The consultants are printing precise AI return numbers, but not the path from a raw document to the number.

McKinsey says focused AI lifts EBITDA about 20 percent. PwC says the AI-fit are pulling roughly 7.2 times ahead. Both figures quietly assume infrastructure the client has to pay to build first.

The week produced two headline numbers, both real and both cited by name. McKinsey's Kate Smaje told a May 9 audience that a cohort of focused adopters saw an average EBITDA uplift of 20 percent, generating roughly three dollars of incremental EBITDA for every dollar invested, largely by concentrating on a handful of high-impact areas. PwC's AI performance study, drawn from more than 1,200 senior executives, found that the most AI-fit companies are pulling about 7.2 times ahead of their peers on combined revenue gains and cost reductions, with the top fifth of firms capturing nearly three-quarters of the value AI creates.

Now read the fine print the slides do not show. The 20 percent uplift is achievable if the semantic layer is already built and the agent already has the correct ontology. The 7.2 times gap is real if the client's data is already unified, normalized, and wired to the layer that measures the output. In other words, each number sits downstream of infrastructure that most buyers do not yet have, and that the consulting firm is happy to bill separately to build. The figure is the advertisement. The setup is the actual purchase order.

The piece none of them put on the table is the extraction path: the traceable route from a raw source document to the number on the slide. That path is either the consultant or the code, and it cannot be both. The question a non-technical buyer can ask in one sentence cuts straight through it. Can you run this from my last board deck today, without a data-maturity project first, and show me how each figure was derived, source by source?

In a market busy repricing everything that used to be bundled with model access, an asserted return is the cheapest thing in the room. What holds value is a path that is structured for review, source-cited, and auditable from the raw input forward. The number tells you what is possible. The path tells you whether you own it.