[AI for Business #13] - Getting AI discovery right
A guide to ideating, validating, and prioritizing your AI use cases
Dear AI friends,
When you are building with AI, complexity adds up - there’s more uncertainty, more unknowns, and more moving parts across teams, tools, and expectations. That’s why having a solid discovery process is even more important than when you are building traditional, deterministic software.
According to recent studies, the #1 reason why AI projects fail is that companies use AI for the wrong problems. These problems can be:
too small — no one cares,
too simple — not worth the effort of using AI and dealing with more complexity,
or just fundamentally not a good fit for AI in the first place.
In this episode, I’ll share how we approach discovery for AI-driven products, breaking it down into three key steps:
I will use the example of a recent project in the automotive industry to illustrate the approach. Some of the points described will be new and specific to AI; others are known from traditional development, but gain even more meaning in the context of AI.
📘 This post is based on my new book The Art of AI Product Development.
Today, the book goes live on Amazon! 🎉
If you’re reading (or have already read) the book and would like to support my work, I’d be highly grateful if you could leave an honest and personal review. Your feedback helps others discover the book and makes a real difference.
Ideation: Finding the right AI opportunities
Let’s start with ideation—the first step in any discovery process, in which you try to collect a large number of ideas for your development. We will look at two familiar ways this plays out: a textbook version, where you follow best practices of product management, and a common real-life scenario, where things tend to get a little biased and messy. Rest assured - both paths can lead to success.
💡 According to Jeremy Utley’s and Perry Klebahn’s book Ideaflow, the single best predictor of the innovation capacity of a business is ideaflow - the number of novel ideas a person or group can generate around a given situation in a given amount of time.
The textbook scenario: Problem-first thinking
In the ideal world, you have a lot of time to explore and structure the opportunity space - that is, all the customer needs, desires, and pain points you’ve identified. These might come from different sources, such as:
Customer interviews and feedback
Sales and support conversations
Competitive research
And sometimes just the team’s gut feeling and industry experience
As an example, here is an excerpt from the opportunity space for our automotive client, whose goal was to use AI to monitor the global automotive market and create recommendations for strategic innovation:
Note that in this example, we are looking at a brownfield scenario. The opportunity space includes not only new feature ideas, but also critiques of existing features, such as “lack of transparency into sources.“
Once you’ve mapped out the needs, you look at the solution space - all the different ways you could technically solve those problems. For example, these can include:
Rule-based analytics
UX improvements
Artificial Intelligence
Adding more domain expertise
…
Importantly, AI is part of the solution space, but it is in no way privileged - it is one option among many others.
Finally, you match opportunities to solutions, as illustrated in the following figure:
Let’s look at some of those links:
If several users say, “I need alerts when a competitor launches new models,” you might consider using AI. However, a simple rule-based system that scrapes competitor offerings from their websites could solve that too.
If the problem is, “I need to create reports and presentations faster,” AI starts to shine. Summarizing large amounts of data or text to reframe it and generate new content is exactly where modern AI excels.
But if the issue is, “I don’t trust this data because I can’t see the sources,” AI probably isn’t the right fit at all. That’s a UX and transparency challenge, not a machine learning problem.
In this scenario, it’s important to stay impartial when matching each need to the right solution. Even if you’re secretly excited to start building with the latest AI tools (who isn’t?), you have to be patient and wait for the right opportunity to surface.
The real-life scenario: “Let’s use AI!”
Now, in reality, things often start on a different note. For example, you’re in a team meeting, and someone says, “Let’s use AI!” Or your CEO makes a magic speech that suddenly puts AI on your agenda without providing any guidance or direction on what to actually do with it. Without further ado, you risk building AI for the sake of AI.
However, it doesn’t have to be a disaster. We are talking about an extremely versatile technology, and you can work backwards from the AI-first imperative and find great opportunities by ideating around the core benefits and shortcomings of AI.
The AI Opportunity Tree: Focusing on the core benefits of AI
When I work with teams who’ve already decided they “want to do AI,” I help them frame the conversation around what AI is good at. In the B2B context, there are four main benefits you can build around:
Automation & productivity
Use AI to make existing processes faster and cheaper.
Example: Intercom uses AI chatbots to handle common customer service questions automatically, reducing response times and freeing up human agents for more complex cases.
Improvement & augmentation
Help people improve the outcomes of their work.
Example: Notion AI assists with drafting, summarizing, and refining content, while leaving the final decision and editing to the human user.
Innovation & transformation
Unlock entirely new capabilities or business models.
Example: Tesla uses AI to shift from selling hardware to delivering continuous software-driven value with features like driver assistance, battery optimization, and in-car experiences via over-the-air updates.
Personalization
Tailor outputs to specific users or contexts.
Example: Spotify uses AI to create personalized playlists like Discover Weekly, adapting recommendations to each listener’s unique taste.
When ideating, you should try to build a rich space of ideas by collecting multiple opportunities for each benefit. This will result in a structured AI Opportunity Tree. For example, here is a small part of the opportunity tree we built in the automotive scenario:
Use the shortcomings of AI as exclusion criteria
It’s also important to recognize when AI is not the best answer. Here are some of the user-facing shortcomings of AI, which you can use to filter out inappropriate use cases:
AI is often a black box - users don’t always understand how it works.
