DE
|
EN
Examples from Practice

Examples of Artificial Intelligence in Business Practice

Text: Shirley Ogolla, Vivien Hard

The following selection of real-life case studies provides examples of smart and autonomous systems and illustrates their growing relevance in very different application areas of business practice. The case studies come from the research project “AI and Knowledge Work – Implications, Opportunities and Risks” (known as the KIWI project in German) conducted by the Alexander von Humboldt Institute for Internet and Society (HIIG) and financed by the AI Observatory of the BMAS (German Federal Ministry of Labour and Social Affairs).

BIO

Assistance in the analysis of documents for mortgage loans

Context:

The ING Bank wants to offer its customers a simple and clear digital experience – anytime and anywhere. Since summer 2019, all organisational units of ING in Germany have followed agile work practices. Agility is the business culture’s response to the digital transformation. The strategy is to become a data-driven business, boost customer satisfaction and remain a digital market leader. In the area of mortgage loans, ING is developing applications for computer-assisted document analysis.

Deployment scenario:

With the help of machine learning, the application is to extract information from documents such as payslips that are submitted by customers applying for property financing. Using natural language processing, this information can then be classified and organised into relevant information fields such as information on the applicant’s actual net income.

Initial situation:

When customers apply for a mortgage loan, they are required to submit an array of documents to provide information about their creditworthiness. The aim is to increase the level of automation in the data processing production processes.

Specific problem:

Originally, the data from these (PDF) documents was entered manually into corresponding tables by staff in the mortgage lending division. This was a very time-consuming process.

„There is a need for internal further training programmes for staff, including opportunities for those who wish to change career and enter IT. “

Initial question: How can we automate simple sub-tasks in document analysis in the area of mortgage lending?

Successes:

  • The application could offer promising support in data processing for simple sub-tasks in the mortgage lending service.
  • Mortgage lending staff would therefore be able to invest more time in the decision to provide a loan.
  • Against the backdrop of potential machine learning bias, the decision to grant a loan is ultimately always taken by the members of staff.

Lessons learned:

  • Internal data science expertise is a key prerequisite for the successful implementation of such projects and can pose a challenge in light of the shortage of skilled professionals.
  • The financial sector is already highly regulated and the possibilities for the large-scale deployment of artificial intelligence (AI) applications are limited accordingly. ING is therefore pursuing an augmented intelligence approach.
  • There is a need for internal further training programmes for staff, including opportunities for those who wish to change career and enter IT.

Assistance in filtering out toxic social media comments

Context:

“Readers of ZEIT ONLINE, the online version of the DIE ZEIT weekly newspaper, like to discuss topics, adding thousands of comments to our articles on a daily basis. While the vast majority of these contributions are constructive, about 200 comments each day are classified as having harmful content and are either abbreviated or removed.”1 (Andreas Loos, Data Scientist ZEIT ONLINE)

“Our community team is responsible for moderating the comments. We monitor content 24/7 to ensure that discussions on ZEIT ONLINE are respectful and constructive. All readers should feel welcome on our site. No-one should prevent others from sharing their opinion and knowledge by posting insulting or ostra­cising comments. This is the ideal scenario – by definition unattainable – which we are striving for.” 2 (Julia Meyer, Community Team Lead ZEIT ONLINE)

Deployment scenario:

The ZEIT ONLINE community editing department develops and tests an AI moderation tool called ”Zoe” for online comments. On the basis of natural language processing, the AI moderation tool recognises toxic Germanlanguage content.

Initial situation:

The ZEIT ONLINE community editing department has a team of moderators who monitor online comments manually around the clock and filter out all comments that breach netiquette3/discussion rules.

Specific problem:

The ZEIT ONLINE community editing department receives up to 80,000 comments per week. 20,000 readers a week post several hundred comments, which are also processed simultaneously by the community editing department. Having an AI application that helps flag comments that probably need to be filtered out would be an enormous help for the moderation team.

Initial question: How can we use AI to provide round-the-clock assistance to the moderation team, particularly during peak loads?


