In the last few years AI invaded our life in many ways through many products and made a real revolution in some domains. Siri, Alexa, navigation apps, detecting medical issues are just a really short list of products and services that made a revolution in their domains. Based on the records funding money going into AI, it is predicted to make an even more dramatic impact during this decade. In this article I will discuss the unique nature of AI-based products and its influence on the development process and the usability.
Applying AI to products is tricky as machine learning and building models have some unique characteristics. These characteristics have an influence on the product users and on the newly formed relationships between product managers, data engineers and data scientists. Following are the unique characteristics of AI products:
1. AI is unpredictable
We gave computers and machines the power to make predictions and decisions based on data patterns. Now we expect them to behave like machines or humans?
Do you remember the last time you visited 3 doctors and received 3 different diagnoses and ways to treat your problem based on the same medical documents?
Unlike classic software engineering where you insert an input and you expect a certain output, machine learning products are probabilistic, and you may get one answer now and the opposite one in another time. Be aware of this.
One thing is predictable — consult your symptoms with Google and you will find out that you have one week left to live 😊
“Anything at all is possible. Some things are unlikely. Some things will never happen. But they always could, at any time.”― Ashly Lorenzana
2. AI projects have uncertain outcomes
When building a machine learning model there are many unknowns and uncertainties during the process. What is it like to manage a product in an uncertain environment? As a product manager you are used to define KPIs and deliverables, but for such cases how do you define them?
- What should be the KPIs for the data scientists’ teams?
- Should we be using agile methods?
- Are they part of the daily scrum?
- How do we define deliverables for these teams?
- When should we start getting deliverables?
All are valid questions and you should be able to answer them and define milestones together with the team.
“We demand rigidly defined areas of doubt and uncertainty!”― Douglas Adams, The Hitchhiker’s Guide to the Galaxy
3. AI is only as good as its data set
Data is king, as we say. But if you want your king to rule (whatever that even means?!), you want it to be rich, deep, historical, continuous, multi-dimensional and other pompous superlatives. The model is as good as the data is. If the data is not rich enough, don’t be surprised that your output won’t be as expected. If you don’t have enough data, maybe you shouldn’t be trying to solve the problem with AI.
Start collecting the relevant and quality data now so that in a few months you will have enough data to start building the model.
“Fall. Stand. Learn. Adapt.”
― Mike Norton, Fighting For Redemption
When you talk to your 2-years-old child you expect a certain level of understanding and you are using simple words to convey your message. If your model have limited data, don’t expect it to act like a grown-up. It still needs to learn.
4. AI requires flexibility
“You don’t know what you don’t know” — Socrates
What Socrates said it very true to the stuff we can find in the data. When data scientists explore the data, they may come up with new ideas that will make great impact on the product. Just like that requirements can be refined and changed during the feasibility discussions with the tech teams, same and even more impactful changes can take place once the data teams review and play with the data. This stage is usually called the discovery and exploration phase.