AI made the work cheaper to produce. It made it more expensive to trust.

The profession has bought. Clio recorded AI use in daily legal work rising from 19 per cent to 79 per cent in a single year (Legal Trends Report 2024). Thomson Reuters' respondents predicted twelve hours a week of recovered capacity (Future of Professionals Report 2024).

The money has not appeared. Not in the form predicted, at the scale predicted, or in the place predicted. Meanwhile the people who sign the work off say they are busier.

Two explanations occupy the field. The pricing one and the strategy one. Both are correct. Neither is sufficient, because both model the work AI does and neither models the work AI creates.

Every unit of AI-produced work creates a verification obligation before it can be delivered. That obligation cannot be discharged by anyone more junior than the professional the AI displaced. So verification migrates up the value spectrum at the same moment production migrates down, inside the same matter, in opposite directions.

The result is a margin scissors. Price tracks production, because clients price what they can see themselves receiving. Cost tracks verification, because the firm pays whoever signs it off. The first blade is well documented. The second, as far as the published research shows, is not yet measured.

Inside the paper:

• Verification as a thirteenth activity: with its own position on the Maister spectrum and its own economics.
• The seniority delta and the verification ratio: the two numbers that decide whether an AI deployment adds margin or destroys it.
• The deskilling loop: why the verification control degrades as it is loaded, so firms pay more for quality assurance that works less well.
• Why the load is uneven: and why adoption clusters by supervising partner rather than by practice area.
• A thirty-year misdiagnosis: the profession's constraint is criticism aversion, not risk aversion. The two need different remedies, and the second is largely learned, so it moves.

Three predictions are stated in falsifiable form, with a measurement protocol any firm can run against records it already holds. The paper contains no primary data. That is a design choice, and the claim-type taxonomy in Section 3 is the mitigation.