Three Reflections on Machine Learning | Product Discussion Series
-
According to data from Crunchbase, a startup database, over 5,000 startups designed products and services based on machine learning in 2018. Just a year later, this number had risen to nearly 9,000.
AI is considered the fourth industrial revolution.
A recent PwC report predicts that AI will contribute $15.7 trillion to global GDP by 2030, which is why we regard AI as the biggest commercial opportunity in today's rapidly evolving economy. Moreover, it will not only significantly impact our economy but also other aspects of our lives.
So, what does the development of AI mean for us as product managers?
First, as business owners recognize the impact of AI and integrate it into key business processes, understanding at least some AI fundamentals will become increasingly important, even for those not directly working with AI products.
Second, a great product manager is often also a builder of stronger teams. We should start exploring how AI can create opportunities for our products.
With this in mind, I’d like to share a few reflections from my own journey of learning about AI and ML.
1. Understanding the Problems We Aim to Solve with ML
Every product development process begins with identifying the right problem to solve: Users don’t buy a drill for the drill itself, nor for the holes it makes, but to hang up the beautiful decorations they’ve purchased.
Introducing machine learning can indeed make our product solutions more innovative, but we must never forget to ask ourselves why we need ML in the first place.
Just as many people rush to buy the latest iPhone more for hype and trend-following than for its actual new features, ML should serve a clear purpose.
From my perspective, ML can help address the following types of problems: Can we make the user experience more customized and personalized?
Imagine visiting a café where the barista knows your name and preferences, and your favorite music is playing—versus another café where you receive the same generic service as everyone else. Undoubtedly, we’d prefer the first one.
Historically, we’ve built products for the masses, but as ML matures, we can envision a world where personalization is achievable at scale. For example, Toutiao uses ML to algorithmically distribute content tailored to users’ interests.
1. Can We Make the User Experience Safer?
Classic examples include spam filtering and banks using anomaly detection to flag suspicious transactions or fake accounts.
With ML, we can analyze and process vast amounts of data far beyond human capacity, making the user experience safer.
2. Can We Help Users Achieve Their Goals More Easily or Quickly?
This is another common application. For instance, email auto-completion speeds up drafting messages. If you buy a product, the system might recommend commonly paired items, helping you complete your purchase more efficiently.
3. Can We Create Previously Impossible Experiences?
For example, the WHO reports over 36 million blind people and 217 million with mild to severe vision impairments globally, many of whom struggle with image-based online social activities.
Facebook addresses this with a feature that uses AI to describe images in text, enabling visually impaired users to participate in discussions.
2. Evaluating Whether ML Is the Best Solution
A startup helping hotels communicate with guests via tablets introduced an ML-powered chatbot to quickly answer common questions, reducing front-desk workload. They found that 85% of questions fell into three categories:
- What’s the checkout time?
- When is breakfast served?
- What’s the Wi-Fi password?
For these, a simple feature sufficed. However, the remaining 15% of questions—like "Can I use an iron in the hotel?"—were too rare for ML to handle effectively. Here, human intervention remained the better solution.
Additionally, ML requires time, quality data, and iterative refinement—sometimes taking a year or more. Until then, it may not be the best approach. For instance, Instagram initially ranked popular posts by overall engagement due to limited data. Only after accumulating sufficient user behavior data did they introduce ML for personalized recommendations.
3. Setting the Right Expectations
At first glance, ML product development resembles conventional processes: identify problems, explore opportunities, assess risks, measure outcomes, and iterate. However, the devil is in the details.
ML involves multiple variables, often lacks universal solutions, and rarely succeeds on the first try. Thus, one of the most critical tasks for an ML product manager is setting the right expectations. ML development is a marathon, not a sprint—requiring persistence, exploration, and rigor.
For some, ML is a mathematical challenge; for me, it’s a behavioral one. Understanding human behavior, emotions, and decision-making is never simple and demands prolonged observation. The same holds for ML, but the reward is a unique product perspective—one that leverages data to inform design and enhance user-product relationships.