Nobody reorganizes because the AI worked. They reorganize because they cannot tell whether it worked, and a company that cannot see the return on a large investment will always reach for a decision it can announce. The announcement is countable. It has a date, a headcount, and a sentence a board can repeat. Whether the software is doing the work is a separate question and a much slower one.
Here is the mechanism. A company spends two years buying AI. The spend is real and the returns are diffuse, so nothing in the reporting stack can isolate them. Pressure builds to demonstrate seriousness. A restructuring becomes the only artifact available that resembles proof, because it is visible, expensive, and irreversible. You can watch this happen in any planning cycle. Somebody asks what the AI investment returned, three teams produce three answers, the answers do not reconcile, and the meeting ends with a decision to go bigger.
A reorganization is not evidence that AI is working. It is evidence that somebody needed something to say about AI.
monday.com gave us the cleanest instance of this in July, and the useful part is that all of it sits in public filings, so you do not have to take my word for any of it. The company cut about twenty percent of its workforce and told everyone that the organization built for its previous chapter was not the organization that fits the new AI era. Three weeks later it reported the quarter. Sales and marketing expense was $162.4 million against $152.6 million a year earlier. The spend went up. Revenue grew faster, so the line fell as a share of revenue, and that is the whole story. Nothing was automated away. A cost grew more slowly than the business funding it.
The rest of the disclosure says what actually happened. Sales headcount is shifting toward enterprise accounts and staying flat. The chief revenue officer described concentrating resources on larger customers. The savings are going into engineers who deploy the product inside those accounts. That is a company moving upmarket. It is a good strategy, it is thirty years old, and it has nothing to do with language models.
The filing contains the tell. The founders wrote that improving margins was not the purpose of the decision, and in the same document the company raised its operating margin outlook and left revenue guidance untouched. Both things can be true at once. They are still one event described twice, and only one of the descriptions was written for a regulator.
The honest version was available the entire time. It reads: we are moving upmarket, the motion that got us here does not serve where we are going, and it costs a fifth of the company. That sentence is accurate and defensible and nobody says it, because it earns no narrative credit. Going upmarket is a decision. Entering the AI era is a destiny. Boards fund destinies.
None of this means the execution layer is safe. I have argued that it is not and I still think that. But there is a distance between work becoming commoditized and people becoming surplus, and the companies collapsing that distance are betting on a number they have not measured. The only controlled trial anybody has run on this found experienced engineers were nineteen percent slower using AI tools while reporting they felt twenty percent faster. That gap is the entire risk. A team that feels faster and cannot prove it is not a staffing decision waiting to be made. It is a measurement problem, and cutting people is an expensive way to avoid solving it.
So here is the test, and I am stating it plainly enough to be wrong in public. If AI is eating the marketing execution layer, the evidence will not be a headcount announcement. Headcount moves for a dozen reasons and the release always credits the most flattering one. The evidence will be a sales and marketing line that falls in absolute dollars while revenue grows twenty percent or more. In dollars. Not as a share of revenue. Nobody has printed that quarter.
If you are looking at a company that just reorganized around AI, ask what the AI replaced. Then ask what sales and marketing spent, in dollars, this year against last. Anyone can answer the first question, it is in the press release. The second one is only answerable by somebody who sat in the meeting where the honest version got proposed and watched the room decide it was not a story anybody wanted to tell.
First touch attribution makes brand look brilliant. Last touch makes paid search look brilliant. Multi touch makes whoever configured the weights look brilliant. Every model is wrong in a direction that flatters somebody, and the somebody is almost always the person who chose the model.
This is not a scandal. It is arithmetic. Attribution assigns credit for an outcome with many causes, and there is no objectively correct way to split a joint cause. The literature has known this for decades. What is strange is that marketing keeps shopping for a model as though one of them is going to turn out true, and then presents the output as measurement rather than as the interpretation it actually is.
The useful question is not which model is true. It is which decision the model is supposed to change.
Start from the decision and the model gets simple fast. If the decision is where to move next quarter's incremental spend, you do not need a model at all. You need a holdout. Turn the channel off in a matched geography, watch what happens to pipeline, and you have a causal answer no weighting scheme can give you. If the decision is which content to make more of, last touch is fine and cheap, because you are ranking within a stage rather than across the funnel. If the decision is whether brand investment is working, accept that you are running a two year experiment and stop asking the dashboard to settle it every month.
Most attribution work fails because it is not attached to a decision at all. It is attached to a narrative obligation. Somebody has to explain the number to the board, and a model produces an explanation on demand, so the model gets built and then defended long past the point where anyone consults it before spending money. You can identify these instantly. Ask a team what they would do differently if the model showed the opposite result. If there is no answer, the model is theater.
I run attribution as a reporting convention, not as truth. Pick a model, write down why you picked it, keep it stable long enough for trends to mean something, and never let it decide anything important on its own. Real budget decisions come from incrementality tests, from sales conversations about what buyers actually said, and from the uncomfortable judgment calls that are the job. The dashboard tells you where to look. It does not tell you what to do.
The seventy twenty ten split holds up better than any model I have used. Seventy percent into what is already performing, twenty into what is showing early signal, ten into novel bets. It requires no attribution to operate and it survives being wrong, which is more than the models can say.