How to Build AI Product Strategies and Teams

This is part 2 in a three-part series on AI digital product management. In the first installment, I discussed types of machine learning (ML), the pitfalls that arise when building AI-enabled products, and best practices for aspiring AI product managers. In this article, I outline how to plan an AI product strategy and build a team to execute it.

AI-enabled products are everywhere. However, companies should consider whether AI capabilities make sense for their product before jumping on the trend. That’s because AI is costly, requiring constant iteration and ongoing investment in specialists and infrastructure. In short, an AI product is never really “done.”

Is AI Right for Your Product?

Projects that can benefit most from AI are those that have ample and robust data available, and that aim to solve a complex problem. Before going any further, your team must ask the following questions about your product—the answer to each one should be “yes.”

Does the necessary data exist? Machine learning models require a lot of data. Ideally, this data should resemble the real world and perform consistently across the development and testing phases. For example, a weather prediction model trained on data from summer months is incapable of forecasting a snowstorm. Comprehensive data allows you to make more accurate predictions. This data must also be accessible, secure, and comply with privacy laws. Customer data, for example, should not include sensitive information such as Social Security or credit card numbers, which could invite legal problems later.

Is the problem your product is trying to solve complex? If you can solve the problem by coding a few dozen rules, then do so. You’ll save time and money. AI products are only worthwhile if other methods cannot solve the problem.

Does the problem change over time? If your problem is slow-moving or inherently static, hold off on an AI solution. Rule-based algorithms or statistical analysis may be all you need. If, however, the problem shifts in real time and needs to accommodate changing variables, parameters, and data responses, an AI solution will pay off. Predicting prices for commodities is a good AI use case, for example, because prices fluctuate.

Can the solution tolerate imperfect results? AI solutions are imperfect because they rely on probabilities. No model will be correct 100% of the time, even after years of optimization. If the product’s users require total accuracy, choose another problem-solving method.

Will the solution require exponential scaling? AI capabilities are a good choice if you expect your solution to scale fast and generate exponential data. Imagine a tool that calculates the freshness of an online grocery store’s apples based on harvest date, location, and transit times. Without AI, this system might work for thousands of orders daily, but the data points would increase exponentially if the tool becomes popular or expands to include other fruits. This problem would be a good candidate for an AI solution.

If you have access to extensive real-world training data and your problem warrants an AI solution, you’re ready to create the product vision.

Define the Product Vision

The product vision is the reason for creating the product and acts as the product’s true north. This common purpose improves collaboration across the team and strengthens resilience in challenging moments.

To create a product vision, ask how the world will be better if your product succeeds. This question might seem romantic, but a compelling answer can inspire you, your team, and your customers for years.

For example, Google’s 2023 product vision statement reads, “Our mission is to organize the world’s information and make it universally accessible and useful.” It is concise, clear, and motivating and will keep Google employees at all levels aligned as they introduce new products and refine existing ones.

Plan a Product Strategy

Don’t worry about the particulars of the AI solution yet—those come after you define the overarching product. At this point, the goal is to identify which problems the product should solve and who will use it. I use an Agile product management strategy that draws from the Lean startup methodology to accomplish this.

Lean startup combines Agile principles with an emphasis on cultivating customers. The “build-measure-learn” loop is at the core of Lean startup. It describes a process in which every new development (build) undergoes user testing (measure), leading to new insights (learn).

This loop repeats throughout the discovery, validation, and scaling stages of your product strategy planning to ensure continuous improvement. Each of these three stages builds on the preceding one. Once you’ve completed them, you should have a sense of the customer, the market, and the product’s growth trajectory.

Discovery Stage

In the discovery phase of the product strategy, you’ll use research to define and prioritize problems and create hypotheses to solve them. Discovery is also the time to identify customer segments, use cases, and business models. You’ll use these elements to write a statement for each minimum viable product (MVP).

The MVP statement should capture the user, pain point, solution hypothesis, and a metric to measure the MVP’s results. Use customer feedback to initiate the build-measure-learn loop, and adjust your MVP statements until you have two or three promising leads.

Suppose an airline has hired you to address stagnating year-over-year (YoY) sales for a specific route. Here are three potential MVP statements:

1. Providing concierge services for senior citizens will increase YoY sales for a specific route by 5%.
2. Enabling 20% more mileage points to business users will increase online YoY sales by 5%.
3. Offering free checked luggage up to 20 pounds will increase YoY sales to families by 5%.

