The 70% problem: what AI vendors won't tell you

Algorithms account for roughly 10% of AI success. The rest is people, process, and leadership. Here's how to address it.

The single most useful number in enterprise AI is a ratio, and it comes from BCG. The firm calls it the 10–20–70 rule:

"10% of a company's efforts should be focused on algorithms, 20% on technology and data, and the remaining 70% — the lion's share — on people and processes."

Put plainly: the model is a tenth of the job. The infrastructure around it is another fifth. Everything else — the part that decides whether anything actually changes — is organizational.

Why the budget runs the other way

Most AI budgets are close to the inverse of that ratio, and there's no conspiracy in it — just incentives. The AI industry sells the 10 and the 20: foundation models, platform licenses, GPUs, data pipelines, connectors. Those are the things with a price list and an order form. The 70 — workflow redesign, role and incentive changes, training, governance, the slow work of getting people to trust and use the thing — has no vendor and no line item, so it tends to be assumed rather than funded.

RAND's 2024 study of why AI projects fail describes the predictable result: organizations acquire AI capability because acquiring AI capability is what you do this year, then look for a problem it fits. More than 80% of AI projects fail, RAND found — about twice the rate of non-AI IT work — and the leading causes are organizational, not technical.

What's actually in the 70%

• Workflow redesign. The pilot has to change how a specific task is done, not run alongside it. A model that doesn't alter a workflow produces a demo.
• Decision rights and governance. Who is allowed to act on a model's output, and who is accountable when it's wrong. McKinsey's 2024 survey found just 18% of organizations had an enterprise-wide body empowered to make responsible-AI decisions.
• Measurement. In the same survey, only 13% of organizations had a clear KPI structure to track the value of their generative-AI work — you cannot manage what no one is counting.
• Literacy and change management. Users have to trust the output enough to rely on it, and change their habits around it. That is a training and communication effort, not a software rollout.

The 20% isn't free either

Data is where the middle tier quietly fails. RAND found that inadequate or inaccessible data is one of the top reasons AI projects collapse — and it's usually discovered after the organization has already committed. Gartner's forecast that at least 30% of generative-AI projects would be abandoned after proof of concept named poor data quality alongside unclear business value as a primary cause. A better model does nothing for a company that can't assemble the data to ground it.

How to rebalance

A rough test: add up what you're spending on models, licenses, and infrastructure, then add up what you're spending on the people-and-process work — training, redesign, a named change owner, adoption tracking. If the first number dwarfs the second, you're buying the 10 and the 20 and hoping the 70 takes care of itself. It won't.

Give the 70 an owner with real authority. Fund it as though it were the majority of the work, because the evidence says it is. And measure adoption with the same rigor you'd apply to the platform contract.