65% claim AI expertise. 6% can prove it.

Ask 750 executives about their AI understanding and 65% will describe it as advanced. Press on whether that understanding is leading-edge and operational, running in the business rather than in the briefing deck, and the figure falls to 18%. Ask who can demonstrate measurable value and P&L impact, and 6% remain. That cascade comes from AlixPartners' Digital Disruption Survey of 750 executives, published in Harvard Business Review (December 2024), and it deserves a permanent slot in every board pack.

The distance between 65% and 6% is fifty-nine points of claimed expertise that evaporates the moment someone asks for a number. Most leadership teams sit somewhere inside that gap, fluent in the vocabulary, sincere in the ambition, unable to point to a line in the ledger that moved. The gap is not dishonesty. It is the difference between knowing about a capability and operating with one.

Each step down the cascade has a price

The first drop, from 65% to 18%, is the price of operations. Understanding AI as a concept survives any meeting; claiming a leading-edge operational understanding requires workflows that changed, roles that changed, and decisions now made differently, and only 18% of the executives surveyed will claim even that much (AlixPartners, published in Harvard Business Review, December 2024). Concepts are free. Operations are not.

The second drop, from 18% to 6%, is the price of measurement. Two out of three executives who claim operational mastery still cannot connect it to demonstrated value and P&L impact, which usually means the work is happening without a baseline that predates it, so there is nothing to compare against. A result without a baseline is an anecdote.

The cascade also explains why internal AI committees stall so reliably. A steering committee staffed from the 65% produces strategy documents; a deployment owned by someone accountable to the 6% standard produces ledger entries. The two activities look similar on a calendar and have almost nothing else in common.

Boards have heard this pitch for two decades

Directors sitting through AI strategy sessions in 2026 sat through digital strategy sessions in 2016 and enterprise software business cases in 2006. Each generation of technology arrived with projected savings that rarely survived contact with the operating model, and the current generation is performing to type: 95% of corporate AI pilots deliver no measurable return (MIT, Project NANDA, 2025). A board's cynicism is not a personality flaw. It is pattern recognition with twenty years of training data.

This is why the 6% matters more than the 65%. In a market where nearly two-thirds of executives claim expertise, the claim itself carries no information, a board hears it from every management team and every vendor in the same quarter. What carries information is proof, and proof is scarce enough at 6% to function as a genuine competitive position rather than a slide.

A board that has approved three waves of software capital expenditure asks three questions, in order: which line moved, when did it move, and who owns it. The questions are old because they work. An answer to all three puts a management team in the 6%; an answer to none of them puts it in the 59 points of air between claim and evidence.

For a mid-market organization, $200 million to $1 billion in revenue, the cascade is less forgiving than it is for the giants above it. An enterprise can park a failed program in an innovation budget and absorb the write-off quietly; a mid-market operator carrying a six-figure AI line item answers for it at the next board meeting, by name. The same constraint cuts the other way, though. With one P&L, a short chain of command, and a CEO who can see the entire operation, a mid-market company can reach the 6% standard in quarters rather than years, provided it aims at proof instead of fluency from day one.

What demonstrating P&L impact actually requires

First, one function, not eleven. A deployment scoped to a single function produces a result a CFO can audit; a deployment smeared across the organization produces a narrative. One function with a clean before-and-after beats ten functions with neither, which is why Milton's M2 engagement deploys into exactly one function over 14 weeks before anything expands.

Second, a documented baseline. The 6% can demonstrate impact because they measured before they deployed, cycle times, error rates, cost per transaction, captured while the work was still entirely human. Milton's M1 engagement is audit-only for 4 to 6 weeks for precisely this reason: the baseline gets written down before a single agent is built, so whatever moves later can be attributed rather than asserted. Skipping that step does not save six weeks; it forfeits the proof.

Third, named ownership of the outcome. A number nobody owns is a number nobody defends in front of a board. In Milton's own operation, 43 named agents work alongside 24 humans, each agent carrying identity and boundary files that define exactly which outcome it is accountable for and to whom (internal operating record). When a result is challenged, a specific name answers, agent and human counterpart both.

The cascade is a map, not a verdict. Claiming understanding is table stakes at 65%, operating is progress at 18%, and proving is the destination at 6%, one function, one baseline, one name on the outcome. Reach it deliberately and the board's two decades of cynicism becomes an asset, because the proof stands out precisely where almost nobody else has any. Those outcomes are design targets, and the design is the point.