The AI Bubble Is a Design Bubble

AI is attracting record investment and delivering record disappointment. According to RAND, 80% of AI solutions fail to deliver on their promises. An MIT study found that 95% of generative AI pilots never generate measurable business value. The algorithms work. The infrastructure exists. The data is there. Yet the outcomes fall flat.

This isn't a technology crisis. It's a design crisis. And until we treat it as one, the failure rates won't budge.

The Circular Trap

Today's AI economy has entered a dangerous loop. Market pressure forces companies to ship AI features fast — "every product needs AI" becomes the mandate. Design teams, under unrealistic timelines, apply AI to legacy systems without addressing fundamental workflow problems or user needs.

The result is software that's technically impressive but experientially broken. Systems that predict accurately but misread intent. They automate dysfunction instead of redesigning it. I call this "putting AI on a pig" — applying sophisticated technology to fundamentally flawed foundations and hoping capability compensates for poor design.

When these products fail, investors tighten funding. But rather than questioning the design approach, the market demands faster returns from the surviving projects. Companies double down on shortcuts. More features, faster shipping, less foundational thinking. The failure rate deepens.

This isn't a technology bubble. It's a design bubble — inflated by the misconception that technological capability alone creates value.

The Mirage of Statistical Success

Most AI systems are built on forecasting — analyzing historical data to predict what happens next. This works well in structured, repetitive environments. Supply chains. CRM systems. Inventory management.

But human behavior is not a supply chain. In entertainment, wellness, learning, and creative work, behavior is nuanced and context-dependent. A prediction that makes perfect statistical sense can miss the mark entirely when it fails to account for shifting moods, life circumstances, or the simple desire for novelty.

Consider the Nest thermostat. It learns your habits beautifully — until life changes. A baby arrives, and overnight, comfort and safety replace energy efficiency as the priority. But Nest, trained on a year of past routines, keeps optimizing for yesterday's goals. It's statistically correct and contextually blind.

Or Amazon Prime. The platform has your entire purchase history, viewing habits, and Alexa data. Yet you open Prime Video and get hit with ads on a service you're already paying for, followed by a maze of "included," "rent," and "buy." The system knows what you watched. It has no idea why. Instead of helping you decide, it overwhelms you.

Forecasting optimizes for probability. But probability alone doesn't create value — because it can tell you what's likely, never what's preferable. It reinforces existing patterns rather than imagining new ones.

Three Missing Human Elements

When AI projects fail, the root cause typically lies in the absence of three human dimensions that forecasting-only design systematically ignores.

The first is intent. Systems overwhelmed by historical data misinterpret why users behave the way they do. They privilege past patterns over present circumstances — confusing what someone did last month with what they need right now.

The second is foresight. AI excels at pattern recognition but fails at purposeful imagination. It can forecast what's likely based on yesterday's data, but it can't envision what should happen or why it matters. Without foresight, predictions become technically accurate but contextually irrelevant.

The third is agency. AI solutions automate too much while explaining too little. Users receive predictions and recommendations with no visibility into how decisions were made. When people can't form a mental model of how a system thinks, trust erodes. The product may be precise, but it violates a fundamental human need: autonomy.

These aren't edge cases. They're the default outcome when the only design methodology is forecasting.

From Probability to Possibility

Breaking this cycle requires pairing forecasting with its opposite: backcasting. Where forecasting projects forward from historical data, backcasting starts with a desirable future and works backward to identify what must happen today to reach it.

The simplest way to explain the difference: forecasting looks at your last five beach vacations and recommends another beach. Backcasting is you saying, "I want to climb a mountain in Peru" — and the system figuring out what needs to happen now to make that possible.

That shift — from "what will you probably do" to "what do you want to achieve" — is the shift most AI products haven't made.

In practice, this means starting every project by defining the preferred future state: who should users become, what should they be able to accomplish, how should their context evolve? Then designing backward, using behavioral science frameworks to ensure interventions actually land — aligning motivation, ability, and timing so the system supports meaningful change rather than recycling predictions.

This isn't about abandoning data. Forecasting tells you where users are. Backcasting tells you where they're going. You need both. But direction must come first. Data second.

The Design Dividend

For executives and investors evaluating AI portfolios, the implications are counterintuitive but clear.

Short-term velocity is killing long-term value. Rushing to deploy AI features reinforces forecasting-only design, which produces the very failure rates that alarm investors.

Design is not a downstream concern. In the current model, capital flows toward "AI capabilities" in the abstract, and design teams figure out the application afterward. This is backward. Design methodology should inform investment decisions, not follow them.

The real competitive moat is methodology, not model sophistication. As AI capabilities commoditize, sustainable differentiation will come from organizations that design with foresight — not just with data.

And pilot failure rates are a design metric, not a technology metric. When 95% of pilots fail, the problem isn't insufficient AI capability. It's insufficient design rigor.

The Path Forward

Today's AI bubble reflects a circular trap: market pressures create design shortcuts, design shortcuts produce failed products, and failed products intensify market pressures. The cycle persists because the industry mistakes the symptom — weak AI capabilities — for the disease: a forecasting-only design philosophy.

The AI that fails won't be the one that got the math wrong. It'll be the one that never asked the right question.

Only when AI systems become humanly right — not just statistically right — will the industry deliver on its promise.

Joana Cerejo is a Senior Experience Designer specializing in AI-driven products and anticipatory design. She is the author of The Anticipatory Design Playbook (CRC Press, 2025).