Some time ago, a zapliance customer approached Alex and me with the idea of starting an Artificial Intelligence compliance project together.
The goal of the project:
to develop a plan on how to evaluate AI use cases - a kind of best practice, so to speak, to classify whether a use case can be rated as AI or not.
For this, our client had prepared three AI use cases from their company.
In our first workshop with our customer, we first defined what artificial intelligence should actually be in the concrete operational context.
We followed the definition of the Gabler Wirtschaftslexikon (2021).
According to this, artificial intelligence is "research into "intelligent" problem-solving behavior and the creation of "intelligent" computer systems.
Artificial intelligence (AI) deals with methods that enable a computer to solve such tasks that, when solved by humans, require intelligence."
In the next step of our workshop, we presented 5 criteria that can be used to check whether a use case is really based on Artificial Intelligence or not.
These include the four criteria (Gehrke, 2019) autonomy, business relevance, learning, and the possibility of dynamic adaptation.
I have subsequently developed the fifth criterion later on: Specificity.
- autonomy: The application decides autonomously by machine or significantly supports a human decision so that there is at least a significant machine influence.
- business relevance: The decision is not merely a decision to process business transactions in a technically correct manner, but the decision has business relevance or an influence on the organization.
- learning: the way the machine makes decisions is not merely based on static rules (e.g. "if-then-else chaining"), but the decision calculus was first learned by the machine through the processing of training data.
- possibility of dynamic adaptation: the way of decision-making could be adapted regularly, as new data for an extended training of the algorithm used is continuously added.
- specificity: it is not a standard AI solution that can be used by everyone without special adaptation (e.g. Google Translator).
After we had presented the five criteria within the workshop, the examination of the existing use cases against these criteria followed together with the customers.
And what was the result?
What followed next was a surprise: because two of the three use cases failed the test - according to the five criteria, the two cases were not use cases for artificial intelligence.
In both cases, the "learning" criterion was not met.
As a reminder, the "learning" criterion states that the way a program makes decisions is not only based on statistical rules.
To be classified as Artificial Intelligence, the program must also be able to develop a decision on its own by processing training data.
The conclusion: both use cases were admittedly not what we classify as AI.
But I am sure that both the cases and their results can provide the company with valuable insights, as they are defiantly based on statistical considerations.
AI is hot - but are all applications that do not fall under AI useless?
Everyone is talking about artificial intelligence and wants to use it.
Because artificial intelligence sounds cool, has recently become a kind of must-have buzzword in the industry and makes a company look good.
But in my experience, many solutions are not based on AI at all when you check them against the five criteria introduced in this blog post.
But what does it mean if an application cannot be classified as AI?
There is nothing wrong with using statistics or complex but targeted static rule sets instead of AI if you and your team are evolving and achieving the goals you have set. Because what matters first and foremost is that the application works and moves your business forward - the categorization is irrelevant in my opinion.
So, if your use case doesn't pass the criteria test, that's definitely not a reason to throw in the towel.
If you still value developing an AI application, then don't give up.
Because you may only be a few steps away and can use our criteria to find out where you could make improvements to the application you already have.
And if someone criticizes your project:
Try not to take it personally.
Unfortunately, however, I can confirm from my own experience that this is often not so easy.
It is best to remind yourself what is really important and how you want to get there.
Educate yourself, actively talk to colleagues or experts, keep finding and trying new ways to get closer to your goal.
The most important thing:
And if you ever need another opinion, email me at firstname.lastname@example.org and we can see how we can help each other!
Gabler Wirtschaftslexikon (2021): Artificial Intelligence (AI). https://wirtschaftslexikon.gabler.de/definition/kuenstliche-intelligenz-ki-40285
Gehrke, N.: Tackling AI and Compliance - How to Approach (2019), in COMPLIANCE Digital - Wie beeinflusst die digitale Transformation das Compliance-Management; S. Behringer, A. Unruh (eds.): Berufsverband der Compliance Manager e. V., Berlin; 95- 103
About the author:
Prof. Dr. Nick Gehrke is a member of the board of the Artificial Intelligence Centre Hamburg, and a professor and course leader at the Nordakademie Hamburg, and Co-Founder and Data Scientist at zapliance. He started zapliance after meeting Alexander Rühle at PWC and set their eyes on the same goal: Changing the way business professionals of the future work with data.