AI can help to filter online comments that violate netiquette and discussion rules. Photo: Apichon_Tee / Shutterstock.com

Successes:

  • Using historical comment data, the AI application has already been trialled and tested, and is currently being trained and undergoing continued development.
  • The tool allows the moderation team to invest more time in difficult comment-related decisions.

Lessons learned:

  • It is much easier for the AI application to filter out toxic comments than particularly good comments, as research has already developed a number of applications to detect toxic content.
  • The AI application can certainly help the moderation team to maintain the comments on ZEIT ONLINE and foster a respectful culture of debate. Far from making the moderation team redundant, the tool allows the team to concentrate on the more acute cases.
  • A high degree of pragmatism together with technical in-house expertise is a key factor in the success of this project.

„It is much easier for the AI application to filter out toxic comments than particularly good comments, as research has already developed a number of applications to detect toxic content.“

Assistance in the evaluation of ideas in improvement management

Context:

“The process of digitalisation affects all business segments at Volkswagen. In addition to the car itself and various mobility-related services, we are also going digital with development, vehicle production and the entire factory and office working environment.” 4

Deployment scenario:

  • Idea management is a department within HR where employees submit ideas for the improvement of all business processes and where these ideas are processed and evaluated on the basis of applicable rules.
  • Between 2017 and 2019, idea management was supported by the IdeenOnline Playground (IOP) software, which analyses the database containing all the improvement ideas already submitted and in doing so allows idea management to see this data from a new perspective. For example, this software enables a real-time search for known improvement ideas and can also compare ideas and assess their similarity. This helps to speed up processes in idea management significantly, as suitable evaluators for these improvement ideas can be found far more quickly, for instance.
  • A few facts about IdeenOnline Playground:
  • IOP was developed entirely by an employee of the strategy and innovation department in the HR division and is built with cutting-edge open-source technology.
  • It uses natural language processing (NLP) to linguistically analyse the database currently containing over one million improvement suggestions, and also deploys unsupervised machine learning techniques to find the relevant improvement ideas in this context from the stock of data in less than one second.
  • Currently, IOP is not fully utilisable due to a change of system in Volks­wagen’s IT infrastructure. As the advantages of IOP are indisputable and the change can produce new useful functions, plans to adapt the software are already in place.

„A pragmatic approach to administrative requirements and effective networking with other staff members can facilitate swift progress and prevent developments from failing before they even get off the ground.“

Initial situation:

Over 20,000 new ideas for improvement are submitted each year, with the result that every member of staff in idea management needs to process over 1,000 new ideas per year.

Specific problem:

In addition to the ideas that already exist, each member of staff in idea management has to process new ideas constantly. To assess an improvement idea, for example, it is necessary to identify suitable evaluators who are experts in their field. Given the size of the company and the wide range of topics, finding these experts is a very time-consuming process.

Initial question: How can we simplify the search for evaluators to assess the ideas?

Successes:

  • In the case of new improvement ideas, the search function coupled with similarity analysis helps idea management staff to find earlier, similar ideas and the idea evaluators. Suitable evaluators for new ideas can therefore be identified far more quickly.
  • Combined with additional, useful functions, the administrative workload of idea management staff can be reduced, and the time saved can be used to support the actual implementation of ideas even more effectively.

Lessons learned:

  • The freedom to develop innovative projects is based on the support and trust of management. A pragmatic approach to administrative requirements and effective networking with other staff members can facilitate swift progress and prevent developments from failing before they even get off the ground.
  • The integration of applications into the existing IT landscape often poses a significant challenge given the broad range of requirements to be met. When developing new applications, particular attention should therefore be paid to application integration from the outset in order to facilitate subsequent production roll-out.
  • Technology-savvy staff members and employee representatives should be the first to be brought on board, as they are open to experiments and are happy to give very useful feedback for the further development of the application. Furthermore, they also act as disseminators and can convince any sceptical colleagues of the benefits of the new technology.

