Because you are measuring the wrong half. The reason a buyer approves a purchase and the reason a user comes back tomorrow are two different things, and only one of them shows up in most dashboards.
The example that makes it concrete
Glean announced $300M ARR on May 28, fifteen months after crossing $100M, with Fortune 500 customer count nearly doubling. Everyone read it as a growth story.
The interesting part is what they were saying at the register. The pitch that carried them was not a feature comparison against Copilot. It was closer to: your AI bill is out of control, and we make it smaller.
That works. A CFO in 2026 approves a subtraction faster than an addition.
Now look at the other numbers in the same release. More than 85% of customers use it across five or more departments. It holds a 45% wDAU/wMAU ratio, more than double the enterprise SaaS norm. Nearly half of monthly actives return on a given day.
Nobody logs in daily because they are pleased with the procurement math.
Permission is not value
The savings pitch is not the value. It is the permission.
Cost got them in the door. Return behavior is what kept expansion compounding across departments, and department five is where a $60K land turns into a seven-figure account.
Most builders instrument exactly one of these, and the dashboard ends up matching the sales deck. Cost per query, tokens saved, hours reclaimed, all modeled, all defensible, none of it predictive of whether the thing survives the next budget cycle. Meanwhile the number that actually forecasts renewal, does a person come back tomorrow without being told to, sits uninstrumented because it never made it into the business case.
Why savings-based value decays
If your value is denominated in someone else's cost, your ceiling moves when their cost moves. Token prices have fallen roughly an order of magnitude a year for three years running. Sell a 30% reduction on a line item that keeps shrinking and you have to win more of a smaller thing every renewal.
There is also a tension inside the pitch itself. You cannot fully sell "we lower your AI bill" while metering the customer per query. The claim that survives is consolidation: one context layer instead of six overlapping tools nobody fully adopted. That one does not deflate when inference gets cheap.
The two-layer fix
Run them as separate instruments with separate owners.
The buyer layer answers why this was approved, and it has to be re-provable at renewal in the same units the original case was made in. If you sold savings, you owe a savings number in twelve months. If you do not produce one, someone else will produce a worse one for you.
The user layer answers whether these people would be upset if it disappeared. Unprompted return rate is the cleanest proxy: daily over monthly, cohorted, no email nudges counted. Departmental spread is the leading indicator underneath it, because a tool that escapes its original team has crossed from budget line to infrastructure, and infrastructure does not get cut in Q4.
The one-hour version
Open your metrics review and mark every number as buyer-layer or user-layer. If everything falls in one column, you have been telling one story to two audiences, and only one of them is going to keep paying you.
Then find your actual permission structure, which is rarely cost. Sometimes it is risk, or a compliance deadline, or a competitor's logo showing up in a board deck. It is almost never the thing your users love you for.