AI initiatives fail more often than people think

When we talk about artificial intelligence in general, we usually adopt one of two perspectives, says ERGO CDO Mark Klein in his current blog on //next: Either we are spellbound by the hype surrounding the technology and welcome innovations that make our lives easier, or – the other perspective – we are sceptical. We fear an artificial appropriation, an unwanted interference of AI in things that are better left to us humans. However, both perspectives only play a very minor role for experts. They are concerned with a much more mundane question: why do many companies find it so difficult to create algorithms that really achieve something? The advancement of AI is nowhere near as great as many people think.

Data scientists have sort of become the rock stars of the digital scene. From mathematical formulas and huge volumes of data, they magically conjure highly complex algorithms that work with astonishing precision. Streaming services such as Netflix, for example, present us with individual film suggestions that we usually quite enjoy. The Google image search function also reliably matches photos to specific photo searches. AI and data scientists make all this possible – and much more besides.

Nevertheless, the “rock stars” are surprisingly often dissatisfied with their jobs and quit. At least, that’s the impression one gets from looking at expert forums and journals. Many of them experience a different reality in the workplace than they had expected. Instead of developing useful algorithms, they are often mainly concerned with preliminary work such as data searches and data preparation.

7 out of 10 companies: AI only has a minimal impact

The dissatisfaction corresponds with the findings of the Boston Consulting Group’s “Artificial Intelligence Global Executive Study and Research Report 2019”. In the report, seven out of ten companies surveyed stated that AI has had little or no impact on them to date. The 2021 edition of the same report added that AI-generated successes in individual units often did not lead to an overall organisational realisation that AI brings economic benefits. And according to IBM’s “Global AI Adoption Index 20212”, only about one-fifth of IT professionals said their company uses AI across the board.

Why is this? Why are experts dissatisfied and companies far less successful than we thought they were, even though everyone believes in the technological “superpower” of artificial intelligence? The answer to this has many dimensions. For example, some initiatives have to be abandoned because they cannot surmount the hurdles of regulation, governance and compliance. Or an outdated IT and data architecture is in the way. The basic IT requirements are often underestimated.

Data is a cultural issue

The most common reason for failure, however, is the data. Our Advanced Analytics division needs huge data sets, and not just for training the “algos”. Data is also continuously needed in daily operations so that the decisions made by the algorithm become more and more accurate.

Data is a cultural issue. As such, it is not enough to structure or label some unstructured data for machine learning. Rather, a systematic understanding of the value of data as a “raw material” is needed. As is often the case, we at ERGO also have to put good blueprints back in the drawer, because the data basis simply does not allow for machine learning. 

We are currently developing a data operating model for systematic value creation. It applies to everyone in the company who deals with data in any way. We need to understand and treat data as an asset, and collect it in such a way that added value is created from the outset. Therefore, the willingness of the specialist departments to think about innovative, data-based solutions is also crucial for success.

Efficient coupling of all subsystems involved

Moreover, an algorithm is always something individual, there are no repeatable recipes. The data science part alone is not nearly enough to “produce” artificial intelligence. Many subprocesses are interlinked. You have to think of it like a car factory. At the very end of the assembly line, the ordered, individual algorithm comes out. But before that, it has already passed through several production halls in several pre-production stages.

Modelling only accounts for a few stops on the assembly line journey. All the stations have to work together interactively, which often makes it long and complicated. Building algorithms is far from being magic, it is absolutely hard work with interlinked systems and processes. The fact that many fail has precisely to do with this: the efficient coupling of all subsystems involved. 

Why we are proud of our AI Factory

All the problems that often lead to failure are well known to us at ERGO. We too are often extremely annoyed and disillusioned and have to stop projects because the intention does not match the reality.

But about two years ago already, we developed an answer to the challenges that is getting us good reviews in the AI scene. We call our solution “AI Factory”. Factory, because it is a kind of template, a framework with which we have connected all the subprocesses and loose ends into a kind of assembly line. While each algorithm is individual, the basic process is the same for (almost) every AI use case. That makes us more efficient and ultimately faster. In the meantime, many “AI cars” have already passed through our assembly line and we are getting better and better.

With the Factory, we have an integrated end-to-end platform that is efficiently connected to the interfaces. For example, the “Data Lab” – our modelling environment – in which the algorithm is trained. In the automotive industry, this would perhaps be called the clean room. But the job also includes integrating the algorithm into the processes and rolling it out “productively”. Since the requirements for modelling and operationalisation are completely different, this is created in different, logically separate units that are, however, closely linked – just like in a real factory.

We work in the Amazon Web Services Cloud

We built our AI Factory in the Amazon Web Services (AWS) Cloud, which has several advantages. We can concentrate on our core business – the service provider takes care of server provisioning services and server stability.

The service also offers us a flexibility that we as ERGO could never guarantee ourselves. We can order the computing power we need for each day. Algorithms are highly complex mathematical entities. When we train them with sometimes millions of data sets, not only are very special programmes necessary, but also exorbitantly high computing capacity.

And the third advantage is the flexibility of the optional services. By moving towards plug & play functionality, we can link applications programmed in-house with open-source modules and AI services from AWS, and integrate them into our AI Factory. This is a working and development environment that we as ERGO could never provide ourselves.

A factory embedded in a high-security wing

However, the decisive advantage of the Factory is not so much the technology – there are also other, very decent systems. The decisive aspect is compliance. Data is sensitive. As a life and health insurer, we have an obligation to our customers to establish the highest safety standards. We have set up every step of our assembly line to ensure the highest level of data security.

Of course, this makes the development of algorithms more complex. However, we have long since made up for the lost speed due to the higher safety standards by combining all processes into an assembly line. Our experts now have so much experience in developing AI in secure environments that their expertise is also being requested by third parties. It is possible that parts of the AI Factory will be opened up to third parties as software as a service (SaaS). 

From art to engineering science

Back to AI, which often fails. Back to the seven out of ten companies that said AI has had little impact so far. I am firmly convinced that this will change. Not because everyone will follow the example of our AI Factory, but because artificial intelligence makes us more efficient, faster and more customer-friendly. The potential of using AI is already huge today and will definitely continue to increase.

Companies will overcome their teething problems. There will be a professionalisation regarding in particular the systematically connected and secure use of data and data models in companies. From being the “art” of “rock stars”, artificial intelligence will evolve to an engineering discipline.

When that happens, I believe the mystification of artificial intelligence will also diminish. The more we engage with it, the less we will fear its omnipotence. An algorithm is mathematics. Not simple, but rather highly complex, it needs a lot of computing power to train. And it does not know that it is showing us a car when we search for “cars” in a Google image search. It is and will remain mathematics.