When Founders Skip Data Work, the Pilot Dies First

Seventy-five percent of AI initiatives miss their expected ROI. That figure comes from IBM's Q4 2025 CEO study, which surveyed executives across industries. The same study found only 16 percent of AI projects had scaled enterprise-wide. These are not early-stage experiments with uncertain outcomes. These are funded, intentional deployments that failed after the money was spent.

The pattern across IBM, Gartner, and MIT is consistent enough to be uncomfortable. Gartner surveyed 782 infrastructure and operations leaders and found only 28 percent of AI use cases fully met ROI expectations. MIT's NANDA initiative reviewed more than 300 public AI initiatives, ran 52 structured interviews, and reported that 95 percent of organizations piloting generative AI showed zero measurable profit-and-loss impact. S&P Global Market Intelligence data integrated into the MIT report shows abandonment before production rising from 17 percent in 2024 to 42 percent in 2025, with 46 percent of projects scrapped between proof of concept and broad adoption on average.

None of these studies blame the models. Gartner names poor data quality as a top driver for the 30 percent of generative AI projects it predicted would be abandoned by end of 2025. IBM links ROI directly to capability maturity in data and governance. MIT's data shows the abandonment happening before production, which means before the team ever got to evaluate whether the model was the right choice.

What the budget split reveals

When AI projects do reach production, the budget allocation follows a recognizable pattern. Roughly 30 to 40 percent of total AI project spend goes to data preparation and integration. Another 15 to 20 percent goes to ongoing operations and infrastructure. Workforce training and change management absorbs somewhere between 5 and 15 percent of program-level spend depending on depth and scale.

These figures come from the research scope document underlying this analysis, which synthesizes cost breakdowns across institutional studies. They describe what successful deployments actually spend, not what founders plan to spend. The gap between those two numbers is where most pilots die.

IBM's IBV study puts numbers on what that gap costs. Average enterprise AI ROI sits at 5.9 percent. Top-quartile organizations, the ones that built mature data and governance capabilities, reach 13 percent. IBM tracked this over time, from roughly 1 percent around early 2020 to nearly 6 percent by end of 2021, with an estimate of 8.3 percent in 2022. Even the improving trend sits below the typical cost of capital near 10 percent for most industries. The organizations reaching 13 percent are not using better models. IBM attributes the separation to capability maturity in data and governance.

Data governance investments, specifically, return over triple their cost on average according to the research. AI training programs run between roughly $800 and $3,500 per employee for meaningful upskilling. These are not rounding errors in a budget. They are line items founders routinely omit from their initial AI cost models.

The lean-startup argument deserves a direct answer

A reasonable founder argues this: before product-market fit, you do not know which data you need, which workflows AI will touch, or which regulatory requirements will apply. Spending 30 to 40 percent of a constrained budget on data infrastructure before those questions are answered is premature optimization. Get to product-market fit first. Fund proper data work from revenue. Build governance for the organization that exists, not the one you are projecting.

This argument has real structural logic. It is the same sequencing principle that tells founders not to hire a general counsel before Series A. The SMB surveys showing "10x" returns from early AI automation give it empirical cover. If narrow, fast-moving AI deployments on imperfect data are producing outsized returns for small teams, the case for front-loading data infrastructure before validation looks like enterprise risk management applied to the wrong context.

The problem with the argument is not the logic. The problem is the assumption it depends on: that deferral is temporary.

MIT's data breaks that assumption. Only 5 percent of organizations piloting generative AI reach production. The abandonment rate before production jumped from 17 percent in 2024 to 42 percent in 2025. These are not organizations that deferred data investment and then built it later. They are organizations that never reached the stage where deferred investment was going to arrive. The failure happens in the window the lean-startup argument treats as temporary.

Gartner makes the mechanism explicit. Poor data quality does not become a problem after validation. It ends the pilot during validation. The deferred investment never arrives because the project is cancelled first.

What a working cost model looks like

If you are building an AI cost model from scratch, the 30 to 40 percent figure for data preparation and integration is the first line item to anchor. Not because it is the most interesting spend, but because it is the one most likely to be missing entirely from your current budget.

The second line item to add is governance. The research documents governance investments returning over triple their cost. That figure does not come from a governance vendor's marketing material. It comes from institutional research tracking actual program outcomes. If you are skipping governance because it feels like overhead, you are skipping the line item with the highest documented return in the stack.

The third number worth anchoring is training. Between $800 and $3,500 per employee for meaningful upskilling is a wide range, but it is a range with a floor. If your current budget for AI workforce enablement is zero, the floor matters more than the ceiling.

IBM's top-quartile organizations reach 13 percent ROI. The average sits at 5.9 percent. The difference is not which foundation model they chose or how many tools they licensed. IBM traces it to capability maturity in data and governance. A founder who builds that maturity into the initial budget is not being conservative. They are allocating toward the outcome that the data actually predicts.