Assistance in answering HR-related questions in HR management

Context:

The HR systems department at Siemens has been using a smart chatbot in human resources since 2017. The chatbot is included in a single point of contact (SPoC). The chatbot component in CARL, as the SPoC is known (named after the son of Werner von Siemens), was co-developed with IBM and uses the IBM Watson components IBM Watson Assistant and IBM Watson Discovery. It provides Siemens staff with round-the-clock information on HR-related topics in an anonymised format in five languages.

“The fact is that the new technologies are now part of our lives and the pace of technological advancement continues to accelerate. As HR, we also have to take an innovative approach and try out, understand and deploy and/or provide new technologies in order to offer new solutions for our own use cases and also to be able to keep pace with our business and deliver effective support in this transformation process.” (Sabine Rinser-Willuhn, HR systems, Siemens AG)

„An agile mindset and team are factors of success, as so much was changed and adapted in the course of the project and the team acts like a start-up within the actual company.“

Deployment scenario:

CARL is an international AI project that uses AI functions of IBM Watson in the project’s chatbot component: IBM Watson Assistant technology is the basic capability for the natural language processing. It is based on a supervised learning model which experts must populate with both potential questions and the answers to these questions. The system’s intelligence lies in its ability to answer variants of questions in the same context in future on the basis of a few initial questions.

Watson Discovery is a cloud-native insight engine that combines data ingestion, storage and preparation using natural language processing in order to extract insights from structured and unstructured data with AI-assisted queries.

Initial situation:

The computer-assisted provision of advice on HR-related issues, such as sick leave, vacation planning, performance management processes, business trips or information on continuing and further training measures, is to be introduced. The aim is to provide 24/7 support for members of staff on all kinds of terminals, and also to reduce the workload of HR departments, giving HR more time and resources to address complex issues by having a machine answer simple, recurring questions.

Specific problem:

The German Siemens HR department alone receives questions on some 360 HRrelated topics from around 120,000 employees. Many of these questions are easy to answer.

Initial question: How can simple HR-related questions be answered auto­matically?

Successes:

  • Currently, the SPoC CARL is contacted over one million times per month.
  • The chatbot gives staff round-the-clock information on around 290 HRrelated topics from all devices. It currently reports an average of 60,000 interactions per month.
  • It now speaks five languages, has gone live in 20 countries and therefore currently reaches around 280,000 employees. The goal is for the majority of Siemens’ 350,000 staff worldwide to communicate with CARL by the end of 2020.
  • With their workloads eased thanks to the chatbot, HR staff have more time to dedicate to more in-depth questions.
  • Staff at Siemens can get answers to their questions more quickly.

Lessons learned:

  • An agile mindset and team are factors of success, as so much was changed and adapted in the course of the project and the team acts like a start-up within the actual company.
  • The chatbot launch was very open; project management addressed users’ questions directly.
  • The trust of management is very important in order to have the necessary freedom and flexibility in the development and implementation process.
  • The successful implementation of a project of this size requires stamina, patience and willingness to experiment considering that completely new technologies are used and supported by new methods (such as Design Thinking, Agile, SCRUM, etc.).

Assistance in image searching in the production of education media

Context:

The Cornelsen Group is one of the leading providers of education media in German-speaking countries, with publishing companies such as Cornelsen Verlag, the Bibliographisches Institut (Duden), VERITAS or Verlag an der Ruhr. People have used Cornelsen education media to teach and learn for over seven decades. Cornelsen fosters educational potential from early childhood through to adulthood and working life. In the development of educational processes, the group is relying on the potential of digital technologies.

Deployment scenario:

Cornelsen has been using an image search technique in its editorial work since 2018. On the basis of a specific image, the image search function searches the internal image database for a similar image without having to rely on indexing.

  • Image search is integrated into the existing editing software, allowing editors to search through the database of images which Cornelsen has purchased for the closest possible match for another image.
  • Based on an open-source image recognition model and using a neural network and a clustering algorithm, the image search function suggests the images that are the closest possible match.

Initial situation:

Educational products, such as schoolbooks, are illustrated in a manner that effectively supports the learning process.

Specific problem:

Editors often already have a particular image in mind and are looking for a similar image the rights to which Cornelsen has ideally already purchased.

