Stibo Systems CONNECT 2023

Stibo Systems is a master data management company based in Denmark. Its 2022-23 growth was 15% over the previous year. It has over 800 employees in 18 global offices. The main Stibo Systems CONNECT (global customer conference) was held in November 2023 in Madrid; it alternates between the USA and Europe each year. There were over 400 attendees at this conference with many Stibo Systems customers including Miele and Marks and Spencer. It has an unusual corporate structure in that it was originally a printing company set up in 1794 and is still privately held. Stibo Systems is one of four companies owned by the Stibo Foundation. Its sister companies operate in the graphics, publishing and magazine industries.

The keynote was given by CEO Adrian Carr. His message was that the world is getting more data centric, and generative AI and the increasing adoption of the cloud are important trends. Multi-channel retailing has always been Stibo Systems’ core market, and enhanced product Information and content is important to engage customers. A demo showed AI generating a product description based on a technical specification of a particular product, in this case a set of headphones. This included descriptive text and images but also a comparison chart with other similar products based on the entries within the Stibo Systems product Information management database.

Microsoft are a major partner (including deployment of the STEP software on Azure) and an executive of Microsoft was on stage with the Stibo Systems CEO for an introductory discussion. She was (unsurprisingly) enthusiastic about the potential of generative AI for business and general life, enabling significant improvements in productivity. GitHub Copilot is used by the 90,000 Microsoft engineers and they claim to have experienced a roughly 50% productivity improvement from using this generative AI code generation tool. She expects companies to be able to produce customised copilots of their own to deal, for example, with many administrative tasks.

I attended several demonstrations and breakout presentations during the conference. AI was a common theme. Artificial intelligence “would have been a good idea”, as Gandhi said about western civilisation. Machine learning has been used for some years by Stibo Systems for auto classification. You may have to train your own language model on your own products for successful product classification; this tends to be effective but quite narrow in scope. Humans tend to be poor at repetitive tasks so using AI for this task can be quite productive. For example, for a particular laptop computer with various attributes in the PIM database, the software can suggest classifications based either on industry standard classifications or on a company specific trained model.

Generative AI is trained on much larger datasets and has broader application than machine learning. In practice, if you train your own model on a few hundred products then a model can produce quite high-quality classifications. A demo showed a mobile phone product being automatically classified based on a trained model, saving time over the previous manual method of looking up hierarchies. Generative AI is heavily dependent on good prompt engineering and can be effective but its output should not be fully trusted without human supervision; also, machine learning models rarely produce identical text repeatedly. STEP enables prompt engineering within the product, for example you can change the AI “temperature” to produce more flowery language for product descriptions. Users need to be aware that bulk use of the generation feature can use a lot of processing power and so can be expensive in a cloud computing context. The demo also showed image generation, for example seeding images to help the AI produce something loose to what you want using Dall E 2. A demo responded to a text description by producing an image (admittedly after an error message, this being a live demo). Other possible uses of AI in Stibo Systems software include anomaly detection.

In the latest software release 2023.3, STEP has improved configuration management (going from development, test through to production). This includes automated transfer between systems without explicit export and import, and more efficient dependency analysis. There is also better defect reporting. In the product information management core there is an improved class editor and easier mapping of products to industry standard classification schemes, including enhanced search capability for non-product data domains such as customers, suppliers and location. For syndication, you can now invite customer retailers to manage aspects of the digital catalogue, and they have added wider API capabilities such as querying products directly from an API.