Example: In financial risk assessments, if a loan applicant gets rejected by an opaque AI model, the bank needs to explain why. Without clear reasoning, the system fails both legally and ethically.
AI introduces uncertainty - the same or similar inputs can produce different outputs.
Example: In legal document drafting, small prompt changes can lead to widely different contract terms. This unpredictability makes it risky for high-stakes, regulated industries.
AI will make mistakes - sometimes in ways you can’t fully predict.
Example: In healthcare diagnostics, a wrong AI prediction isn’t just a bug—it could lead to harmful decisions with life-or-death consequences.
If your use case demands full accuracy, explainability, or predictability, move on - AI is likely not the right solution.
With your AI opportunities and use cases laid out, let’s now see how you can add more flesh to your ideas and specify them for further prioritization and development.
Specification & validation: Iterate yourself to the optimal system design
Once you’ve mapped out your use cases and potential features, the next step is specification and validation. Here, you define how you are going to build an AI system to address a specific use case. Before we dive into the frameworks, let’s pause and talk about process, and specifically about the power of iteration in the context of AI.
Adopting the ritual of iteration
The cover of my book The Art of AI Product Development features a dervish. Just as this dancer rotates in an endless and focused motion, you need to build the habit of iteration to get successful with AI.
At the beginning of any AI journey, uncertainty is high:
You are exploring a new land. Compared to “traditional” software development, where we have a lot of historical wisdom to build upon, the solutions and best practices aren’t figured out yet.
AI systems will make mistakes, which are a major risk for trust and adoption. From the start, you should allocate a lot of time to understanding, anticipating, and preventing these mistakes.
Your users will have different levels of AI literacy. Some will know how to handle errors and uncertainty; others will blindly trust AI outputs, which can lead to problems down the line.
Through iteration, you reduce this uncertainty and build confidence both within your team and for your users. The key is to specify and validate in small steps: run quick experiments, build prototypes, and create feedback loops to understand what’s working and what’s not.
Most importantly, get real feedback early. Today, it’s tempting to cocoon yourself in the world of AI-driven research and simulation. However, that’s a dangerous comfort zone. If you don’t talk to real users and put your prototypes in their hands, you risk a hard clash when your product finally launches. AI is AI, humans are humans. To build something successful, you need to understand and connect both worlds.
Specifying your system with the AI System Blueprint
To make an AI idea more concrete, we use the AI System Blueprint. This model represents both the opportunity and the solution, and its beauty lies in its simplicity and universality. Over the last two years, I was able to use it in literally every AI project I encountered to clarify what was being built. It helps align everyone around the same vision: product managers, designers, engineers, data scientists, and even executives.
Here’s how to fill it out:
Pick a use case from your AI Opportunity Tree.
Map out the value AI can realistically provide to this use case:
How much of it can you automate? Often, only partial automation is possible (and sufficient).
What will the cost of the mistakes made by the AI be? Start with a rough estimate of the frequency and potential cost of mistakes, and correct as you get more information from prototyping and user testing.
Do your users actually want automation? In some contexts—especially creative tasks—users might resist automation. They might prefer to do the task by themselves, or welcome lightweight AI assistance instead of a black-box system taking over their workflow.
Specify the AI solution:
Data will be the raw material powering your AI system.
Intelligence, which includes AI models and your larger architecture, will use AI algorithms to distill value from your data.
The user experience is the channel that transports this value to the user.
Thus, the initial blueprint for our use case of creating presentations and reports can look as follows:
Avoid narrowing down your solution space too early
The following figure shows a high-level solution space for AI:
A full description of this space is out of the scope of this post. If you are interested in an overview, you can find it in chapter 3 of my book, while chapters 4-10 provide deep dives into specific branches.
Here, I would like to guard you against a common mistake - defining your solution space too narrowly. This limits creativity, leads to poor engineering decisions, and can lock you into suboptimal paths. Watch out for these three anti-patterns:
“Let’s build an agent.”
Right now, every other company wants to build their own AI agent. But when you ask, “What exactly is an agent in your context?”, most teams don’t have a clear answer. That’s usually a sign of hype over strategy.
“Let’s pick a model and figure it out later.”
Some teams start by selecting a model or vendor, and scramble to find a use case afterward. This almost always leads to misalignment, iteration dead-ends, and wasted resources.
“Let’s just go with what our platform offers.”
Many companies default to whatever their cloud provider suggests, skipping critical architectural decisions. Cloud providers are biased toward their own ecosystems. If you blindly follow their playbook, you’ll limit your options and miss the chance to develop AI craft and build something truly differentiated.
Thus, before you decide on tooling, models, or platforms, take a step back and ask:
What are the high-level decisions we need to make about data, models, AI architecture, and UX?
How do they interconnect?
What trade-offs are we willing to make?
