Betting on AI? You must first consider product-market fit

The AI boom isn’t going to plan. Organizations are struggling to turn AI investments into reliable revenue streams. Enterprises are finding generative AI harder to deploy than they’d hoped. AI startups are overvalued, and consumers are losing interest. Even McKinsey, after forecasting $25.6 trillion in economic benefits from AI, now admits that companies need “organizational surgery” to unlock the technology’s full value. 

Before rushing to rebuild their organizations, though, leaders should go back to basics. With AI, as with everything else, creating value starts with product-market fit: Understanding the demand you’re trying to meet, and ensuring you’re using the right tools for the task. 

If you’re nailing things together, a hammer is great; if you’re cooking pancakes, a hammer is useless, messy, and destructive. In today’s AI landscape, though, everything is getting hammered. At CES 2024, attendees gawped at AI toothbrushes, AI dog collars, AI shoes and AI birdfeeders. Even your computer’s mouse now has an AI button. In the business world, 97% of executives say they expect gen AI to add value to their businesses, and three-quarters are handing off customer interactions to chatbots.   

The rush to apply AI to every conceivable problem leads to many products that are only marginally useful, plus some that are downright destructive. A government chatbot, for instance, incorrectly told New York business owners to fire workers who complained about harassment. Turbotax and HR Block, meanwhile, went live with bots that gave bad advice as often as half the time. 

The problem isn’t that our AI tools aren’t powerful enough, or that our organizations aren’t up to the challenge. It’s that we’re using hammers to cook pancakes. To get real value from AI, we need to start by refocusing our energies on the problems we’re trying to solve.

The Furby fallacy

Unlike past tech trends, AI is uniquely prone to short-circuiting businesses’ existing processes for establishing product-market fit. When we use a tool like ChatGPT, it’s easy to be reassured by how human it seems and assume it has a human-like understanding of our needs. 

This is analogous to what we might call the Furby fallacy. When the talkative toys hit the market in the early 2000s, many people — including some intelligence officials — assumed the Furbys were learning from their users. In fact, the toys were merely executing pre-programmed behavioral changes; our instinct to anthropomorphize Furbys led us to overestimate their sophistication. 

In much the same way, it’s easy to wrongly attribute intuition and imagination to AI models — and when it feels like an AI tool understands us, it’s easy to skip over the hard task of clearly articulating our goals and needs. Computer scientists have been wrestling with this challenge, known as the “Alignment Problem,” for decades: The more sophisticated AI models become, the harder it gets to issue instructions with sufficient precision — and the greater the potential consequences of failing to do so. (Carelessly instruct a sufficiently powerful AI system to maximize strawberry production, and it might turn the world into one big strawberry farm.) 

The risk of an AI apocalypse aside, the Alignment Problem makes establishing product-market fit more important for AI applications. We need to resist the temptation to fudge the details and assume models will figure things out for themselves: Only by articulating our needs from the outset, and rigorously organizing design and engineering processes around those needs, can we create AI tools that deliver real value.

Back to basics

Since AI systems can’t find their own path to product-market fit, it’s up to us, as leaders and technologists, to meet the needs of our customers. That means following four key steps — some familiar from Business 101 classes, and some specific to the challenges of AI development. 

Understand the problem. This is where most companies go wrong, because they start from the premise that their key problem is a lack of AI. That leads to the conclusion that “adding AI” is a solution in its own right — while ignoring the actual needs of the end-user. Only by clearly articulating the problem without reference to AI can you figure out whether AI is a useful solution, or which types of AI might be appropriate for your use-case.

Define product success. Discovering and defining what will make your solution effective is vital when working with AI, because there are always trade-offs. For example, one question might be whether to prioritize fluency or accuracy. An insurance company creating an actuarial tool might not want a fluent chatbot that flubs math, for instance, while a design team using gen AI for brainstorming might prefer a more creative tool even if it occasionally spouts nonsense. 

Choose your technology. Once you understand what you’re aiming for, work with your engineers, designers and other partners on how to get there. You might consider various AI tools, from gen AI models to machine learning (ML) frameworks, and identify the data you’ll use, relevant regulations and reputational risks. Addressing such questions early in the process is critical: Better to build with constraints in mind than to try to address them after you’ve launched the product. 

Test (and retest) your solution. Now, and only now, you can start building your product. Too many companies rush to this stage, creating AI tools before really understanding how they’ll be used. Inevitably, they wind up casting about in search of problems to solve, and grappling with technical, design, legal and other challenges they should have considered earlier. Prioritizing product-market fit from the outset avoids such missteps, and enables a process of iterative progress toward solving real problems and creating real value.

Because AI seems like magic, it’s tempting to assume that deploying any AI application in any setting will create value. That leads organizations to “innovate” by firing off flurries of arrows and drawing bullseyes around the spots where they land. A handful of those arrows really will land in useful places — but the vast majority will yield little value for either businesses or end-users. 

To unlock the enormous potential of AI, we need to draw the bullseyes first, then put all our efforts into hitting them. For some use-cases, that might mean developing solutions that don’t involve AI; in others, it might mean using simpler, smaller, or less sexy AI deployments. 

No matter what kind of AI product you’re building, though, one thing remains constant. Establishing product-market fit, and creating technologies that meet your customers’ actual wants and needs, is the only way to drive value. The companies that get this right will emerge as winners in the AI era.

Ellie Graeden is a partner and chief data scientist at Luminos.Law and a research professor at the Georgetown University Massive Data Institute.

M. Alejandra Parra-Orlandoni is the founder of Spirare Tech.