AI chatbots: the sparring partner at the workplace.

Only a few hours to go until the phone call with the defaulting consumer. By then, EOS colleague Petra must have a complete picture: How often has the consumer already been contacted? Have they already been offered payment by installments? If not, what installment amounts would be realistic? Are there indications that they could refuse to pay completely? Previously, this has meant that Petra has had to go through all entries in the consumer’s electronic file, including brief notes by colleagues that often contain unclear abbreviations. “This is a highly time-consuming process,” says Paul Baltag, Operations Manager at EOS KSI Romania, “and every colleague has to start from scratch every time they need to contact the consumer.”

A project by Paul’s team is showing what the future could look like. Just one click – and all relevant facts and figures should appear on the screen, clearly structured and compactly summarized. This is made possible by ChatGPT, which is meanwhile the best-known generative artificial intelligence worldwide. “We wanted to develop tools based on ChatGPT that our colleagues can use for typical routine tasks,” says Paul. Like the chatbot that can summarize digital customer files in a structured way. Paul calculates that this saves colleagues up to 80 percent of the working time spent on analyzing previous actions.

Nevertheless, there was also a challenge to be overcome. “Naturally, we must ensure data privacy,” says Paul. Because ChatGPT stores data from the prompts in order to learn from them. This is why the project team initially made copies of the documents, replaced sensitive data like insurance numbers or names of the consumers with dots, and used these amended documents to produce the prompts for ChatGPT. In the chatbot’s replies, the dots were then replaced by names and numbers. However, that only works up to a certain data volume, says Paul. “If you anonymize too many details, at some point the ChatGPT answers become inaccurate.”

Other tools developed in the project included a “legal assistant”, for example. If consumers ask EOS legal questions like “On what legal basis can EOS take over the claim?”, it provides quick answers, including the relevant clause from the legislation. Another tool was a chatbot that evaluates anonymized customer data, and based on predefined parameters predicts the likelihood of the customer ceasing to make payments at some point.

Already, generative AI can be used for a wide range of routine tasks. Paul uses chatbots, for example, for the following tasks:

• As a substitute for googling. “If I ask Google something, I only get a compilation of websites more or less related to my question,” says Paul. “When I ask ChatGPT a question, I get an answer.”
• To develop ideas for the structure of a presentation from just a few key words or sentences.
• To give a stylistic polish to hastily scribbled texts like emails and formulate them in a more complex way.
• To find suggested topics and help with wording for topics on which we can connect with business partners and customers on LinkedIn.
• As a sparring partner for developing ideas, improvement suggestions and strategies.

Meanwhile, there is a whole range of AI chatbots for various fields of activity, adds Paul. Some of those he finds useful are:

1. ChatGPT 4: Alongside the conventional chatbot function, this software also enables precise data analyzes, the automatic revision of text files, and optical character recognition (OCR), which means that text can be read from photos or PDFs, for example. 
2. Cohere Generate can summarize texts in various styles.
3. Synthesia can help you produce personalized videos. For an avatar, all you need is a photo. 
4. GitHub Copilot is geared towards programmers. This tool can improve and correct code – or at least give indications of possible errors.

In future, many other use cases are conceivable, says Paul, including for other companies in the financial sector. Chatbots, for example, that can provide a decision-making tool for investments, e.g. by providing answers to fundamental questions like “Which parameters are critical in this industry? What legal risks may occur?” Such tools could not just help EOS to decide, for example, what price would be reasonable for a package of non-performing loans, but could also help investors wanting to estimate the development of securities or exchange rates. Until now, too little data has been taken into account for this purpose, explains Paul. “Often, you just compare with the purchase prices for previous transactions,” he says: “However, generative AI can also incorporate data on the parallel development of wages, inflation, unemployment and other factors.” Auditing company EY also envisages numerous use cases for generative AI for staff in finance departments, including:

• To make it easier to prepare documents with operating results for decision-making in areas like benchmarking, due diligence or market segment analyzes.
• To simplify access to the financial data necessary for business intelligence or performance management processes, i.e., the analysis and control of the company’s performance.
• To generate data on the reference values and drivers crucial for financial planning and the forecasting of financial developments.

Many people share this positive view. In a worldwide survey of employees by auditing company EY, 63 percent of respondents expected AI to simplify their work. However, for others, the issue did not inspire enthusiasm but anxiety about losing their jobs.

Wrongly so, Paul believes: “I do not believe that AI will replace human workers.” In the past, the arrival of a new technology – whether in the form of machines or computers – has often triggered fears that it would make human beings superfluous. Paul predicts: “Generative AI will lead to employees improving their efficiency and having more time to deal with innovations.” He says that for companies, this is a great competitive advantage. Paul’s advice to his colleagues: “Use AI, and don’t give up immediately if you get poor or incorrect results, but keep trying. When your prompts get better, then answers will get better too – and can make your work easier.