You’ll refine these statements further in the validation phase of planning.

Validation Stage

The validation phase uses minimum viable tests (MVT) to determine the viability of an MVP hypothesis. An MVT confirms or discredits the core assumptions of the hypothesis by measuring customer interaction with an MVP prototype. This process will save you from overinvesting in faulty concepts.

Begin by prioritizing MVPs according to which product is most feasible to build, desirable to customers, and viable as determined by growth and revenue potential.

Next, create prototypes to enable customer interactions and to collect data on one or two important metrics. Do this using the lowest degree of functionality possible. For instance, if the MVP statement’s core assumption is that senior citizens will pay more for concierge services, a landing page about this feature or a rudimentary chatbot would likely provide enough data to validate or disprove the hypothesis.

This MVT process constitutes a build-measure-learn cycle in which you build something fast, measure the results with actual users, and learn more about the product you should develop.

Scaling Stage

Scaling begins once the MVP statements meet your minimum viable test standards. I break scaling into three customer development activities: get, keep, and grow. The activities you focus on will depend on the company’s size and longevity, as well as the product’s strategic purpose.

For example, a startup’s core product will require customer acquisition, which could entail optimizing the pricing model, adding features, and expanding the product development team. In an established company, the product’s purpose might be to grow the lifetime value of existing customers, which might entail cross-selling or upselling.

In our airline product example, imagine that a concierge AI chatbot for older customers succeeded during validation. In the scaling phase, you would use the build-measure-learn loop to identify new features (which would then cycle through the discovery, validation, and scaling process), explore revenue models, and evaluate how to structure and grow your team. As you iterate, the AI chatbot hypothesis will grow into a comprehensive strategy.

The goal of any product management strategy is to ensure that you don’t build the wrong product. As you scale the MVP, you should have clear measures of success for each iteration. Concrete goals ensure that all changes add value to the business and align with the product vision and customer needs.

Once you have a well-positioned MVP concept with a sound business plan, you’ll start planning for the product’s technical demands with an AI strategy.

Plan an AI Strategy for Your MVP

After defining your product vision and selecting a product MVP, assess its technical feasibility with an AI strategy. An AI strategy identifies the problem that AI must solve. It accounts for unique data and operating environments, and ensures seamless and constant iteration across the technology team.

You can break down an AI strategy into four steps:

Define the AI Problem

Be as specific as possible in your problem statement. Your team will use it to identify and access the necessary data, select features, and choose the appropriate learning algorithm. An effective problem statement will answer the following questions:

• What problem are you trying to solve with AI—and for whom? To boost flight route sales, you must identify which customers you’re targeting before starting work on a solution.
• What measurable goal do you want to achieve with the AI solution? You may want to increase route sales by 5% within six months.
• What use cases will affect that goal? Historical purchase behaviors for a given route based on holidays, school breaks, or business travel might lead you to reconsider your target audience for those six months.

Choose a Data Strategy

As I mentioned in part 1 of this series, AI needs vast quantities of training data to recognize patterns and identify the next course of action based on those patterns. With that in mind, more than half of an AI product team’s effort should be devoted to data processing.

To build your data strategy, answer the following questions:

• What data is available, and what data is missing? You might have access to historical inbound and outbound flight data, booking data, and customer data. To build an accurate model, you’ll need data from every season of every year in the development set. However, let’s say that the most recent year’s data from October through December is missing, which leads us to the next question.
• Where does the missing data reside? Missing data will likely be strewn across departments or organizations. Due to departmental siloing, the sales team may own the missing data for one region, while the operations team owns the data for another region. Or you might need access to regional data from multiple airlines.
• How do you get access to missing data? Obtaining data from a different business unit or organization requires planning what to ask, whom to ask, and how to share the data.
• How do you identify and discard irrelevant data? A data dump often contains extraneous information, so plan for the time your data engineers will need to vet and organize it. For instance, if another business unit sends you flight sales data, it may be interspersed with information about passengers and crew or with data whose meaning is unclear.

Create a Tech and Infrastructure Strategy

After obtaining the data, you’ll need the right tools and structures to process it, run the models, build the AI services, and ensure that everyone, from your internal team to your customer, can access the product.

Here are some prompts to guide your infrastructure strategy:

• Will all product team members have access to secure data? How will the customer access the data once the solution launches? You might build your AI model in a test environment accessible from a secure connection within an organization. However, the customer or other team members might reside in different countries. In that case, you would need to host the data in a cloud environment.
• How will you support scaling once the initial infrastructure is in place? AI workloads need massive-scale computation and enormous amounts of data. In the case of the airline product, you might build your AI model using a few million test data records, but scaling would mean storing and processing tens of millions of records. Ensure that you have enough data storage and enough computational power to process the new data streams as your use case scales.
• Are workloads portable across core, edge, and endpoint deployments? The modeling team will need regular access to the data. The customer might want to access the solution through a mobile app, which may not be part of your network. Your infrastructure should be able to support portability across environments.
• How do you balance information quantity and computing power to ensure timely access for all stakeholders? AI workloads require large-scale computing resources. The model you build on millions of test data records might have a processing time of three minutes but would take much longer when confronted with tens of millions of records.

Build a Skills and Organizational Strategy

To build a great product, you’ll need a skilled and cohesive team, and strong organizational support. Use these prompts to ensure you have the resources you need:

• Do you have the appropriate team composition? AI product teams require domain, data science, machine learning, and product design experts. I’ll discuss the function of each role in the next section.
• If not, how will you obtain the necessary personnel? Will you assemble teams from an internal candidate pool or hire them? (For the record, I am a great proponent of engaging outside experts.)
• Which business vertical will house the solution? Say a sales channel responsible for the eastern United States initiates and funds the concierge AI chatbot. The product is successful, and the airline wants to scale it across the organization. Should all sales channels share the product’s scaling and maintenance costs, or should the company’s technology team take over? It could take a lot of meetings to figure this out.

Assigning responsibility for the AI solution at the start of the project will reduce bureaucratic discord and ensure that the product grows seamlessly.

The Ideal AI Product Team

A successful AI product team believes in the mission and takes ownership of its success. These five personnel categories will ensure you build a high-quality product your customers love.

Domain experts: These are industry subject matter experts who help determine what problem is worth solving and offer feedback on the product’s utility throughout its development.

Engineers and architects: This category of technical experts collects, processes, and presents the data. Data engineers wrangle, preprocess, and transform the data. Software engineers then code it into a readable format to present to stakeholders and customers. Infrastructure engineers ensure that the environment is up, running, and scalable. If you follow DevOps methodology (and you should), this role can be interchangeable with a DevOps engineer. Architects will help you design the various components that coordinate the interactions between the model and the external environment.

Product designers: Designers transform the product’s vision into a customer-facing interface. They are responsible for determining the customer’s needs, how to organize features, and the product’s overall look and feel. Product designers work closely with digital product managers and connect them to the target customers.

Data and research scientists: Data scientists extract actionable information from the data to make informed business decisions. They finalize which features get attributed to the variables you want to predict and which algorithm is best suited for the predictions. As the product grows, data scientists will gather new information for predictions. Research scientists ensure that the AI solution’s results are consistent and always improving. As the ML model ingests larger quantities of varied data, its accuracy will fluctuate. The research scientists continuously adjust the model to account for these fluctuations.

Business representatives and analysts: In an organizational setting, business representatives will be members of the business unit, such as finance or marketing, that sponsors the product. They also link company decision-makers to the product team. Business analysts act as translators between technical experts and business representatives or end users. For example, a business analyst might keep a representative from the finance team apprised of how customers react to MVP tests or how much revenue the MVP generates. Or the business analyst might work directly with the marketing team to understand what data it needs to target customers and work with the ML team to collect that data.

Prepare to Scale Your Team

You may need to scale your team as you accumulate data or use cases to solve. I recommend Agile-based team structures, such as Scrum or Kanban teams, to enable efficient tracking and scaling. In part 3 of this AI product series, I’ll offer a tutorial on implementing your strategy within an Agile framework, including how to run sprints on a cross-functional AI product team.

Want in-depth product management guidance? Mayank’s book, The Art of Building Great Products, offers step-by-step instructions for digital product managers and entrepreneurs looking to turn ideas into successful products.