An interesting use case was a customer called Thieme, a medical journal and digital health provider with 800 employees in Germany and revenue of €182 million. They are transitioning from traditional print journals to digital products. They have a complex application landscape with seven different CRM systems alone, making it hard to get a grip on their customers. Even knowing the sales for one customer is not easy given how diverse are their systems with various bits of customer data scattered across the enterprise. They use STEP to integrate the customer data into one place and so get better understanding of customers and improve the data quality. For example, they have a customer (Pfizer) in their SAP system but also in Microsoft Dynamics as one of their 400,000 customer records. The data records do not match (with very different names in the systems) and they want to match up legal entities. They used ChatGPT 4.0 to fill in customer attributes to make it easier to compare customer records. They use OpenAI on Azure cloud and built a prompt engine and caching system (for efficiency) to populate a new temporary database with extended attributes of the customer data and then better match records to populate their new MDM system. This was done via an API so it was a fully automated process. Careful formulation of the prompt to ChatGPT enabled matches of the legal entity names in German with a high degree of success. Additional instructions to the AI were provided, such as “never invent a legal entity”, to try and minimise AI hallucination. They reckoned that this achieved an overall 96% accuracy of the AI output, found by validating sample data with human experts. 85% at least of the AI suggestions were semantically correct. This further enabled 89% matching accuracy of records in STEP compared to around 30% previously. The process cost around 8 cents per record compared to 20 cents per record doing it manually. This project saved 500 work days. It was interesting to hear about a generative AI project being used with structured data rather than for generating text of images.

Miele is a German high-end domestic appliance manufacturer dating back to 1899 and still family owned, despite turning over well over €5 billion annually with steady growth. Notably, its manufacturing capacity is entirely based in its factories in Germany and Eastern Europe. Miele has been using STEP since 2015 to provide consistent product and marketing content. They operate across multiple sales channels, including direct to consumer. Their products are designed to last at least twenty years so there are many product versions and models to deal with over their lifetime. Miele use SAP for transaction processing but STEP is their central “product cockpit”. In 2016 they evaluated several MDM products including Hybris from SAP but selected STEP and embarked on a major migration project. In February 2023 they finished replacing their SAP product catalog globally with STEP. They have 100,000 SKUs, 30,000 content objects, two million assets in 75 categories and 20,000 attributes. Miele has to deal with forty different languages across its over fifty markets. Data comes from various core systems such as material master in SAP, and also CRM and a video database. Channels include direct sales, mobile sales, portals, business partners and even printed material. Data quality is important to them for regulatory reasons amongst others. They found it crucial to engage business lines with the data model, not just leave it to the IT folk. This is a quite large-scale master data management implementation.

BSH is a €16 billion European domestic appliance supplier with 63,000 employees and brands such as Bosch, Neff and Gaggenau. They put in a master data system as far back as 2001, drawing data from their core SAP ERP system. An important part of this has been the implementation of data governance with a high degree of business engagement and clear data ownership. Numerous challenges had to be overcome in this process including limited data governance and poor data quality. In particular they have been migrating from SAP R/3 to the latest S/4 version, including consolidating seven separate R/3 implementations. They considered having a single centralised hub, but increasingly data now comes from suppliers or even directly from consumers. They have consequently implemented a federated approach where master data can be created either centrally or at a local level. There has been a gradual approach with tandem master data systems bridging the old and new worlds. They are three years into a large migration project that is scheduled to take until 2030 before R/3 should be finally switched off. BSH has already achieved carbon neutrality and uses their master data hub as part of the foundation for their sustainability initiative. Stibo Systems has built in some specific support for ESG (environmental, social and governance) data in its product, with data cards for various aspects of this such as recycling and certification data.

MMS is a German retailer operating in twelve countries, and is Europe’s largest electronics retailer, with their main brand being MediaMarkt. They have many sales channels including physical stores through to social media. Countries have their own systems which are fed into a central product hub. Products were maintained at the country level, which made it easier to localise and adapt to specific local markets and legislation, but of course this makes standardisation difficult. The speaker showed an example of an iPhone and how descriptions and measurements were inconsistent eg the physical dimensions of an iPhone appear, according to their systems, to be different in Spain and Germany, which of course is not true. The company of course wants to be able to monitor sales at a corporate level, not just a local level. The company started by carefully documenting the existing processes associated with data and set up a central data governance team to standardise across the various countries. They then implemented a new global product identifier across the whole company and are in the process of putting in a new centralised master data management system to replace the current landscape. A core unique data record for each product will have many attributes such as local descriptions in local languages as well as technical specifications and product classification. Local countries can enrich the core data eg there might be a local video promoting a certain product in a local language. They use machine learning to predict the classification of a new product, which achieves a 97% accuracy. STEP is the core of this new architecture.

