Why AI Initiatives Fail (and the 10-20-70 Rule)

Start with the uncomfortable number. In MIT NANDA's The GenAI Divide: State of AI in Business 2025 (July 2025), roughly 95% of enterprise GenAI pilots delivered no measurable P&L return — only about 5% reached rapid value or production — despite an estimated $30–40B of investment. The pattern repeats across every major 2025 study, which is why it's a finding, not an outlier.

These are time-stamped findings from 2025 surveys, not timeless laws. The numbers will move; the shape of the gap — wide adoption, thin value — is the durable lesson. Re-verify each figure against the live source before you quote it.

The term practitioners use is 'pilot purgatory': an organization runs demo after demo, each one impressive, none of them changing a number on the P&L. The critical reframe for a leader is that this is not a technology failure. The models work. The failure is organizational — and that means it is yours to fix, not the vendor's.

The single most useful framework for understanding why initiatives fail is BCG's 10-20-70 rule. It describes how effort, cost, and attention should be split in a successful AI transformation:

The insight that changes behavior: winners invest in the 70%; stalled pilots invert the ratio, pouring attention into model selection and infrastructure while starving adoption and process redesign. The 70% is precisely the part a non-technical leader is best positioned to own — and the part that gets neglected because it's harder, slower, and less exciting than a demo.

The corroborating evidence is consistent: Deloitte found only ~37% of organizations invest significantly in change management, even though ~70% of the value (and the failure) sits in the people-and-process layer. You cannot buy your way out of the 70% with a better model.

MIT NANDA gave the failure a precise name: the 'learning gap.' Pilots fail not because the underlying models are weak, but because the tools don't learn, adapt, or integrate into how work actually gets done. There is no memory of the organization, no feedback loop, no redesign of the surrounding process — so the tool sits beside the workflow instead of inside it, and the value never compounds.

McKinsey describes the same phenomenon from a different angle as the 'gen AI paradox': horizontal copilots and chatbots scale fast but deliver diffuse, unmeasurable gains, while the transformative vertical, function-specific use cases — where the real money is — stay roughly 90% stuck in pilots. The easy wins are shallow; the deep wins require integration work that most firms skip.

The practical implication is liberating: the bottleneck is rarely the next, better model. Waiting for a more capable model to rescue a stalled pilot is usually a category error. The fix is integration and adoption — wiring the tool into a redesigned workflow, giving it access to the right data, building the feedback loops, and getting people to actually use it as part of their job.

Stalled programs fail in recognizable, repeatable ways. Learn to spot these five — each has a one-line tell and a one-line fix.

Numbers 1, 2, and 4 are the ones leaders most often cause personally. Number 3 is the one that hides the other four — without honest measurement, a program can look healthy on a dashboard while creating no value, exactly as it did at Klarna.

A note on the fifth mode, because it is the 2025–2026 trap. Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027 — driven by unclear value, cost, and weak risk controls. Agents add autonomy, and autonomy multiplies both value and risk. They belong after you have an integrated, measured, well-governed foundation, not as the opening move.

If you change only one thing, change this. McKinsey's research finds that workflow redesign has the single biggest effect on the EBIT impact firms get from AI — and yet only about 21% of firms have fundamentally redesigned their workflows. The overwhelming majority 'bolt AI onto' legacy processes and then wonder why the returns are diffuse.

Re-verify the 21% figure live before teaching: McKinsey, The State of AI: How organizations are rewiring to capture value (2025). The finding — redesign beats bolt-on — is durable; the exact percentage will drift.

The distinction in plain terms:

This is why McKinsey's four pillars for agentic value all point the same way: (1) pick 2–3 high-value vertical use cases; (2) embed AI into reimagined workflows — not legacy ones; (3) scale with infrastructure, cross-functional teams, and strong governance; (4) build a new skills mix. Pillar 2 is the whole ballgame. Redesigning a workflow is, by definition, a 70%-layer activity — which is exactly why it is the executive's job and exactly why it gets skipped.

Across roughly 25 organizational attributes McKinsey tested for correlation with bottom-line impact, one stood out: CEO oversight of AI governance is the attribute most strongly correlated with EBIT impact, especially at large firms. It is not a coincidence that the most-cited 'leaders get it wrong' signal is the inverse — accountability for AI is diffuse, with only a minority of firms saying the CEO or board owns AI governance.

Re-verify live before teaching: McKinsey, The State of AI (2025). The ranking is the durable point; the exact attributes and percentages refresh annually.

Why would governance ownership — not, say, technical talent or budget — be the strongest predictor? Because CEO oversight of governance is a proxy for genuine ownership of the whole transformation. A CEO who owns governance is, in practice, the kind of leader who:

• treats AI as a strategic transformation, not an IT line item;
• funds and protects the 70% (workflow redesign, change management, reskilling);
• role-models use — and BCG found high performers are far more likely to have senior leaders who personally use AI daily;
• sets and tracks real KPIs, so the program is measured on value, not activity;
• and engages the C-suite — deeply engaged C-suite teams are reported to be many times more likely to be top AI value creators.

Governance here is not a brake. Done well, it is the mechanism by which a leader steers the program toward value and away from the five failure modes. The opposite — delegating AI to a central lab and hoping it diffuses upward — is the single clearest predictor of a stalled program.

The themes converge into a short diagnostic any leader can run on any AI initiative. If the answers trend right, you have a transformation; if they trend left, you have a pilot heading for purgatory.

The through-line of this entire lesson: the failure is organizational, not technical, and that is good news — because organizational levers are exactly the ones an executive controls. You do not need to become technical to fix a stalled AI program. You need to fund the 70%, redesign the workflow, concentrate the bets, measure value honestly, and own it from the top.