Walk a mid-market operator through last year’s software budget and you will find a familiar pattern: a generative AI line item that grew faster than any other category, a slide that promised 20% time back, and a quiet admission in the hallway that nobody can find the 20%.
This is not a model-quality problem. Frontier systems write decent first drafts, summarize noisy tickets, and clear the kind of boilerplate that used to clog Monday mornings. The paradox is operational. Companies bought a capability and installed it on top of processes that were never designed to absorb a faster first draft.
Where the hours actually went
In interviews with finance, IT, and line managers at companies roughly $80 million to $800 million in revenue, three sinks show up again and again.
First, review replaced production. A junior analyst who used to spend three hours building a memo now spends two hours checking whether the model invented a customer, a covenant, or a date. The calendar still shows a three-hour block because review is harder to batch than drafting. The work got different, not shorter.
Second, pilots never became defaults. Tools live in a browser tab that power users love and everyone else forgets. There is no system of record for “this is how we write a QBR.” Without a default path, you do not get compounding. You get anecdotes.
Third, the metric was vanity. Seat counts and prompt volume look like adoption. They do not tell you whether close, cycle time, ticket resolution, or win rate moved. If the scoreboard is licenses, the organization will buy licenses.
A faster first draft in a broken workflow is a more expensive broken workflow.
The firms that did get the hours back
The exceptions are boring, which is why they work. They picked one workflow with a clean definition of done — monthly close commentary, support macros, bid/no-bid memos — and they changed the standard operating procedure, not just the toolbar.
They also staffed a human owner. Not an “AI committee.” A named manager whose bonus includes cycle time for that workflow. Vendors will not do this for you. Neither will a center of excellence that only runs lunch-and-learns.
What to cut, what to keep
If you cannot point to a before-and-after number in 90 days, you do not have a deployment. You have a subscription. Treat it like any other underperforming SaaS: consolidate tools, kill overlapping copilots, and put the remaining spend next to a process map.
Keep the models. Kill the theater. The productivity story in 2026 will not be won by the company with the most seats. It will be won by the company that rewired the few workflows that actually set the P&L.