Beyond Buzzwords: How to Build Successful AI-Powered Products
The only thing that might surpass the buzz around AI is the potential of AI. Whether you are a CEO, founder, product leader, product manager, or involved in any way in building products, you’ve surely pondered the impact of AI on your products.
Building AI-Powered Products: Old wine in a new bottle?
When thinking about transforming your products with AI, a natural and important question that comes up is: how do you build successful AI-driven software products? Framed another way, how different is building a successful AI-powered software product from building a successful traditional software product?
A provocative answer is that it is fundamentally different and none of what applies to a traditional product is useful. At the other extreme, a conservative answer is that it isn’t different at all – it is simply old wine in a new bottle. From my experience leading teams in the creation of several AI-driven products, the truth lies somewhere in the middle. The core principles are the same, but the specifics are different in significant ways. In this article, I'll share my reflections and practical insights on the nuances of building AI-driven products.
The 4 Pillars of Success: Valuable, Usable, Feasible, and Viable
Marty Cagan, the author of INSPIRED: How to Create Tech Products Customers Love, identifies four big risks to the success of a tech product: value, usability, feasibility, and business viability. These risks serve as a useful framework within which to consider how to build successful AI-powered products and the nuances specific to AI-powered products.
Value: Start with the problem
As with any product, start by developing a deep understanding of customer problems. Pick the right problem to solve with AI – a problem that is real, a problem that is important to users and customers, and one where either satisfaction with existing solutions is low or your solution is likely to be disruptive.
So, how do you pick the right problem to solve with AI?
Brainstorm with your cross-functional team. If you begin with brainstorming how you could leverage AI in your product, you risk being drawn into solutions in search of problems. I’ve found it to be far more effective to first define or clarify the problem space, the target users and the jobs to be done; and then brainstorm use cases from that shared context and starting point.
Assess the ideas you generate on value, differentiation, defensibility, and alignment with the company and product vision.
Validate your top ideas with potential customers. This is a critical step that will not only help you validate value, but will help you prioritize your efforts and effectively position the product.
“The value is in what gets used, not what gets built.” – Kris Gale
Usability: Tune, Test and Iterate
AI-powered products demand a heightened focus on usability. Here are a few suggestions to ensure that your users find the product usable rather than confusing and frustrating:
Design the user experience to guide users in their interactions with your AI-powered product.
Be transparent with users if you are using AI-powered agents to interact with them. This will set proper expectations and help cultivate trust with your users.
Continually tune the model to gain and maintain sufficient confidence that the model generates high-quality output even with broad variations in the input.
Relentlessly test with internal teams. Involve internal users from across functions and geographies to simulate variation in input and usage. An effective way to obtain feedback quickly is to host testing parties with internal teams. Be sure to create a tight feedback loop with internal users and iterate on improvements quickly.
Pilot the AI-powered product with proofs of concept, early adopter programs and betas to test with customer data, obtain user feedback, tune the model, and refine the product. I would argue this is even more important for AI-powered products than it is for traditional products.
Once you release the AI-powered product, monitor and audit the model output. If appropriate, allow users to provide feedback on the model output, so that your team can quickly identify and address issues.
“Usability is about people and how they understand and use things, not about technology" – Steve Krug
Feasibility: Expand Frontiers
AI undoubtedly expands the frontiers of what is feasible. However, in building AI-driven products, you need to ask two key questions to assess feasibility, and partner with your Data Science and Engineering teams to answer them for the specific AI-powered product you are considering:
Do you have access to the necessary, high-quality data? AI models often need large amounts of high-quality, labeled data for training. If you don’t have the necessary data, how can you generate that data?
Do you have access to the skills and knowledge on the team to develop, train, tune, maintain and optimize AI models? If not, how can you acquire or develop the necessary skills and knowledge on your team?
Viability: Navigate Concerns
AI-powered products bring unique viability risks. At a minimum, you will need to consider and address the following concerns:
Generative AI models are susceptible to hallucinations – imagine AI getting creative in all the wrong ways. AI going off the rails is a real concern, especially to enterprise customers who prize brand integrity. If this is a concern for your AI-powered product, you can minimize the risk by using high-quality and relevant data; limiting responses that fall below a confidence threshold; and providing an escape hatch, for example, you might escalate to a human when AI reaches a pre-defined threshold of unsuccessful attempts.
How do you ensure regulatory compliance? How do your terms of use, SLAs, and contracts need to change to limit legal liability?
What will it cost to train and operate the AI-powered product, and what are the implications on product pricing and packaging?
How will you maintain data privacy? How will you ensure data security so that users can access only the data they have permissions for? While these questions are relevant for any product, they are particularly salient for generative AI-powered products because of the large amount of data required to train these models and the risk of exposure of personal and privileged information given the non-deterministic nature of such AI models.
What, if any, ethical concerns are raised by your AI-powered product? How will you address those concerns?
In summary, like any other product, a successful AI-powered product must be valuable, usable, feasible and viable. However, as we’ve seen, building successful AI-powered products demands a different approach – one that addresses the important nuances specific to AI-powered products. With that in mind, how will you transform your products with AI?
