Run both, with different jobs. A monthly read on an AI enablement program is an operating check for the person who owns it: a few numbers, no audience above the person who can act this week. A quarterly read is the decision document — it goes to finance, carries the dated evidence, and settles renew, renegotiate or stop.
Pick one and you lose something specific. Monthly reporting hides meaning, because at the group sizes a 50-to-500-person company trains, a month-to-month percentage move is mostly sampling noise and gets read as a trend anyway. Quarterly reporting hides the window where intervention was still cheap, because habits form and collapse inside a quarter and the first report lands after the fact.
Neither failure is a calendar preference. Both are measurement problems with numbers attached, and the numbers are what every cadence template circulating on this topic leaves out.
The two cadences, side by side
| Monthly | Quarterly | |
|---|---|---|
| What it sees well | Direction of weekly use, support questions, who has gone quiet | Depth, breadth across functions, one dated output change |
| What it hides | Whether the move is real at your group size; anything that takes a quarter to appear | The weeks when a fading program could still have been rescued |
| Reliable at | Counts and events: sessions run, workflows named, questions logged | Rates and outcomes: sustained use, time on a named deliverable |
| Right audience | The program owner and function leads | CFO, executive team, whoever renews the spend |
| Decision it supports | Run a clinic, chase a quiet team, fix a blocked workflow | Fund, defund, renegotiate or stop |
| Typical failure | A six-point move triggers a reorganization of the program | The first honest look arrives in month four |
That is a division of labor, not a ranking. The monthly report prompts small corrections by the person already doing the work. The quarterly report has to survive someone outside the program poking holes in it, which is the test our measurement guide is built around.
What monthly reporting hides
Monthly reporting hides whether anything happened. It is a precise instrument pointed at a quantity that does not move that fast, and three things go wrong.
The number moves for reasons that are not the program. A month holding a quarter-end close, a four-day week or a product launch produces a figure about the month, not about capability. Finance teams in close do not experiment.
The denominator drifts. Two people left, six joined, three transferred in. A monthly percentage computed over a shifting roster is a different measurement each time, and the shift is usually larger than the effect being tracked.
Monthly cadence selects for the cheapest metric. Something has to fill the page every month, and the only numbers that reliably move that fast are seat counts, logins and session totals. The report drifts toward activity and away from capability, which is the gap our case against treating licenses as capability was written about. Nobody chose that drift; the cadence chose it.
The noise floor nobody publishes
Here is the arithmetic every dashboard page skips. Weekly active use among a trained group is a proportion, so it carries a sampling error set by group size. For a proportion near 50%, the 95% bound on the difference between two independent monthly readings works out like this.
| Trained group | Standard error on one reading | Month-to-month move that is indistinguishable from noise |
|---|---|---|
| 25 people | ±10.0 points | Up to 28 points |
| 40 people | ±7.9 points | Up to 22 points |
| 80 people | ±5.6 points | Up to 16 points |
| 150 people | ±4.1 points | Up to 11 points |
At 40 trained people, a drop from 52% to 44% is not a decline. It is the same program measured twice. Comparing the same roster month over month tightens the bound, because the people are not re-drawn, but not nearly enough to rescue a six-point move at that size.
Two rules follow. Never put a monthly rate in front of an executive at a group under about 80 people without stating the bound. One confident percentage with no error attached invites the first skeptic to dismiss the program when next month’s figure falls back.
And report counts monthly, rates quarterly. Workshops delivered, workflows named, support questions logged, people quiet for three straight weeks: four workflows named is four workflows named, with no confidence interval to argue about. Percentages belong in the quarterly document, where the comparison is a quarter against a quarter.
The depth number you cannot compute monthly
This one is a hard constraint rather than a judgment call, and it settles most cadence arguments.
The best published definition of depth comes from Microsoft’s analytics reference, and it has a twelve-week window built in. A habitual user averages 1 to 14 assistant actions a week and was active in at least 9 of the past 12 weeks; a power user clears 15 or more actions a week on the same nine-of-twelve test.
You cannot compute a nine-of-twelve-weeks metric monthly. A monthly depth figure is therefore either a different metric wearing the same name or an invention. Most of the monthly depth numbers in the review decks we read are four-week averages that one enthusiastic week can carry.
