Three Questions Before You Build Anything with AI

Eighty-eight percent of organizations McKinsey surveyed use AI in at least one business function. Fewer than one-third have started any meaningful scaling program. That is not a model quality problem. The models are better than they were two years ago. The failure is upstream, in which problems founders choose to solve with AI before they understand whether the data exists to solve them or whether the outcome is worth measuring.

Accenture estimates 80–85% of companies are stuck in what they call a "proof-of-concept factory," where AI work stays siloed, IT-led, and disconnected from business outcomes. Founders building AI products often mirror this inside their own companies without noticing. They respond to customer curiosity with demos. The demos produce enthusiasm. The enthusiasm produces another pilot. Nothing ships.

What RAND found when they asked why

RAND interviewed sixty-five data scientists and engineers with at least five years of industry experience and asked them to describe why AI projects fail. Eighty-four percent cited leadership-driven causes as the primary driver. Not the model. Not the infrastructure. The absence of a clear problem definition and the absence of a measurement criterion for whether the thing worked.

The fix is not a framework in the consulting sense. It is three questions you answer before you commit to building.

The three questions

First: what is the business value if this works? Not "it will make us more efficient." A number, or a named process with a named cost. If you cannot name it, you are selecting on model capability or market enthusiasm, which is exactly the pattern MIT NANDA found in the 95% that produced no P&L return.

Second: does the data exist today? RAND ranked weak data quality and utility as the second most cited failure cause, and MIT's funnel data shows that among organizations evaluating enterprise AI tools, only 20% reached the pilot stage and only 5% reached production. The drop-off happens before the technical work begins. Founders who discover mid-build that their CRM, their product logs, and their billing system each hold a different version of the same customer record are not learning something new. They are paying to learn something they could have checked on a Tuesday afternoon.

Third: what breaks if this goes wrong? Not a formal risk register. Just a sentence. Who sees the output, what do they do with it, and what is the cost if the model is wrong 15% of the time?

When fast pilots produce slower paths to production

A reasonable objection: early-stage founders lack the historical data, the usage records, and the process documentation to score these questions with any precision. The most useful signal about expected value and data readiness comes from running a fast, scoped pilot, not from scoring criteria before the pilot exists. Some wasted pilots are the price of exploration.

This argument has force at the margins. It overstates what the three questions require. RAND's interviews did not identify missing risk registers or incomplete data maturity scores as the cause of failure. They identified absent problem definition and missing measurement criteria. A founder who writes down what business problem the AI use case solves, names a metric that would confirm it worked, and checks whether the data needed to run the model currently exists is doing the work RAND found absent in failed projects. That takes an afternoon, not a sprint.

The "healthy pruning" interpretation of high failure rates also runs into S&P Global's Voice of the Enterprise data, which documented a sharp rise in AI initiative abandonment between 2024 and 2025 across more than one thousand IT and business professionals in North America and Europe. S&P frames that abandonment as a problem. Accenture frames the proof-of-concept factory as a failure of governance design. Neither source reads the pattern as productive exploration.

What the 5% did differently

MIT NANDA found that only 5% of organizations piloting generative AI extracted significant value. The separating variable was not model selection or engineering quality. Organizations reaching production shared common data infrastructure characteristics. Those that did not were disproportionately blocked at the data integration stage, before any model ran.

Pick the AI opportunity where you can answer all three questions before you build. If you have five ideas and only one clears all three, that is your answer. If none clear all three, the most valuable thing you can do this week is fix the data problem in the opportunity with the highest expected value, not start a new pilot.