xingtwittersharerefreshplay-buttonpicture as pdfLogo_DPM_SCHRIFTZUGinstagram icon blackShapeGroup 3 Copy 2Group 2 Copydepartment_productdepartment_datascienceuserclosebasic clockblogShapearrows slim right copy 3arrows slim right copy 3arrows slim right copy 3

3 Thoughts On: Successful AI

Lorenzo

Lorenzo |

23. Apr. 2020 |

- min Lesezeit

3 Thoughts On: Successful AI

A warm welcome to the new episode of our series "3 Thoughts On" – this time especially for data scientists and data-driven managers who want to implement a successful strategy around AI.

Most AI products fail. This is not a luddite statement from someone who wants to resist at any cost emerging technologies. This is a fact, which can be confirmed by reading a couple of articles out there in the Web, like this one, for example. The good news is that they do not fail because of AI in itself: following some good practices in the development phase can help to increase the chances to succeed. The common denominator between these practices is the creation of a shared knowledge between data scientists and product people that enables an area of discussion where engineers talk about users and managers have a high-level perspective about algorithms. The following three thoughts should clarify why this is important, and how it works in practice.

1. AI is industry-agnostic, data is not

You might have probably heard how great AI is, this magnificent science that can be applied to anything and will make your business numbers magically shine. You might have also heard how often data science teams fail delivering value to the businesses that hired them. How can these two stories both be true? They reflect two perspectives on AI, which if not aligned, lead to frustration and failure. The first perspective is that of the data scientist: the theory they handle is beautifully abstract and has (apparently) nothing to do with user behaviour or business performance. The managers on the other hand focus their attention on making their users happy and for them it doesn’t really matter (or at least it seems so) whether the engineering team implemented a linear regression or a deep neural network. They’re both missing the connection between their two worlds: data, a bunch of numbers that instruct the algorithms to actually solve some real problem the users face. A common understanding of data is the key: data scientists can align their strategy to the business goals only if they know the right questions to ask to the data they are provided, and these question can only come from people with a deep understanding of the product and its environment. In this way businesses can unravel information that is hidden in the abstract numbers (aka data) and use it to impact the “real world” (aka users).

2. Agile yes, but not the usual way

Agile practices revolutionized the way companies develop software, creating a framework in which product people and engineers work together in an effective and pleasant way. Implementing Kanban as it is won’t make your data team successful and happy, though: Data Science is science, its outcome is subject to high uncertainty, and moving tickets from left to right is not just a matter of time. A task may not be completed just because it doesn’t work, the same way a substance might not cure the rats in a medical experiment. On one side, data scientists need to take into account uncertainty when doing time estimates, manage the expectations of their stakeholders and document each experiment (for the team as well as for external communication). On the other side managers need to understand the risk to take upfront, in view of the high value that data science and AI can deliver when successful. Within our team we are testing a modified Kanban board for this purpose, have a look at it and start being comfortable with the word “experiment”!

3. Improved experience vs automation

It is not uncommon the idea that adoption of AI requires lots of efforts and investments. This is due to the fact that often the first examples of AI that come to mind are full-automation challenges (e.g. Autonomous driving) that indeed require full R&D teams working for years on the topic. Also, the difficulty to explain what state-of-the-art algorithms do makes it difficult to trust them taking automated decisions on delicate topics (a lot is going on in this field). AI is more than moonshots, though: it is a powerful tool to enhance human activity, and in just days you can build data-driven solutions that improve significantly your users’ experience. A deep understanding of your users’ workflow opens the door to many applications of fancy smart algorithms: you will be able to identify the little burdens of your customers, and relieve them from annoying repetitive tasks (partial automation), give them tailored suggestions, and in general optimize their experience while still granting control on the most key decisions. Automatic gear is no autonomous driving, but still it makes driving in the city a more enjoyable experience.


In short

To summarize, these little dictionary combines in three words some key elements of a successful AI strategy:

  1. Data - Know how the numbers connect to your users, in order to implement the right algorithms for the right problems.
  2. Science - Account for the uncertainty of AI, in order to make realistic plans and to work effectively as a team.
  3. Experience - Dive into your users’ product journey, in order to understand where AI can assist them (not replace them!).

If you want to read more on these topics:


Ähnliche Artikel

Ähnliche Artikel