Why companies aren’t using artificial intelligence anymore

AI is an amazing piece of technology that can defeat human masterminds, guide cars through traffic jams, and accurately translate text into other languages ​​in the quiz show Jeopardy. Some researchers compare the influence of AI with that of electricity 100 years ago. According to the management consultancy Capgemini, companies that make decisions based on data achieve 22 percent more profit and 70 percent more revenue per employee than their competitors.

But using it in business is often not as easy as technology providers and management consultants would like you to believe. In a recent study by industry representative Bitkom, managers unanimously affirm the importance of AI, but only nine percent say they actually use it. Association President Achim Berg called the result “sobering”. Internationally, the picture is not fundamentally different. IBM learned this the hard way.

Some optimistic forecasts are therefore no longer valid. Market research firm Forrester expects sales of AI software to grow from $33 billion last year to $64 billion by 2025 — that’s impressive, but not the gigantic deal some experts were expecting just a few years ago.

Of course, the major crises of recent years have had a detrimental effect. According to Bitkom President Berg, there is “little room to think about new technologies and business models for the future” in permanent crisis mode. However, there are also tangible problems with the technology. There is a lack of skilled workers and data, and sometimes also money, as the companies reported to Bitkom. IT security concerns and concerns about data protection violations weigh the heaviest.

Last but not least, there is a lack of the right ideas. “My most important work with customers is finding the right application,” says Eng Lim Goh, AI technology chief at Hewlett Packard Enterprise (HPE). Customers often have false expectations: “AI is neither super smart, as some believe, nor completely incapable of what others assume.” Only those who know what the technology can and cannot do can use it sensibly.

Problem 1: Integration into business processes

In order for the use of artificial intelligence to be worthwhile, companies must integrate the technology into their business processes. The Watson system impressed the public with quick-witted answers to quiz questions, but the doctors in the partner clinics ultimately failed to provide any answers.

DB Fernverkehr AG wants to avoid something like this. The company is setting up a “Competence Center” to promote exchange between the development department and the rest of the group. “An essential task is to make the benefits and possibilities of AI tangible for the organization,” says Axel Schulz, who heads the AI ​​systems and customer analyzes department.

Various exchange formats are intended to give employees ideas on how the technology can improve their everyday lives, whether in the factory or in on-board service. It’s not the developers who should define what’s good – it’s the colleagues who work in shifts. “If AI models achieve demonstrably good results, we can quickly integrate them into existing business processes and tools.”

Example passenger service: Customer feedback – whether by e-mail or Twitter – is automatically processed by the Deutsche Bahn subsidiary. Algorithms identify the topics and extract content, such as specific complaints or suggestions, from the texts. This works so well that selected opinions can be played out to the train attendants in just a few minutes, says Schulz. And that leads to “a much better travel experience”. The data is processed faster and more purposefully than the personnel capacities in the service allow.

This brings you to a core topic of AI – data. And at the same time with a core problem.

Problem 2: Collecting the right data

What is now referred to as artificial intelligence is a hodgepodge of different technologies. What most of them have in common: They use data as learning material, i.e. derive rules from texts, photos or vibration patterns.

The rule of thumb is: the more data, the better the AI. For example, machine learning requires at least 10,000 data points from the same process, for example from a machine and its production behavior. “Many small companies do not have such data sets,” says Katharina Zweig, AI expert and professor at the University of Kaiserslautern.

Even when there is enough data, it is often poorly available. In companies, data is distributed across numerous computer systems and is available in different storage formats – on server drives, local hard drives, USB sticks, CDs, possibly floppy disks. Everything that has accumulated over the years and decades.

“In no other company have I found the data processed as if from a single source,” says Bryan Harris, head of technology at analysis software provider SAS. Last but not least, it is important to sort out incorrect or duplicate data records. “Data preparation is the biggest part of what we do,” says Goh, HPE chief technology officer.

The data must also be complete. Some time ago, Eon planned an artificial intelligence to help the service department deal with customer issues. The idea: based on existing cases, relieve employees of some of the work through automation and find a solution for the customer.

“We were able to draw on many millions of cases, each of which was correctly classified into the appropriate problem category,” says Christian Essling, global data and analysis manager at the utility. However, information about the individual solution steps was missing. “However, since this data was absolutely necessary, we had to take a step back and adapt the data-generating process so that it also documented the steps on the way to the solution in addition to the request and the classification.” Even if the data is available and complete, it can Data protection or legal reasons make it difficult to use.

Example patient data: From them, researchers and AI models can gain valuable information on disease control. But medical histories should not be given away so easily, privacy must be protected.

Google developed a partial solution in 2017 with “Federated Learning”. The technical term of “federal learning” describes that neural networks are trained on several participating devices or “nodes”. The data records remain stored locally, which guarantees data protection.

Problem 3: Finding the right application

The selection of the data is of great importance, even small details can be decisive. For example, when analyzing social networks: If the AI ​​was trained about a month ago to recognize important geopolitical trends, it cannot react to world events that happened only two to three weeks ago.

For example, a machine trained in early February 2022 cannot know that Russia invaded Ukraine at the end of February. It is correspondingly difficult for her to classify posts on the Internet. If basic data and assumptions are not consistent, an AI can quickly come to wrong conclusions.

The danger showed up five years ago with the “Libratus” program: the AI ​​defeated four of the best poker players in the world. When the players were asked what they noticed about the way the computer was played, everyone involved answered: The program played like an “alien”, for example betting extremely large sums “like no human would do”, says Goh, who was in charge of the IT project at the time.

Such behavior can become a problem in everyday business. “Customers need to understand that AI is not always accurate,” says Sebastian Bluhm, head of IT consultancy Plan D. “It’s just a statistical approximation.”

For example, an AI can correctly recognize an image 98 percent of the time, says HPE technical director Goh. According to him, that’s more than a human can do. But there is an important difference: in the 2 percent of cases in which the AI ​​misrecognizes an image, it can behave almost absurdly compared to humans.

Goh gives an example: The AI ​​easily recognizes an apple and a glass of water. But if the apple is behind the glass of water, then some neuronal networks produce the answer “apple juice”. A gross mistake that a human would not make. And which could cost a fortune in important production decisions.

It’s all about finding the application where the AI ​​can play to its advantage. “It’s a challenge,” says Goh. One of his recommendations: “Start with the simplest application and not the one you want to impress the boss with.” And then slowly increase the level of difficulty and the range of applications, for example starting with the evaluation of the images from ten cameras, and then later to 100 or more cameras to go.

Good advice: seek help

In summary: start small and don’t expect a miracle solution. AI is a complex and exhausting thing. The US paper company Georgia-Pacific, for example, operates a total of 15,000 models to control and optimize its production and machines. Therefore, companies should seek help, especially smaller companies. “Small companies have the most problems with AI,” says Bitkom President Berg, “they often lack know-how and technical capacities”.

But according to Berg, burying your head in the sand is not an option: “Artificial intelligence is a key technology.”

More: “Seeing the immense opportunities alongside the risks”: German business fears overly strict AI regulation