Another customer case study was on how a pub can undergo a digital transformation. Mitchells and Butlers has been running since 1898 and has many brands including Harvester, Toby Carvery and All Bar One. Their annual revenues are £2.2 billion with 44,000 employees, with 1,600 pubs and restaurants and 16 brands; they serve 300 million meals a year and 500 million drinks. Their head of data management explained how master data management was implemented at the pub chain. Product and location data are the main data domains managed. In 2019 they started a digitalisation project. This involved better data governance, and enabled initiatives such as being able to order food at the table digitally and also to be able to deliver food as well as serve it in dining rooms. Each physical site can now log in to STEP and enter information such as local opening hours, which is fed to Google to ensure that customers see accurate opening times for each site when searching. Other data that is mastered includes when certain menus are served eg when does breakfast end and lunch begin, crucial for digital orders. One operational issue is if a kitchen is temporarily out of action the pubs would previously call a central Helpdesk to ensure their digital services were temporarily turned off (this happens 700 times a month, to give a sense of scale). One service that is now offered is the ability for each site to adjust their availability via the MDM system; this data is automatically transferred to the ordering mobile app. Collecting data at source in this way improves efficiency and gives a better customer experience. At present the menus of dishes are entered locally but this will shortly also be held in the MDM system, enabling a new menu to be rapidly implemented and avoid data being rekeyed into multiple systems. There are still constraints due to the need to print menus, as in the UK customers are quite resistant to QR code menus.

Wurth Group makes and sells fastening equipment including screwdrivers, fasteners and adhesives. It has 400 subsidiary companies turning over €20 billion in total in 2022, with 86,000 employees in 80 countries and over 4 million business customers. Their head of eBusiness explained that they had been using STEP for 13 years. They have a centralised product information system for the whole group. Each subsidiary can see their own data and global data only, but product data maintenance is decentralised across all companies. Products can be initiated locally but they must then maintain that product data. There is a uniform data model with a central pool of technical attributes, roughly 30,000 in all. They have clear data governance rules and can share product data with partner companies in a controlled manner. Suppliers can be onboarded within the central system. Currently there are twenty million products defined in the system and over thirteen million assets. Language translations are also carried out centrally. Despite the large numbers of data assets being handled, the actual performance of the system has generally proved quite good; currently this runs on premise rather than in the cloud. Their experience has shown the importance of a stable data model and the need for clear data governance rules.

The last session that I attended was from Stibo Systems itself, its Chief Product Officer Nede Nia and Product Innovation Director Jesper Grode discussing generative AI. Stibo Systems already use machine learning in their product but generative AI is a newer field. One application is the use of AI in customer service to provide a more personalised chatbot experience, based on a large language model that has been trained on Stibo Systems customer experience data. Stibo Systems are also looking at using the Microsoft CoPilot initiative to improve productivity in programming. They have also established a dedicated research team to investigate brand new AI possibilities beyond these applications. One example is generating product descriptions from the master data hub attributes. One key to success in using AI is careful prompt engineering, and there is still a need for human interaction and review. A demonstration was given of a generated product description of, to take an example, a bottle of mineral water. A generic request generates a quite generic response from the AI. However, a much more precise prompt generated a much more specific product description. Not surprising perhaps, but it shows the need for very detailed prompts. Stibo Systems have built a prompt engineering prototype tool that enables a user to build a prompt using a template that helps you build a more precise prompt, for example giving various pick lists of elements to help ensure that you give enough detail. This contains various prebuilt examples to help you. There is a trade-off between productivity and quality, as human review will improve quality but of course takes time. You can ask the AI to explain its reasoning and provide references where appropriate, which will help a human domain to review machine generated material. Customers need to be aware of privacy issues in using public AI, and to consider whether their data is or high enough quality to train an AI model. AI is not always the answer to every question!

Overall, the event was very well organised and was of a manageable size. There was a large contingent of customers, who seemed highly engaged with Stibo Systems and its product journey.