Building AI Products: My 3 steps.
(para a versão em português, por favor clique aqui)
Hi!
Here I want to share a little bit about my experience developing AI Products. I tried to resume it on a framework. My objective here is to share some thoughts and perceptions, and see if othr AI Product Managers face the same problems, and use the same or distinct solutions. Any suggestion, comment or feedback will be so welcome, and certainly will help me a lot! Thanks in advice!
So, why a new framework?
Because AI is a distinct kind of product.
Why is it so distinct?
Because involves concepts that are hard to communicate, and is expensive to develop.
Why AI concepts are hard to communicate?
Math. Statistics. Linear Algebra. Calculus. Forget about the movies, that is what AI is all about. Behind all the glamour and the fancy companies, you have a lot of numbers and complex concepts trying to emulate the most complex system on earth, the human brain. Unless you are a neuroscientist, you will face some hard times explaining why you think the way you think, how the neurons interact and so one. The same with AI.
And, as Product Manager is all about building bridges between people, areas, is all about communicate with people of distinct skills, roles and backgrounds… The way of an AI PM is a little bit different from a PM.
Why is it expensive?
I will highlight 3 main factors:
· People — Opportunity Cost. An AI team is rare and expensive. We are on a worldwide shortage of Data Scientists. Therefore, if you have a team, you will expend their time in the best projects. If you expend on a project with low return potential, or so high risk, you may be losing the opportunity of get more money to the company in other projects.
· Data — To get data is not easy. You need to generate or adapt systems in order to get data, clear the data, analyze it, exclude any bias… a lot of work. And, last but not least, you need to get computational power to deal with all that data, and it costs money.
· Time — Build a model, train, test, put in production, and analyze the results. It costs time. Moreover, is hard to make a MVP. Lean is a little bit different here, you can expend a little more time on validating hypothesis, and this directly affects the cost, especially on opportunity cost.
Distinct, hard to communicate, and expensive. This affects the way we deal with those products, and so the strategies we use in order to develop them.
Let’s start.
My three steps are:
· Understand the needs
· Identify opportunities
· Define the roadmap
And here I will try to describe it a little bit better:
Understand the needs
Is important to say that my experience is about work on tech companies, close to become unicorns or decacorns. This can drive a little bit the problems I faced.
Here are the three main needs I tried to understand before (or while) developing an AI Product:
· Process
Remember about the opportunity cost? Before test your bright new Ideas you probably will find big opportunities on existing process, that already tested the Product Market Fit. So, how to deal with it?
All starts with people. Understand the people in charge of those processes. Why do they do what they do. Their pains, what can generate gain. Do interviews. Build empathy, with people and process. In this part, I took some insights from Design Thinking, more precisely from Stanford Bootcamp Bootleg. I will share the link in the end of the article. Understanding the process is the begin, not the end.
· Company
If a process already exists, probably is because of a company need. Is important, while mapping the process, to identify why that process exists. Is possible that your AI product replace the process, so a lot of the mapped pains and gains will simply disappear. However, in order to do this, you will need to generate more value, better satisfy the company needs.
· Market
Your product must increase the company´s product market fit. You need to generate value for the company, considering the market. This can be translated in a lot of decisions, starting from build a product from zero in-house or buy/adapt an existing one, to invest in the best process, those who will generate the biggest impact on the companies market fit and strategy.
Identify the opportunities
Here, I adapt the Opportunity Solution Tree, developed by Teresa Torres, to the needs I faced. In the end of the article, I will share a link to her website.
· Need
Here is important to model the need in a common language between stakeholders and Data scientists. In other words, is important to translate the problem you mapped in a way that can be understood and generate insights, statistical insights. Modelling is an art, but do not forget about the main idea: Translate the need in a common language.
Tip: A thing that helped me a lot about is was… mockup. A poor and simple mockup. Maybe using a spreadsheet to emulate a concept, or sample images to emulate the results of a visual system. The point is to communicate, and be shure that everybody is talking the same language, facing the same need.
· Opportunity
With the needs properly mapped and shared, you can start to think (and conduct brainstorms about) possible opportunities to address those needs, generating gains and reducing pains. Here is extremely important that Data Scientists deeply understand the needs, as they will be the main source of insights about opportunities that AI can address.
· Solution
Here we start to draw solutions that can address those opportunities. Usually we can start to search for benchmarks or theories that addressed those opportunities in distinct markets, or research about the theme.
Define the roadmap
Here is where the project start to become so serious. We are talking about investment, return and risk. I based my decisions mainly on concepts that I have learned on Stanford Strategic Decisions Course, I will share the link in the end. In order to measure if the project is viable, you need to analyze the investment (highlighting opportunity cost) and the expected return. As usually AI products involve new solutions, it represent a lot of risk, so you need to face the return on risk.
Usually, in order to reduce the total project risk, I use to divide it on dependent steps, if possible with delivered value in each step.
So, first you will generate value to the business at each step, and even if the full product fail to be released, some steps or parts should be used.
And at each step we will validate hypothesis, so if any hypothesis fail we stop, get the learned lessons and drive the workforce to another project. This reduces risk too.
Well, this is an overview about what I´m doing J
I want to thank you very, very, very, much for read until here, and will be extremely grateful for any comments, suggestions or feedbacks.
References:
Stanford Bootleg — Design Thinking: https://dschool.stanford.edu/resources/design-thinking-bootleg
Opportunity Solution Trees — Teresa Torres
Stanford Strategic Decision and Risk Management
https://online.stanford.edu/strategic-decision-and-risk-management
