. Data is the basis of value-creating use cases in all areas of artificial intelligence.
But companies are often unaware of what data they have. In many cases, there is not even a data directory. In this case, identifying the right data is a good opportunity to start building a comprehensive data directory.
Firstly, internal sources within the company should be used to identify data sources. It should not be expected that these internal sources can be identified along the individual departments and parts of the company. This is because data relating to a specific area is often distributed across a wide variety of systems. For example, customer data is initially the responsibility of sales. Their first port of call is the CRM system. However, there will also be other sources that are not under the control of Sales.
When identifying data sources, external sources that are beyond the company's control should also be taken into account. These include social networks such as LinkedIn or X. This is because supplementing existing customer data with customer activities on LinkedIn or X can add valuable insights to the customer database.
Another potential source of external data can also be sensor or tracking data from the Internet of Things (IoT). Such data not only provides valuable information on processes, but also predicted maintenance and the condition of maintenance-intensive parts within production.
In this first step, lean solutions such as Power BI can be used to carry out an initial preliminary analysis of the data in order to obtain an indication of the quality and completeness of the data as well as possible variances.
2 Prioritize your data.
AI applications are only worthwhile if the database contains patterns with these two characteristics:
- The patterns can be used for classifications or predictions, but
- they cannot be identified with expert knowledge alone.
In order to use the potential of your own data to create value, data with these characteristics must be identified and prioritized accordingly. This step is fundamental: identification and prioritization reduces the risk of putting unnecessary effort into data consolidation, data preparation, data enrichment and training the model. This increases the probability of success of the next steps.
3. Determine use cases.
The foundations for AI applications have been laid with the identification of relevant data sources, the creation of a data directory and the prioritization of data. Now you can turn your attention to possible use cases.
This should be done together with the experts from the respective specialist departments in the company.
This is because new ideas for use cases that you had not yet considered often emerge in discussions with the specialists. Above all, however, experts from the specialist departments know exactly what conclusions can be drawn from their data. And this is information that can be used to substantiate, supplement and, last but not least, validate your plans for AI applications.
4. Cleanse and enrich the data
Even if you now have the data that is worth analyzing automatically, the data is usually only available in a raw form. This means that the data must be cleaned up for further processing and, if necessary, enriched with certain tags. Ideally, this process starts with the smallest possible data extract. "Small" here does not mean the volume of data, but the dimensions of the attributes: The fewer attributes required for the first step, the less time-consuming it is to cleanse and enrich the various data attributes. Additional attributes can then be added successively in later steps.
This approach has two advantages:
- It minimizes the effort required for data cleansing - and therefore ultimately also the risks. However, it could still turn out later that the data does not have the expected value.
- The effect of the attributes on the result can be assessed much better using the iterative approach.
If the data is unstructured, it must be enriched with tags before further processing. Paradigmatic for this is text written for humans. Methods from the field of natural language processing (NLP) are then helpful. Open source solutions, such as SpaCy, already take on a large part of the work here by cleaning, segmenting and enriching text with tags in process pipelines.
5. Prepare the data for machine analysis.
Even cleaned and tagged data must usually still be optimized for machine analysis - for example by neural networks. Image data, for example, is not simply dumped one-to-one into an AI model. It is prepared beforehand so that the model can better recognize certain features. This addresses the challenge that objects can appear in different positions in an image. Here, the position of the object must be prevented from influencing the result.
Different approaches have been established for processing depending on the field of application - image processing, speech recognition or something else. In the case of company data in particular, this means finding out which aspects are relevant for the use case and preparing the corresponding data for pattern recognition. It may often be necessary to create a dictionary with company-specific terms and/or meanings as a first step. This dictionary can then be used to specialize a ready-made NLP model, usually based on free sources such as texts from the web.
6. Create a suitable model.
The core of an AI application is a model tailored to the respective use case. This model contains the patterns that are used for classifications or predictions. The implementation of an AI application therefore differs from traditional programming: instead of translating business rules into algorithms, suitable models are created.
The first step is to choose a suitable architecture for the model. Just as the architecture of a railroad station building differs from that of an opera house, the architecture of AI models also differs depending on the area of application. A model for image recognition needs a different architecture than one for speech recognition.
The second step is to optimally configure and train the model. As there is a high proportion of trial-and-error loops, it is important to keep the development process as lean as possible. The Python programming language with its well-documented libraries in the areas of data processing, statistics, neural networks and visualization is often used as a technology for the development of AI applications. Important to know: Once a model has been successfully trained, it can be integrated into almost any technology.
7. Optimize the AI model with expert knowledge.
If the performance of the trained AI model does not yet meet the requirements, it can be further optimized with the expertise of experts. This is often done according to the principle of reinforcement learning. The correctness of a classification or prediction is evaluated positively or negatively and the model is adjusted accordingly. If experts are involved, this looks like this:
- The system creates classifications or predictions on a trial basis.
- Experts evaluate their correctness.
- The model is adjusted based on direct feedback.
- This procedure is repeated until the system meets the requirements.
8. Use experience from productive operation.
The optimization is not over with the productive introduction: If you draw on experience from productive operation, the AI system can be further optimized even after its introduction. This allows you to iteratively add further relevant attributes to the data and use the existing productive data to automatically check whether this addition improves the system. In this way, you get closer to an optimal model with every iteration.
Conclusion: Without data, everything is nothing
For value-creating AI applications, an understanding of your own data is essential. This requires
- several passes and
- with the support of experts from as many areas of the company as possible
- identify relevant data, process and supplement.
Those who approach AI projects as marathons run the risk of running out of breath along the way.
It is more promising to plan such projects as a relay race with small sprints: an iterative approach leads to new insights more quickly and experience has shown that it has a higher success rate.
This is how AI applications become a competitive advantage.