The same reference gives one legitimate high-frequency signal: the share of this week’s active users who also used an assistant the previous week. Week-over-week return is a real weekly number and the honest monthly stand-in. Report return rate monthly as a trend and sustained use quarterly as a level, and both hold up.
What quarterly reporting hides
Quarterly reporting hides the only period when a weak program is still cheap to fix.
Habit formation is the clock that matters, and it has been measured. A 2024 meta-analysis in Healthcare covering 20 studies and 2,601 participants found median times to habit formation of 59 to 66 days, with means of 106 to 154 days. Those medians sit inside the first quarter. A program that reports for the first time at week 13 gets its first honest look after the window in which a refresher clinic or a manager conversation would have changed the outcome.
Three more things hide in a quarter.
Function-level divergence. The aggregate can hold flat while operations climbs 20 points and finance falls 20. One number, two opposite stories, no action available.
Who went quiet. People stop in a recognizable way: a few lighter weeks, then nothing. On a monthly list of names that is obvious and recoverable. In a quarterly rate it is already baked in.
The follow-up nobody scheduled. Most enablement work needs a second pass where people hit the limit of what the first session taught, and finding that point takes more than four looks a year. It is also how you avoid the trap in our rules for evaluating a pilot before scaling it: a program judged once, and judged late.
Decay moves faster than a quarter
Decay is the specific failure that quarterly reporting cannot see, and it is not hypothetical.
Recon Analytics surveyed more than 150,000 US paid AI subscribers between July 2025 and January 2026 and watched one major assistant’s share of those subscribers fall from 18.8% to 11.5% across seven months, a 39% contraction. The same work measured how many people with workplace access convert into actual users: 83.1% for ChatGPT, 35.8% for Copilot, 34.0% for Gemini.
Sample that slide twice and you get two points on a line. You cannot separate a decline from a bad fortnight, and you cannot tell when it started. Federal statistical practice handles this by pairing windows instead of choosing one. The Census Bureau’s business survey, fielded every two weeks, asks firms whether they used AI in the last two weeks and, in the same instrument, whether they expect to use it in the next six months.
That pairing beats the weekly-monthly-quarterly ladder the dashboard vendors recycle. One short window for behavior, measured often, read as a trend. One long window for outcome, measured rarely, read as a level.
The split that works: an operating read and a decision read
Here is the arrangement we recommend for a trained group of 30 to 100 people.
| Monthly operating read | Quarterly decision read | |
|---|---|---|
| Length | One page, five minutes | Three to four pages, a scheduled meeting |
| Audience | Program owner, function leads | CFO, executives, budget holder |
| Content | Sessions run; workflows newly named; week-over-week return rate as a trend; names of people quiet three weeks or more; blocked requests | Sustained use at nine of twelve weeks; spread between strongest and weakest function; one dated before-and-after on a named deliverable; cost against hours moved; renewal recommendation |
| Rates allowed? | Only with the bound stated | Yes, quarter against quarter |
| Decision it produces | A clinic, a conversation, an unblocked workflow | Fund, renegotiate or stop |
| Time cost | Under an hour to assemble | Half a day, once a quarter |
Two notes on proportion. A workshop program in the $2,000 to $15,000 market range does not justify a monthly executive deck, so put the effort into the one quarterly document somebody will act on. And the quarterly read is the one that needs a defensible financial model behind it, which is why we treat building an ROI model finance cannot dismiss as quarterly work rather than something assembled the night before a renewal.
Report to the decision, not to the calendar
The clearest rule in this field predates AI entirely. The Talent Development Reporting Principles, a learning-measurement standard modeled on accounting practice, hold that management reports should be produced at least quarterly and, in most cases, monthly — and that the reporting rhythm has to match the rhythm of the decisions. Report quarterly into a monthly decision cycle and the data lands too late to use. Report monthly into a quarterly cycle and you generate work nobody reads.
So start from the decisions. List every one this program could produce in a year — schedule a follow-up session, move a license, retrain a function, redirect freed hours, renew or cancel — and mark how often each can genuinely be made. The monthly report carries the inputs to the monthly decisions. Everything else waits for the quarter.
Most review decks fail this test the same way: ten numbers every month, of which two could change anything that month, and nothing the quarterly decision turns on.
