Case studies in Machine Learning and AI

Pricing was heavily dependent on the market, yet at the same time only partially data-driven. Competitors were monitored manually, and prices were adjusted occasionally. The problem was that the market moves faster than we can keep up with. Furthermore, products are rarely directly comparable. A flyer is not simply a flyer; paper weight, format, finishing, print run, and many other factors come into play. This makes it correspondingly difficult to compare prices meaningfully. Decisions were often based on partial information or experience, which inevitably led to inconsistencies.

The goal was to make price monitoring and adjustments significantly more systematic. The aim was to automatically analyze competitors, identify relevant products, and compare them with the company’s own catalog. This process was intended to generate well-founded, transparent pricing recommendations. However, the final decision remains with a company employee.

The solution is a system that collects publicly available product and price information from defined, authorized sources, using HTML parsing or available APIs depending on the source. An AI agent analyzes the products found and compares them with the company’s own product range. Descriptions, attributes, and specifications are semantically compared to identify meaningful matches even when names and structures differ. In a Laravel-based web interface, suitable competitor offers, price ranges, and a rule-based recommendation are displayed for each of the company’s own products. The calculation bases can be maintained via Excel import. An employee reviews the suggestions, adjusts them as needed, and approves them. Only then are the new prices updated in Shopware via an API.

A key issue was the legal classification. Not every online store permits the automated extraction of data, and depending on the provider, such solutions can quickly end up in a legal gray area. Here, it was important to proceed with care and selectivity, as well as to develop a sound data acquisition strategy. Before connecting to external competitor data sources in production, we work with the client to determine which sources are technically and legally viable. In doing so, technical protective measures on the part of competitors are never circumvented. Competition law also played a role, particularly regarding the extent to which prices may be automatically aligned with the market. Accordingly, the system was designed to provide recommendations but not to make automatic pricing decisions.

The effort required for market monitoring has decreased significantly. Instead of manually checking individual prices, a structured overview is now available that is continuously updated. Price changes in the market are identified more quickly and can be addressed in a timely manner. At the same time, the predefined rules ensure that margins are not undersold. The bottom line is that a robust foundation for data-driven pricing decisions has been established, paving the way for more advanced, dynamic pricing strategies.

Product data was available, but rarely where it was needed at the moment, and, above all, not in a standardized format. At the same time, demands were growing from sales partners and resellers, whose online stores require structured, clean product data, ideally in a format that can be imported directly. Preparing this data manually is extremely time-consuming for a product range that has evolved over time. In this case, “historically evolved” meant that the creation and maintenance of product data had never been standardized, but was instead handled on a case-by-case basis by the respective employee who was working on it at the time, in whatever way they deemed appropriate. On top of that, the employees were already fully occupied with day-to-day business and had no time available for data preparation. At the same time, the implementation of a PIM system was still in its infancy.

The goal was to generate usable product data from the existing unstructured information. Relevant technical features and characteristics were to be automatically identified and extracted, regardless of their source. This data was then to be converted into a uniform format and made directly usable in the PIM system. It was important that the process be largely automated, but remain verifiable by employees at critical points.

In a Laravel-based web app, documents and texts can either be uploaded directly or imported via interfaces. An AI agent extracts relevant content, identifies technical characteristics, and formats the information into a defined, structured format. What makes this unique is the semantic understanding of the content, for example, recognizing that “max. pressure 10 bar” is a specific product attribute. The results are displayed in the web app and can be reviewed by the relevant department and corrected if necessary. This step is intentionally included to ensure quality and trust. After approval, the data is automatically transferred via API to the Akeneo PIM system, where it is immediately available for further processes, such as for web shops or catalogs.

Technical details were scattered across data sheets, catalogs, ERP entries, and sometimes even emails, and were often incomplete. Some of the documents were technically complex, containing tables, inconsistent formats, and, in some cases, highly specialized technical language. The data processing system had to be robust enough to handle even imperfect input. In addition, not all information was always complete or unambiguous. The system therefore had to not only extract data, but also recognize when something was missing or uncertain and flag these cases accordingly.

Within a short period of time, product information for several hundred items was automatically processed. Data quality has improved significantly, primarily due to the standardized structure and the improved accessibility of information. The manual effort required per product has decreased significantly. Instead of researching data from scratch, the focus is now primarily on verification and fine-tuning. At the same time, a solid foundation for the PIM system has been established, upon which further processes can be built seamlessly.

Many calls go unanswered and end up on the answering machine. Listening to and following up on messages takes a significant amount of time each day, time that is actually needed elsewhere in the day-to-day operations of the practice. On top of that, not every message is immediately clear. Information is missing, misunderstood, or lost in the midst of the chaos. This leads to follow-up questions, delays, and ultimately, dissatisfied patients.

The goal was to make managing voicemail messages more efficient without relinquishing control. Incoming messages should be automatically transcribed into text and their content analyzed. Relevant information, such as the nature of the request, urgency, or a request for a callback, should be presented in a structured format. Based on this, requests should be prioritized, categorized appropriately, and initial follow-up actions should be triggered automatically.

From a technical standpoint, we’ve integrated the existing voicemail system via Sipgate with webhook support. Each new message is converted to text as an audio file using WhisperAI. A language model then performs semantic analysis. It identifies the subject matter, the urgency of the request, and which information is relevant. Based on this, predefined workflows are triggered, which control the messaging system of the existing practice software to schedule appointments or send notifications to the practice staff.

The biggest challenge was integration. The existing practice software is functional, but it wasn’t primarily designed to communicate seamlessly with external systems. This required us to develop an interface to enable communication. A second key consideration was data protection. Especially in the healthcare sector, the handling of sensitive information is non-negotiable. Accordingly, in close consultation with the client’s data protection officer, we ensured strict segregation, minimal data storage, and controlled access paths during processing.

The manual effort involved in managing the answering machine has been significantly reduced. The team no longer has to listen to every single message; instead, they work with structured, clear information. At the same time, fewer details are lost, and urgent matters are identified more quickly. This is also noticeable from the patients’ perspective. Feedback is provided more quickly and reliably, which is directly reflected in patient satisfaction. Overall, a process has emerged that runs noticeably more smoothly, even though the volume of calls has remained the same.