Here is a number to start with. BCG published its 2026 Applied AI Index on September 30, built on more than 1,300 executives in 20-plus sectors. Only 7.5% of companies reached the top maturity tier. And 95% of that tier does something almost boring: it uses clear KPIs, or it tracks the value of AI straight in the P&L.

I put that one first because the other statistic going around has a date that should be said out loud.

The PwC figure is nine months old

You have probably seen it on LinkedIn again this month. PwC says 56% of CEOs have seen no return from AI and only 12% have both higher revenue and lower costs. True. But the Global CEO Survey was released on January 20, 2026 in Davos. The 4,454 CEOs answered between September 30 and November 10, 2025.

A year-old reading, in a field where a year is a long time. So PwC is my second witness here, not the headline.

One clarification. PwC’s wording is “no significant financial benefit to date”. Another 33% report a gain in revenue or in costs, just not both.

What BCG measured about the 7.5%

BCG’s report gives four numbers that point the same way.

• 61% of the most mature companies have moved from isolated pilots to a multiyear program across the business. Among laggards, 17%.
• Companies that track AI value directly in the P&L realize three times the value of those that do not measure it formally. BCG’s figures are 3.6% against 1.2%.
• Companies that pair strategic clarity with applied AI get five times the value of companies with neither.
• The top tier shows 2.4 times the top-line growth of the bottom half.

PwC gets to a similar place from a different direction. CEOs whose organizations have strong foundations, meaning responsible AI frameworks and a technology environment that lets AI plug into the whole enterprise, are three times more likely to report meaningful financial returns.

One program. One number. A foundation before the scaling.

I should be straight about the limits. These are surveys, not experiments. Companies that already win may simply measure better, and not the other way around. What I can say is that two studies, with different samples and different authors, land on the same shape.

Where the KPI comes from

PwC’s release does not mention a KPI. Sonali Minocha added it when she summarized the study on LinkedIn: “The 12% used AI across the business, and tied it to a specific KPI.” It is a fair reading. But the hard evidence that measuring matters comes from BCG, so I would rather credit it there.

A KPI is not an adoption target

“80% of the team has a license” is an inventory, not a KPI.

An AI KPI is a number your company already reports, that one named person owns, and that you can read in the P&L or next to it. Average days to collect an invoice. Cost per order processed. Hours between a customer request and the first reply.

The foundation is what follows the choice of number: the order in which opportunities get attacked, how the work connects to the systems you already run, who approves what. I have written about why buying tools is not playing offense and about the decisions you cannot hand to an implementer, so I will not repeat them.

How we handle it in AI Maestro

AI Maestro is a two-month program built so the question of the number comes before the question of the tool. First, we map how the company really works (the Process Reality Map). Then each opportunity gets an AI Opportunity Score and a place in a ranking. It all ends at a Go/No-Go gate with a return projection.

If no opportunity moves a number someone in your company already reports every month, the gate says No-Go. The three deliverables stay with you and you have committed to nothing further. If it says Go, you hold a program with an owner, a KPI and an order of attack, which is what BCG found in the 7.5%.

Before your next AI meeting, write one line: the number your biggest AI project is supposed to move, who reports it today, and how often. If you cannot finish that line, you already know why the pilot never leaves pilot.

Pick your number with AI Maestro before you scale