Monthly theater and quarterly amnesia
Both cadences have a signature failure, and they are easy to spot once named.
Monthly theater is a report that exists to prove the program is being managed. Its tell is a chart with no decision attached, a green or red arrow on a move inside the noise floor, and a narrative written to explain whichever way the arrow pointed. The fix is subtraction: cut it to the three things that could change next week.
Quarterly amnesia is a report assembled in the week it is due from whatever the admin console still retains. Its tell is a baseline nobody wrote down, a definition that changed between quarters, and a comparison against published adoption benchmarks drawn from companies ten times the size. The fix is to set definitions and baselines before the program starts and collect monthly, so the quarterly document is assembled rather than reconstructed.
One habit prevents both: at kickoff, write a single page naming every metric, its definition, its window, who pulls it and which decision it feeds. Change it only deliberately, and note the change. Definitions that drift silently cannot produce a trend at any cadence.
Frequently asked questions
How often should we report on an AI enablement program?
Monthly to the program owner, quarterly to the budget holder. The monthly report carries counts and names, so corrections happen while they are cheap. The quarterly report carries rates, a dated output change and a renewal recommendation — the numbers a quarter of data can support.
Is monthly reporting on AI adoption a waste of time?
No, but monthly percentages usually are. At 40 trained people the 95% bound on a month-to-month move runs to about 22 points, so most monthly rate changes are noise. Count events instead: sessions delivered, workflows newly named, people quiet for three weeks, requests blocked.
What should go in a monthly AI enablement report?
Five items on one page: sessions or clinics run, workflows where an assistant became standard this month, week-over-week return rate as a trend, names of people who have gone quiet, and anything blocking use the owner cannot fix alone. No forecast, no color-coded arrows.
What should go in the quarterly report?
Four things: the share of the trained group clearing nine active weeks out of twelve, the gap between your strongest and weakest function, one dated before-and-after on a real deliverable, and program cost against hours moved. Close with a recommendation to continue, renegotiate or stop.
Why does our AI usage number jump around every month?
Partly sampling, partly the calendar. A group of 40 carries about eight points of error on a single reading, and closes, short weeks and launches suppress experimentation regardless of skill. Read three-month rolling figures for direction and keep single months for the counts.
Can we measure depth of AI use monthly?
Not with the standard definition, which requires a user to be active in at least nine of the past twelve weeks. The honest monthly substitute is the share of this week’s active users who also used an assistant the previous week, labeled as what it is.
How quickly should we expect to see an output change?
Later than reach, and usually after the first quarter. Capacity appears before anyone redirects it, so the month-three report should name one candidate deliverable and the month-twelve report should carry a measured one. Demanding a moved metric in quarter one produces a flattering number.
Does a quarterly cadence catch a program that is failing?
Only after the cheapest fix has expired. Habit formation runs to a median of 59 to 66 days, so the weeks when a refresher clinic would have worked are gone before the first quarterly report is written. The monthly list of who went quiet is the early warning a quarterly rate cannot give you.
Should the CFO see the monthly numbers?
Not as a standing report. A monthly rate at these group sizes will fall at some point for reasons unrelated to the program, and that month becomes the reference point in every later conversation. Give finance the quarterly document with its assumptions stated, and the definitions page whenever they ask for it.
Key takeaways
- Run both cadences with different jobs: a monthly operating read for the owner, a quarterly decision read for finance.
- At 40 trained people a month-to-month move of up to 22 points is indistinguishable from noise, so monthly reports should count events rather than quote rates.
- The standard depth metric needs a nine-of-twelve-week window and cannot be computed monthly. Week-over-week return is the honest stand-in.
- Quarterly-only reporting arrives after the habit window closes, with habit-formation medians of 59 to 66 days falling inside the first quarter.
- Decay runs on a sub-quarter clock: one published series contracted 39% in seven months, which two distant readings cannot detect.
- Set metric definitions and baselines at kickoff. Silent drift destroys more trend data than any cadence choice.
If you want a second opinion on what your current reports can and cannot support, book a free AI readiness call. We will look at your group size, your definitions and the decisions you need to make, and tell you which numbers are carrying real information and which are decoration.
Arthur Wandzel
