How to Design and Manage AI Products?

baoshi.rao

What are AI and ML? Why is AI product management more challenging than general software?

In the article "AI Development Guide: What is a Machine Learning Product?", we discussed the fundamental understanding and challenges of managing AI products.

For product managers (PMs), AI or ML (Machine Learning) product management is more challenging than general software because it involves greater uncertainty. It requires not only technical changes but also organizational adjustments.

For example, if you want to teach a machine to recognize cats:

Through software engineering, you might list rules like "a cat has four legs and two pointy ears." The more explicit and complete the rules, the better, as the machine relies on them to make judgments.
In contrast, with deep learning, you don’t provide explicit rules. Instead, you feed the machine a set of labeled photos (indicating which are cats and which are not), then build an ML model or neural network to let the machine learn and derive the rules on its own.

(source: IBM Research Blog)

Your team’s tasks include: defining the problem, preparing data, building the ML model, iteratively testing and refining until you achieve a model that delivers the desired results.

Given that ML product development requires more trial and error, as a PM, you need to give engineers and data scientists more space and time to explore.

But how can you help your team navigate uncertainty? How can you balance clearly defining problems and success metrics while allowing flexibility for experimentation?

Here are key considerations:

ML is a tool, not a goal: If the problem doesn’t require ML, don’t use it.
Start with the problem: Identify a market need and a technically feasible customer pain point. Conduct market research to validate demand, then determine if ML can solve the problem.
ML is best suited for decision-making or prediction tasks.

Avoid using ML in your product when:

The problem can be solved with simpler methods.
Data quality or quantity is insufficient.

After identifying the right problem, the next step is to define requirements clearly.

ML product development is iterative. Skipping planning and problem definition may waste time without yielding results.

Unlike traditional software engineering, ML is experimental and uncertain, making it hard to predict what works. Thus, engineers and data scientists need exploration time.

As a PM, help your team stay focused during exploration by:

Defining the expected outcome (e.g., what is the model predicting?).
Identifying ground truth for accuracy validation (e.g., comparing weather predictions to actual data).
Setting success metrics early and testing frequently.

Treat the ML model as a black box: define inputs and outputs, but don’t assume you understand the inner workings. Start with simple prototypes, validate core functionality, and avoid overcomplicating solutions prematurely.

Model accuracy alone isn’t enough. Consider precision (true positives/all positive predictions) and recall (true positives/all actual positives). Trade-offs depend on user needs.

Data Strategy:

ML models require high-quality data. Plan data acquisition early (e.g., purchasing, partnerships, customer collection).
Consider competition, regulations (e.g., GDPR), and costs.
PMs, not just data scientists, must drive data strategy.

For startups: Avoid markets dominated by giants (e.g., e-commerce vs. Amazon). Find niches with untapped data opportunities.

Compliance: Ensure data collection aligns with privacy laws (e.g., GDPR). Involve legal and operations teams early.

Real-world challenges (e.g., robotics, self-driving cars): Leverage simulation, transfer learning, and meta-learning to reduce data needs.

Cross-functional collaboration:

ML products require UI/UX, software, and hardware integration.
Build teams with ML engineers, data scientists, designers, and backend engineers.
Adapt workflows (e.g., daily stand-ups may not suit ML teams).

ML development isn’t just technical—it demands organizational change. As a PM, bridge gaps between teams and resolve conflicts.

Communication between PMs and internal teams as well as clients is crucial. The performance of ML products improves over time.

However, this also means that clients may not receive the best results initially—can users accept this? How can we mitigate user risks and ensure acceptable minimum performance? How should products be designed to reduce uncertainty and provide the best user experience?

ML can be used to make products more personalized or customized; for example, making it easier for users to find the most relevant results or applying ML to improve prediction accuracy. These are considerations for potential applications to meet the needs of internal or external users (clients).

Do colleagues or clients need to repeat processes that could be automated? Automating repetitive tasks can save time, costs, and resources, even creating a better user experience. If a process is too complex, can parts of it be automated or can users be assisted in completing tasks more efficiently? Gmail's "Smart Compose" is a great example: now, Gmail can automatically complete sentences for users, eliminating the need to manually type repetitive words or phrases like "Hello" every time.

Are there new business opportunities or user problems that were previously unsolvable but can now be addressed with ML? For instance, in warehouses, sorting goods typically requires manual labor because it's difficult to program robotic arms to recognize and handle millions of products. But with ML, robots can learn to identify various objects with minimal human assistance. This capability of ML and AI opens up vast business opportunities for warehouse robotics.

Sometimes, the methods used to manage software products may not be suitable for ML products.

I often remind myself of the following points:

Regarding the management of AI products, I believe the most important aspects include: