A great AI product begins with a strategy for AI products. This step-by-step guide will help you plan and execute complex development projects.

Authors are experts in their field and only write about topics where they have experience. Toptal experts who are also in the field review and validate all of our content.

AI-enabled product is everywhere. Before embracing AI, however, companies need to consider whether it makes sense for their products. AI is expensive, as it requires constant iteration along with ongoing investments in infrastructure and specialists. An AI-based product, in short, is never “done.”

What is the right AI for your product?

The projects that will benefit the most from AI are those with a lot of data and a goal to solve a complicated problem. Your team should ask these questions before moving forward. The answer to each question must be “yes.”

Exists the data required? Data is needed to build machine learning models. Ideal data would be representative of the real world and should perform consistently throughout the testing and development phases. A weather prediction model that is trained using data from the summer months will not be able to forecast a snowstorm. You can make better predictions with comprehensive data. These data must be secure, accessible, and adhere to privacy laws. Customers’ data should not contain sensitive information, such as Social Security numbers or credit card numbers, that could lead to legal issues later.

Is the product you are trying to create complex in nature? You can solve the issue by simply coding a dozen rules. Save time and money. AI products only make sense if no other method can solve the problem.

Is the problem changing over time? Hold off on AI solutions if your problem is static or slow-moving. You may only need a rule-based algorithm or statistical analysis. AI can be useful if the problem changes in real-time and requires that variables, parameters, and data responses change. For example, predicting the price of commodities is an excellent AI use case because prices fluctuate.

Can the solution tolerate imperfect outcomes? AI solutions are flawed because they rely heavily on probabilities. Even after years of optimization, no model can be 100% accurate. Choose another method of problem-solving if the users are looking for 100% accuracy.

Does the solution need exponential scaling? If you want your solution to generate exponential data and scale quickly, then AI is a great choice. Imagine a tool that calculates the freshness level of apples in an online grocery store based on their harvest date, their location, and their transit time. This system could work without AI for thousands of daily orders, but data points will increase exponentially as the tool grows in popularity or is expanded to include other fruit. AI could be the solution to this problem.

You can create a product vision if you have extensive data from real-world training and your problem requires an AI solution.

Define your product vision

The vision of the product is what drives the creation and the direction in which the product will go. This common purpose strengthens teamwork and improves collaboration.

Ask yourself how the world would be different if you succeeded with your product. A compelling answer to this question can inspire your team and customers for many years.

Google’s product vision statement for 2023, for example, reads: “Our mission is to organize the world’s data and make it accessible to everyone.” This is concise, and motivating, and will keep Google staff at all levels on track as they launch new products and refine those already available.

Plan a Product Strategy

Only worry about the specifics of your AI solution once you have defined the overall product. The goal at this stage is to determine which problems the product will solve and who it will be used by. To achieve this, we use an Agile Product Management strategy based on the Lean Startup methodology.

Lean Startup is a combination of Agile principles and an emphasis on cultivating clients. The “build-measure-learn” loop is at the core of Lean Startup. It is a process where every new product (build) goes through user testing (measure), which leads to new insights.

Every product development stage incorporates a loop of build-measure-learn.

To ensure continuous improvement, this loop is repeated throughout the discovery, verification, and scaling phases of your strategic plan. Each of these three phases builds on the previous one. After you have completed these three stages, you will have a better understanding of your customer, your market, and the growth trajectory of the product.

Discovery Stage

During the discovery phase, you will use research to identify and prioritize problems. You’ll also create hypotheses for solving them. Discovery is the time when you identify customer segments and use cases. These elements will be used to create a statement about each minimal viable product (MVP).

The MVP statement must include the user, the pain point, the solution hypothesis, and the metric used to measure the MVP results. Customer feedback can be used to start the build-measure-learn cycle. Adjust your MVP statements once you have two or more promising leads.

Imagine that an airline hired you to fix stagnant year-over-year sales on a particular route. Here are three possible MVP statements:

1. Providing concierge services to senior citizens can increase sales YoY for a particular route by 5%.
2. Online sales will grow by 5% YoY if business users are given 20% more mileage.
3. Families will see a 5% increase in sales YoY if they receive free luggage up to 20 pounds.

These statements will be refined further during the validation phase.

Validation Stage

Minimum Viable Tests (MVT) are used to validate MVP hypotheses. An MVT measures customer interaction with an MVP prototype to confirm or discredit the core assumptions. This will prevent you from investing in faulty ideas.

Prioritize the MVPs based on which products are most feasible to construct, desired by customers, and viable in terms of growth and revenue potential.

Create prototypes for customer interaction and collect data about one or two important metrics. Use the least amount of functionality. If, for example, the MVP statement assumes that seniors will pay more money for concierge services than younger people, then a simple landing page or chatbot could provide enough information to prove or disprove this hypothesis.

The MVT cycle is a build, measure, learn process in which you quickly build something, then test it with real users to see what the results are. You can also use this information to determine the type of product you need to develop.

Scaling Stage

Once the MVP statements have met your minimum viable test standard, scaling begins. we divide scaling into three Customer Development Activities: Get, Keep, and Grow. You will choose the activities that you want to focus on based on the size of the company, its longevity, and the strategic purpose of the product.

A startup’s main product may require Customer Acquisition. This could include optimizing the pricing, adding features, and expanding the team responsible for product development. The product might have the purpose of increasing the lifetime value of existing customers. This could be achieved by cross-selling and upselling.

Imagine that the validation of a concierge AI bot for older customers was successful. In the scaling phase, you will use the build, measure, and learn loop to identify new features, explore revenue models, and evaluate the structure and growth of your team. The AI chatbot hypothesis grows into a comprehensive plan as you iterate.

Lean startups encourage early and frequent feedback from customers and incremental development.

As you scale the MVP, it is important to have clear measures of success for each iteration. You should be able to measure the success of each iteration as you scale up your MVP. Concrete goals will ensure that changes are aligned with customer needs and the vision of the product.

After you have an MVP concept that is well-positioned and a business strategy, it’s time to start planning the technical requirements of your product with an AI plan.

Plan your AI strategy for your MVP

A strategy for AI can be used to assess the technical feasibility of your product after you have defined its vision and selected an MVP. A strategy for AI identifies what problem AI is supposed to solve. It takes into account the unique data and operating environment and ensures constant and seamless iteration amongst the technology team.

AI strategies can be broken down into four simple steps.

With a dedicated AI strategy, you can be sure that your AI use is justified and that you have the necessary data, infrastructure, and personnel in place to implement it.

AI Problem Definition

Your problem statement should be as precise as possible. Your team will use it to access and identify the required data, choose features, and select the most appropriate learning algorithm. A good problem statement should answer the following:

• What is the problem you are trying to solve using AI, and for whom? You must first identify the customers you are targeting to boost sales of flight routes before you can start working on a solution.
• What is your measurable goal with AI? You might want to increase your route sales by 5% in six months.
• What are the use cases that will impact this goal? You might want to rethink your target audience if you’ve seen historical purchase behavior for a route, based on holiday, school break, or business travel.

Choose a Data Strategy

I have mentioned in Part 1 that AI requires vast amounts of data for training to identify patterns and determine the next course of action based on these patterns. To that end, the AI product team should devote more than half its effort to data processing.

Answer the following questions to build your data strategy:

• What data are you missing and what data do you have? You may have historical data on inbound and outbound flights, bookings, and customer information. You’ll need to have data for every season in every year of the development set if you want to build an accurate model. Let’s assume that data for the last year from October to December is missing. This leads us to the next question.
• Where are the missing data located? The missing data is likely to be scattered across different departments or organizations. A departmental division of labor may mean that the sales team is responsible for missing data in one region while the operations group is responsible for data in another. You might also need to access regional data from different airlines.
• How can you access missing data? To obtain data from another business unit or organization, you need to plan what questions to ask and who to ask. You also need to know how to share data.
• How can you eliminate irrelevant data? Plan for the time it will take your data engineers to organize and weed out extraneous data. If, for example, another business unit sends flight sales data to you, the data may contain information about crew and passengers or data that is unclear.

Create a Tech and Infrastructure Strategy

You’ll also need to ensure that your team and customers can all access your product.

Here are some suggestions to help you with your infrastructure strategy.

• Do all members of the product team have access to the secure data? What will happen when the solution is launched? Your AI model could be built in a secure test environment within your organization. The customer or team members may reside in a different country. Then, the data would have to be hosted in a cloud-based environment.
• How do you plan to scale AI workloads once the initial infrastructure has been established? AI workloads require massive computations and huge amounts of data. If you were building an AI model for the airline product using only a few million test records, scaling would require storing and analyzing tens of millions of records. As your use case grows, you will need more data storage space and computational power.
• Can workloads be deployed across core, edge, and endpoint deployments? The modeling team needs regular access to data. Your customer may want to use a mobile application that is not part of your network. Your infrastructure must be portable across different environments.
• AI workloads demand large computing resources. A model built on test data with millions of records could take three minutes to process but would be much slower when faced with tens or hundreds of millions.

Create a Skills-based and Organizational strategy

You’ll need an organized and skilled team to build a successful product. These prompts will help you ensure that you have all the resources needed:

• Are you using the right team members? AI product teams need domain, data science, and machine learning experts, as well as product design specialists. In the next section, I’ll explain the functions of each role.
• How will you find the required personnel if not? Are you going to hire or assemble a team from your internal pool of candidates? (For the record we are a big fan of bringing in outside experts.
• What vertical business will be the home of the solution? Let’s say a sales channel in charge of the eastern United States funds and initiates the concierge AI chatbot. The airline has a successful product and wants to scale it throughout the company. Should the technology team of the company take on the scaling and maintenance costs for the product, or should all sales channels? This could require a lot more meetings.

By assigning responsibility for AI solutions at the beginning of a project, you can reduce bureaucratic chaos and ensure that your product will grow smoothly.

The Ideal AI Product Team

A successful AI team believes in its mission and owns it. These five categories of personnel will help you create a product that your customers will love.

Domain experts These experts are subject matter experts in the industry who can help identify what problems to solve and provide feedback throughout product development.

Engineers & Architects: These technical experts collect, process, and present the data. Data engineers preprocess and transform data. Software engineers code the data into a readable form to be presented to stakeholders and clients. Infrastructure engineers make sure that the environment runs smoothly and is scalable. This role is interchangeable with DevOps engineers if you are following DevOps methodologies (and you should). Architects can help you create the components that coordinate interactions between your model and the external environment.

Product Designers: Designers translate the product’s vision into a user-friendly interface. They determine the customer’s requirements, the best way to organize the features, and the overall feel and look of the product. Product Designers work closely with Digital Product Managers to connect them with the target customers.

Data scientists and researchers: Data Scientists extract actionable data from the data to make informed business decisions. They decide which features are attributed to variables that you wish to predict, and which algorithm will be most suitable for making predictions. the data scientist will collect new information as the product develops. Researchers ensure that AI solutions are always improving and delivering consistent results. The accuracy of the ML model will fluctuate as it ingests more data. These fluctuations are accounted for by the research scientists who continuously adjust their model.

Business Representatives and Analysts: In a corporate setting, business representatives are members of the business unit that sponsors a product, like finance or marketing. They are also the link between company decision-makers and the product team. Business Analysts translate between technical experts and end-users or business representatives. A business analyst could, for example, keep a finance representative informed about how customers respond to MVP tests and how much revenue is generated by the MVP. The business analyst could work directly with marketing to determine what data is needed to target customers, and then work with the ML to collect this data.

Prepare Your Team to Scale

As you collect data or solve use cases, your team may need to be scaled. we recommend agile-based team structures such as Scrum and Kanban teams to allow efficient tracking.