Artificial intelligence as a door opener? Expectation vs. reality

At present, artificial intelligences still work relatively often in secret and therefore unnoticed by many users, for example in the control of robotic arms in the automotive industry, in medical assistance systems, in toy robots or in the cloud of AFI Solutions in training with invoice documents.

But they have not been or are not successful everywhere, according to an IDC study. Failed experiments are frequent and exist across all industries as well as targets. People also play a crucial role.

Artificial intelligence is not smart per se

It is therefore highly doubtful that AI systems or intelligent machines, as they are used today, are capable of taking over world domination. To date, they are still too dependent on their human teachers, who tell them the question, the goal or their purpose. Moreover, it is these teachers who feed them with selected data sets, strongly bound to the task, with which they are allowed to train within narrowly determined system and parameter limits.

Otherwise, if there is too much leeway for the AI, the results can only be evaluated with difficulty or not at all. When it comes to results, the human being, the data scientist or analyst specializing in data, is needed to interpret the results of the AI and make them usable. 

Looking around at AI projects, today's use cases are rather sobering, especially in the B2B environment, or at least strongly tied to a specific application area. The two AI systems developed by IBM, "Deep Blue" (1996 winner in chess against Gary Kasparov) and its further development "Watson" (2011 winner in Jeopardy against two human opponents), for example, are "intelligent" only within a certain rule world or with appropriately trained data sets.

More complex tasks are still difficult for artificial intelligences to solve. B2B projects often also lack resources and the amount of data required for comprehensive training.

AI can also do things differently

Some AI projects exist with rather disturbing results. In a communication experiment of a well-known social media giant, two AI-supported chatbots suddenly conversed in a "secret language" that the human researchers could no longer understand. To do this, the two bots developed a more efficient language from the human terms in order to negotiate more quickly.

However, AI-optimized languages are not uncommon in research. The secret language was also not the reason why the experiment was aborted. Since the AI was to interact with humans in its later field of application, it naturally had to use human language for this purpose. However, the scientists had failed to specify this crucial restriction for the two bots. 

More courage to experiment with AI

Ultimately, the example given above is an experimental set-up under laboratory conditions in which both the possibilities and limitations of intelligent systems and machines are explored. And these experiments are important because they often produce surprising results, some of which show us our own thought processes and human weaknesses.

The task for any artificial intelligence is clearly formulated: It should enable machines to initiate learning processes independently, to react adequately to new information and to perform tasks that require human-like thinking and a human set of values.

The training of AI-controlled systems plays a key role, comparable to the human learning process. This determines how well an intelligent machine performs under the given conditions - or not.

To this end, technologies such as machine and deep learning, natural language recognition, etc. are used to capture recurring patterns and rank them according to certain probabilities, in other words: AI learns. However, it is only when a correspondingly large amount of data is available with which to train that reliable paradigms can be identified in this data that can be interpreted. This is where the understanding or autonomous learning process begins, so to speak.

Who adjusted the new sales tax values?

In July 2020, there was a change regarding the existing VAT rates from 19 to 16 and from 7 to 5 percent. After these changes, companies sometimes had to take into account five different tax rates (plus the 0 percent), depending on whether the invoices were issued before or already after the adjustment. Both companies and their service providers feared new expenses in invoice processing.

In this regard, Bernd Kullen, AI expert at AFI Solutions, has noticed a surprising phenomenon in AFI Solutions' cloud service: "In the cloud-based AFI DocumentHub, documents are processed with the help of our AFI AI. For our customers, we have noticed that the VAT rates have already been recognized correctly, although the addition of the new rates has not yet been made in the recognition."Was artificial intelligence at work here? 

"Due to the fact that we have our AFI AI in use in the DocumentHub, there are candidates for many invoices that can also be applied to the recognition of invoice totals" Kullen said.

"There, no contents are stored, but geographical information - i.e. what is located where on the document. Thus, the contents of a trained value can differ or change, which led to the automatic adoption of the new VAT rates.."

One surprising result: no one told the AFI AI to do it. But via the automatic document training in AFI DocumentHub, it simply taught itself. 

Thanks to such results, artificial intelligences are given a raison d'être and at the same time provide arguments for being used by other customers or in other products and application areas. It is not for nothing that AI is considered a key technology to open doors for future developments such as self-driving cars or intelligent surgical robots.