COLUMN
When “people are using it” doesn’t get past the board,
the first three metrics to measure
“People are using it” can’t decide a renewal. The only numbers you can put in front of the board are usage rate, answer quality, and correction rate.
Check first whether the three metrics can be written →No plan document, no sales deck after the call
The generative AI renewal case has no KPI number in it
The generative AI licenses are in. Ask the floor and you get “we’re using it.” But when the board asks “do we renew at the same amount next year,” all you have is a login count or a satisfaction survey.
The numbers that can carry a renewal decision are, first, usage rate, answer quality, and correction rate. That doesn’t dismiss cost savings or audit trails. It isn’t a call to add dashboard tiles either.
McKinsey’s The State of AI (2025) reports that fewer than one in five organizations track clear KPIs for generative AI; definitions and scope are as in the original. “Adopted” without measurement is a feeling. A feeling can’t cut a renewal or a reduction. Adding a login count doesn’t change the board’s question.
Logins and satisfaction don’t get a generative AI approval through
| Common number | Why it fails in an approval |
|---|---|
| Seats issued / login count | Can’t tell “opened it” from “the work changed” |
| Satisfaction survey | Only users answer. Non-users’ reasons never surface |
| PoC accuracy | In live work, you can’t see how much people are correcting |
What you need is not “expected benefit before rollout” but measured results after. More pre-rollout assessments don’t produce renewal evidence.
Usage rate — three layers: monthly, sustained, high-frequency
Don’t say “30% usage” on one line. The same 30% made of people who tried once and left, versus people using it every week on the same task, points to opposite renewal decisions.
- Monthly: used it at least once that month
- Sustained: used it on the same family of tasks for two or more months
- High-frequency: it is in the weekly flow of their work
What goes to management is a table with these three layers side by side. Before anything, cut whether “people are using it” refers to the high-frequency layer or the monthly one.
Answer quality — turn “feels good” into kinds of error
“Reasonably usable” can’t be written into an approval. On the floor, quality mostly splits three ways.
- Outright wrong: facts are incorrect. In financial or legal judgment, this is the cut-off
- Unsupported: looks right, but wasn’t grounded in the internal source of truth
- Usable with edits: the skeleton is good; a person always touches it
Absolute scores vary by task and by scorer. What you decide first is who tolerates which kind of error.
Correction rate — how much people are fixing is the substance of adoption
When something is used but the benefit never becomes a number, the correction rate is usually hiding. If a person rewrites half of it every time, the license is a “drafting device.” That isn’t bad. But the basis for a renewal amount needs before-and-after on how much gets fixed.
What brings the correction rate down is more often handing the tool that task’s source of truth (manuals, past cases, terminology) than switching models. Buy “the next AI” without measuring this and the same correction rate shows up in the next product.
Once you have all three, make one fix
With three metrics side by side, the stuck points separate into authority, source of truth, procedure, and evaluation. Try to fix all of them and the floor stops. Implement the one fix with the largest effect, then re-measure on the same three metrics two weeks later. Only what shows a difference becomes renewal evidence.
Don’t reach renewal month with nothing but a login count.
We only check whether you have numbers you can put into a renewal case. That single line is the entry point to our Metamon diagnostic. One line on the form (a team name is enough if you’re not sure). No plan document.
Separate whether the three metrics can be written →
References
Source: McKinsey, The State of AI (2025) — fewer than one in five organizations track clear generative-AI KPIs
Our services: Consulting services (CPP)
Our services: Metamon — generative AI adoption
Related: Three structural reasons AI agent projects stall at PoC
Related: For whom is a project plan a product?
Related: A renewal case with no usage number — what to cut first
Related: Scope the external assessment to teams you can’t measure in-house
This article is a general framing, not legal, regulatory, or contractual advice. Decisions on renewals, budgets, and contracts follow your own internal rules.