A CFO showed us a slide last quarter with eleven AI projects on it. Nine were live. Two had been killed. The line that mattered was at the bottom: spend was up 340% year over year and the company could not name a single thing it now did that it could not have done in 2023, just slightly slower.
That is not a failure of models, budget, or talent. It's a failure of a single design decision made at the very start of each project — one nobody wrote down, because it didn't look like a decision at the time.
The adjacency trap
There are two shapes an AI project can take, and from a roadmap they are indistinguishable. Both are called "AI." Both have a model in them. Both demo well.
AI-adjacent work bolts a model onto a process that already existed. The summarizer on top of the ticket queue. The drafting assistant beside the CRM. The classifier that pre-sorts the inbox a human still reads. Value is real but bounded — it's the labor you removed, once, and it never grows again.
AI-native work puts the model inside the loop the business runs on. The output changes what happens next, what happens next is observed, and the observation improves the next output. Value starts smaller and does not stop.
The two-minute test
We run this in the first hour of every CRFT Scan, out loud, with the person who owns the P&L in the room. Pick the project. Delete the model from it — not degrade it, delete it. Then describe what the team does on Monday.
If the answer is a sentence about doing the same thing by hand, you're holding an efficiency play. Price it as one. Fund it as one. Do not put it on a slide with the word transformation on it.
If your process survives the model being deleted, the model was never the product. It was staffing.
Three questions that separate a loop from a feature
The delete test tells you whether you're adjacent. These three tell you whether the thing you want to build instead can actually compound.
Where the plateau happens
Adjacent projects follow a shape we've now seen enough times to predict. Weeks one to six look excellent: an obvious manual task disappears and everyone is delighted. Months two and three flatten — the easy task is done and the next one is 40% as valuable and 200% as hard. By month six, the recurring model spend, the integration maintenance and the vendor seat count are all still on the books, and the savings curve has stopped moving.
That's the point where most teams conclude "AI is overhyped." What actually happened is narrower: they bought a one-time cost reduction with a recurring cost structure. The economics were decided at design time, not at renewal.
Rebuilding an adjacent project as a loop
You rarely have to throw the work away. A support summarizer is adjacent — it drafts, a human sends, nothing is recorded. The native version of the same build routes the ticket, drafts the reply against the specific customer's history, watches whether the reply closed the ticket or produced a second contact, and weights the next draft accordingly. Same model, same integration surface. The difference is that the outcome now has somewhere to land.
In practice the retrofit is three pieces of plumbing: an outcome event, a store that keeps it attached to the original input, and a scheduled pass that folds it back into retrieval or scoring. None of it is exotic. It is simply never in scope, because the demo didn't need it.
What to do Monday
Take your current list. Run the delete test on each line — two minutes each, no deck. Sort into two columns. Keep funding the adjacent ones if the payback still clears, but stop calling them strategy and stop expecting them to grow. Then pick exactly one item from the native column and ask what the outcome event would be. If nobody in the room can name it, that's the project's real first milestone.