„The development of AI applications requires digital business understanding, experimen­tation and pilot phases, extensive data access and exchange with domain experts.“


The faster way to get the right picture: AI applications can help to find similar pictures which helps in the process of producing didactically meaningful educational media. Photo: S_L / Shutterstock.com

Initial question: How can the closest possible matches for a particular image be identified in the Cornelsen image database?

Successes:

  • AI-assisted image searching accelerates the process of producing educational material.
  • This gives editors more time to invest in content.

Lessons learned:

  • The development of IT products in general and AI applications in particular requires digital business understanding, experimentation and pilot phases, extensive data access and exchange with domain experts.
  • Therefore, constant exchange with the expert users – in this case the media managers and editors – is crucial in order to understand the work processes and be able to support them effectively.
  • Here, it helps to prioritise according to potential time savings for routine tasks in order to get the necessary time for support from the expert users.
  • The use of open-source models can significantly speed up the development of AI applications.
  • Early exchange with IT administrators and developers is important for sub­sequent production roll-out to existing systems. One challenge in this context is to be able to offer approaches and prototypes as quickly as possible as the basis for discussion.

Assistance in the digitisation of historical library collections

Context:

“As Germany’s biggest academic universal library, the Staatsbibliothek zu Berlin (Berlin State Library, SBB) is a central source of national and international literature. More than 11 million volumes of printed material alone have accumulated since the library was founded over 350 years ago. Furthermore, the collection comprises over 2.2 million additional printed works and other, often unique, ma­terials in the special collections – including western and oriental manuscripts, music manuscripts, autographs, unpublished works and papers, maps and historic newspapers. The library’s collection also contains more than 10 million microforms and, in the photographic archive, over 12 million motifs.”5

Since 2010, the SBB has been digitising its public domain collections in the Digitisation Centre specifically created for this purpose in Berlin. Here, the SBB creates digital collections from analogue documents (manuscripts, prints, maps, sheet music, etc.) and makes them available to the public online.

Deployment scenario:

A total of three AI applications are being developed in the QURATOR project. A layout analysis of documents is performed using deep learning procedures for image recognition (1); text recognition is performed using neural networks (2); and semantic analysis is performed using natural language processing (3), e.g. for named entity recognition.

Initial situation:

The aim is to apply the promising approaches of AI to the particular challenges of historical documents. The SBB’s repository contains around 2.5 petabytes of data, providing a vast quantity of training material.

Specific problem:

The SBB is digitising all the documents that are free of copyright in its collection (15th century–1920) and making them available online6. However, several com­plex processing steps and technical challenges need to be overcome before digitised sources are just as convenient to work with as digital-born documents. Existing AI solutions are not suitable for the historical particularities of writing styles and spelling, as they contain many distinctive features and differences compared with the standardised language of today.

Initial question: How can the quality of the digitised content be improved by AI-based processes?

Successes:

  • Thanks to the AI processes, it has been possible to make a decisive quality improvement in the further processing of digitised content, both with regard to the quality of layout and text recognition and the semantic enrichment of the documents.
  • In future, more documents can be made accessible more quickly and effectively and will therefore also be more easily searchable.
  • The scholars can now apply methods from digital humanities, such as text mining, as in terms of text accuracy the quality of the digitised content is close to the analogue originals and the digitised objects are available globally online.

Lessons learned:

  • The SBB has been involved in the research field of document analysis and text processing for many years, so it had developed a good foundation for this project.
  • It is difficult to scale how much computing power will be needed for the project on the long term. For example, the training of AI models will place considerable demands on specialised hardware at times.
  • The machine learning engineers can incorporate the necessary domain knowledge such as knowledge of certain types of handwriting from the Middle Ages into the AI development process directly from the scholars and experts on site at the SBB.

For more information on the research project Artificial Intelligence & Knowledge Work – Implications, Opportunities And Risks visit the KIWI project website on the Alexander von Humboldt Institute for Internet and Society website.

https://www.hiig.de/das-institut/

Contact information:

shirley.ogolla@hiig.de

share page