Updated: Feb 17
I'm often being asked questions about the day-to-day activities and responsibilities of AI Product Managers. While there is no standard definition for this role, over the last decade, I've observed and experienced specific success patterns in adapting AI in technology corporations, as well as mid-sized businesses through my own startups. In this article, I want to share a summary of these success patterns.
If you want to learn more about the full range of frameworks and best practices, please join me at my upcoming Stanford Continuing Studies class at Stanford campus or online Live virtual classroom at AI Product Institute.
Fundamentals of AI Product Management
Long-time ago, when I decided to change my profession from a software architect to product manager, I asked one of my product manager peers at Yahoo, what product management really means. The answer was so simple, but to the point that I still carry it with me every day. She stood up, took a marker, and draw two small circles on each side of the whiteboard. Then she connected the circles with an arc and said something like this: "Product management is about understanding where your product is, defining where your product should be and assuring that you get there on time and under budget."
The fundamental idea behind AI Product Management, IoT Product Management, and Software Product Management is this simple, and I believe that no matter what prefix will be added, it will not change. Therefore, 101 of AI Product Management is to be an excellent product manager first.
Over the years, I've learned that there are more aspects to take care of than solely the product. Therefore, I'm extending this elegant description for future product managers by adding the team, company, and industry aspects as well as naming it the "As-is" and "To-be" states.
However, fundamental PM know-how is not enough to lead AI initiatives to success. Then, what else does a PM do in AI projects? What are AI Product Manager skills?
It all depends on the type of AI project. There are fundamentally four different types of AI initiatives:
1) AI for non-AI product/service (e.g., Netflix, Uber, Adobe)
2) AI for AI product/service (e.g., Google, Amazon, NVidia, Microsoft, IBM)
3) AI consulting (e.g., Accenture, Ernst & Young, McKinsey & Company)
4) AI in education (e.g., Coursera, Udacity, Stanford, MIT)
First of all, regardless of what initiative you work in, you need to concentrate on adding value to the customer and, therefore, the company instead of thinking which algorithm has a better relevance or accuracy. This, of course, doesn't mean that an AI Product Manager should not know anything about the technical aspects. On the contrary, technical understanding is needed. However, in practice, the data scientists and software engineers are responsible for the technical implementation, not the AI product manager.
It turns out that the further away you are from any ML algorithm code, Python script, or any software code, it is better. There are three reasons for this:
First, regardless of how successful your technical solution is, in the end, if it doesn't move the needle in business or customer value, it is useless. Basically, the success of an AI project depends on the correlation between the objective function and business metrics. I've seen highly accurate ML models that degrade user experience as well as one line of an if-then rule that improves user experience. Therefore, work on gaining a quantified understanding of your customers and business. Renée Mauborgne and W. Chan Kim's Buyer Utility Map from the Blue Ocean Strategy is an excellent framework to understand your customer's journey, your company's positioning, and even the competitive landscape. You can learn more about the buyer utility map at https://www.blueoceanstrategy.com/tools/buyer-utility-map/ .
The second reason to be on the business side rather than code is that you need a broader perspective when the time comes to consider trade-offs between an entire tech team and off-the-shelf solutions. Amazon Kendra, for instance, is a highly accurate and easy to use enterprise search service that’s powered by machine learning. $5K a month cost is a catch compared to investing millions of dollars in search engine development and hundreds of thousands of dollars to keep it running. Or take Amazon Personalize, for example. For $1K - $2K a month, you get a fully functional recommendation solution.
On the contrary, if you would try to build it in-house, it would take you one year with at least one data scientist and two software engineers to develop a personalization engine. Furthermore, you would need at least a software engineer and a data scientist to maintain it. Not counting the IT cost.
And, the third reason is that you won't have time to look at the code or even participate in technical architecture discussions. The amount of work to understand the customer, define precise requirements, analyze industry trends, define strategies, become the customer's voice, acquire project buy-in, create roadmaps, track product development lifecycle, collaborate with cross-functional teams, form partnerships, track product performance, manage experimentation and report to c-suite is enormous. I don't believe any product manager who does their job to the fullest has any time to get technical at work.
AI for non-AI product/service
If you are working on an AI project that solves some problems in a non-AI product or service, such as a movie recommendation at Netflix or the last one-mile problem at Wallmart, then you need to focus on the industry/domain side. Without knowing the domain and its dynamics, it is nearly impossible to add value to the customers and company by using AI. Product recommendations in e-commerce, for instance, has more in common with marketing and sales than an unsupervised learning algorithm. It requires knowing anything about the sales cycle, market dynamics, inventory, or pricing model for a specific industry. Without knowing these, it is even not possible to define business metrics and therefore optimize upon.
Therefore, the day-to-day activity of an AI product manager working on a non-AI product or service includes tasks such as understanding customers, market, inventory, identifying AI opportunities, defining the to-be state, creating roadmaps, translating requirements to technical teams, following through AI solution development lifecycle, managing experiments, determining go-to-market strategies and releasing products.
AI for AI product/service
If you are working on an AI product or service that is part of the AI ecosystem, such as AI platform, AI hardware accelerators, or AI frameworks, then you need to focus on the specific AI solution domain as well as any cross-cutting concerns.
The AI solution domain can be divided into four main categories:
1) AI industries, such as robotics, where AI became a mandatory component.
2) AI applications, such as computer vision, machine translation, natural language processing, or automated speech recognition, where many machine learning techniques are utilized in combination to achieve human-level performance in vision, sound, and language problems.
3) AI platforms, such as ML algorithms, ML frameworks, AutoML, or distributed system, where data scientists and ML engineer productivity is the upmost importance.
4) AI infrastructures, such as GPUs, TPUs, ASICs, supercomputers, specialized storage, and networking components, to support the heavy ML training and inferencing workload.
Unless you are working on an AI platform, like me, it is less likely that you will need to involved in every AI solution domain aspect. In the industry, AI product managers usually work on specific solutions such as AutoML or a GPU accelerator. Therefore, having a broad knowledge of the AI solution domain and a more in-depth understanding of the particular solution is necessary.
Similar to other software solutions, there are, of course, cross-cutting concerns such as responsible AI and security. These concerns need to addressed regardless of the specific domain.
Therefore, the day-to-day activity of an AI product manager working on an AI product or service includes everything I've mentioned in the non-AI product part, in addition to leading the specific AI initiative. One notable difference is that the field of AI is expanding exponentially compared to any other domain, which requires continues learning and unlearning.
I hope this gives you a high-level idea of what an AI Product Manager does. I share more details and frameworks in my on-campus Stanford Continuing Studies class and online Live virtual classroom at AI Product Institute. If you are interested, please feel free to drop me a message or join https://www.aiproductinstitute.com.