Also, make sure your entire team understands the whole solution space. In AI, cross-functional dependencies abound. For example, UX designers need to be familiar with the training data of an AI model because it largely determines the outputs users see. On the other hand, data and AI engineers need to understand the UX so they can put the AI system together in a way that allows it to serve the different insights and interactions. Therefore, everyone should be on-board with a shared mental model of the potential solutions and the final specification of your AI system.
Stay up-to-date with the AI solution space with our AI Radar: The more concrete your specification gets, the more difficult it is to keep up with moving parts and new developments. Our AI Radar monitors the latest AI publications, models, and use cases, and structures them in a way that makes them actionable for product teams. If you're interested, please sign up for the waitlist here.
Prioritization: Deciding what to build first
The last step in our discovery process is prioritization - deciding what to build first. Now, if you’ve done a solid job in specification and validation, this will often already point you to use cases with a high potential, making your prioritization smoother. Let’s start with the simple prioritization matrix and then learn how you can refine your prioritization criteria and process.
The prioritization matrix
Most of us are familiar with the classic prioritization matrix: you define criteria like user value, technical feasibility, maybe even risk, and you score your ideas accordingly. Then, you add up the points, and the highest-scoring opportunity wins. The following figure shows an example for some of the items in our AI Opportunity Tree:
This kind of framework is popular because it creates clarity and makes stakeholders feel good. There’s something reassuring about seeing messy, hairy ideas turned into numbers. However, prioritization matrices are highly simplified projections of reality. They hide the complexity and nuance behind prioritization, so you should avoid overrelying on this representation.
Adding nuance to your AI prioritization
Especially when you are just about to introduce AI, you’re not just ranking features, but making long-term bets on your product direction, tech stack, and positioning and differentiation. Instead of reducing prioritization to a spreadsheet exercise, sit with the complexity. Take the time to work through the subtle details, weigh the trade-offs, and make decisions that align not just with what’s easy to build now, but also with the longer-term vision for AI in your business.
1. Pick the low-hanging fruits first
The AI Opportunity Tree from section 1 provides a first hint for your prioritization. Normally, you are better off starting on the left of the tree and moving to the right as you gain more experience and traction with AI. Here’s why:
On the left side, you have simple automation tasks. These are usually low risk, easy to measure, and a great way to start.
As you venture to the right side, you see more advanced, strategic use cases like trend prediction, recommendations, or even new product ideas. These can add more impact, but also more risk and complexity.
Starting on the left helps you build trust and momentum. It delivers quick wins, gives your company the time to get comfortable with AI, and builds the foundation for more ambitious projects down the line.
2. Work on strategic alignment
Before you decide what to build, think about the role of AI in your business. While your company might not have an explicit AI strategy (yet), you can infer important information from its corporate strategy. For example, is AI a potential differentiator, or are you just playing catch-up with the market? If you want to gain a competitive edge with AI, you will want to move fast along your opportunity tree to implement more advanced and differentiated use cases. Your engineering decisions will lean towards more custom and crafty alternatives like open-source models, custom pipelines, or even on-premise infrastructure. By contrast, if your goal is to follow competitors, you might focus on automation and productivity for longer, and choose safer, off-the-shelf solutions from large cloud vendors and model providers.
3. Define custom criteria for prioritization
AI projects often require custom prioritization dimensions beyond the usual trio of user value, business impact, and feasibility. Consider factors like:
Scalability & generalization power
Will your AI solution generalize across different user groups, markets, or domains? For example, if you need to inject heavy domain expertise for every new customer, that limits your scaling curve.
Privacy & security
Some AI use cases are tightly bound to data governance and privacy concerns. If you’re in finance, healthcare, or regulated industries, this becomes critical.
Competitive differentiation
Are you building something truly new, or are you following industry trends? If AI is part of your differentiation strategy, prioritize novel use cases or unique capabilities, not just features everyone else is shipping.
4. Plan for spill-over effects
Another important consideration is spillover effects and the long-term value of building reusable AI assets. When you design and develop datasets, models, pipelines, or knowledge representations with reuse in mind, you’re not just solving one isolated problem, but creating a foundational AI capability. It will enable you to accelerate future initiatives, reduce redundancy, and unlock compounding recurring returns in your business. This is especially critical if AI is a strategic differentiator in your business.
That’s it for today, and I hope this post helped you better understand the value of a structured discovery process in the messy, complex world of AI product development.
Key takeaways:
Use the AI Opportunity Tree to collect, map, and prioritize a broad set of potential AI use cases.
Rely on iteration and continuous feedback to reduce uncertainty and refine your AI product over time.
Leverage the AI System Blueprint to align your team around a shared vision and avoid cross-functional disconnects.
Explore the full AI solution space - don’t fall into the trap of limiting yourself to specific tools, models, or vendors too early.
Treat prioritization as strategic alignment, not just feature scoring. It’s a way to gradually surface, shape, and refine your larger AI strategy.
If you have questions, ideas, or feedback, just hit reply—I’d love to hear from you. And if you have colleagues around who struggle with selecting and defining their AI use cases, share this with them.
Thanks for reading!
Best wishes
Janna
Thanks, that's very helpful to bring some structure into our discovery and ideation!