Back to Guides
Enterprise Software 15 min read

Inside the baseline: what to measure the week before training starts

Inside the baseline: what to measure the week before training starts

An AI training baseline is six items, and the week before the first session is the last moment to collect four of them cleanly. The other two are already compromised, because the session is on people’s calendars and everyone knows their work is about to be read. Collect them anyway, and write down what they are exposed to.

The guides that rank for this say to set a baseline. None of them says what day to do it, who pulls which number, or which items survive being collected late. Here is the schedule, the owner per item, and the vendor documentation that changes what your usage number means depending on the day you pull it.

What an AI training baseline has to contain

Six items. Each has one owner who is already closest to the system it comes from, so none of this needs a new role.

# Item Where it comes from Owner Time
1 Time on one named deliverable, three recent instances per person The function that produces it The function lead 45 min
2 One dated work sample per person, collected by date Shared drive or mail archive Whoever already has access 10 min each
3 Reach and habit: licensed seats, people active in the window Tool admin console IT or the tool admin 30 min
4 Safe-use snapshot: what people paste where, and whether they know the rule A five-question form, no names Policy owner 45 min
5 Cost basis: headcount in scope, loaded hourly cost by role, current license spend Finance Finance 60 min
6 The failure line: the result that would mean this did not work The sponsor The sponsor 45 min

Four to six hours in total for a group of 40. Scoring the samples is a separate method, laid out in our guide to scoring a team before and after; this page is about collection.

Completion rates and satisfaction scores are not on the list. Both run near-universal and predict nothing. If your learning system produces them at no cost, keep them off the page that goes to finance.

The two items you cannot rebuild later

Items 3 through 5 are recoverable. A console report pulled in March still shows you February, finance can hand you a loaded hourly rate any week of the year, and the safe-use form can be re-sent. Items 1 and 2 cannot be recovered, and they are the two the whole case rests on.

The before-time on a deliverable is gone the moment people know what answer is wanted. Ask someone in April how long a client update took them in February and you get a number shaped by what they now think of the training. No archive exists to check it against, because nobody was timing anything.

An uncontaminated work sample is gone faster still. Once a person knows their output is being read against a standard, they hand you the version they think the standard wants. Take the most recent instance by date, from the drive, yourself.

Item 6 is a third case. You can write a failure line in March, and it will not be evidence, because a threshold set after you have seen the result is a sentence about the result. Date it before the first session.

The console lag that decides what your usage number means

The obvious source for reach and habit is whatever admin view ships with the tool. Pull it, then read the fine print on the window, because the number is not about the day you pulled it.

Microsoft’s documentation for the Copilot Dashboard says it plainly: the readiness, adoption and impact pages show the previous 28 days, with a delay of up to six days. Newly licensed people take up to seven further days to appear at all.

So a report pulled on the Friday before a Monday session covers roughly day minus 34 through day minus 6. Whether that window is clean depends on when you announced the program. Announce on Monday and pull on Friday, and the window is almost all pre-announcement. Announce three weeks out and two thirds of that window already contains people who knew training was coming.

Write the window’s end date on the record, not the pull date. If the announcement falls inside the window, say so and treat the number as a floor.

Two more constraints from the same document:

  • Small groups return nothing. Metrics are suppressed for any group below the minimum group size your own administrators set, so a team under that line comes back as a blank row. Decide your reporting groups before you pull, and check where the line sits.
  • Two Microsoft reports will disagree. The dashboard uses a 28-day window with up to a six-day delay. The admin center uses rolling 7, 30, 90 or 180-day windows and produces within 72 hours. Same month, same people, two numbers that do not reconcile. Pick one source, record which, and never switch between the before and the after.

One metric there is not a measurement. Copilot assisted hours is computed from “multipliers derived from Microsoft’s research on Copilot users,” and the documentation calls it a general estimate whose “underlying calculation methodology will evolve over time.” That is a vendor’s model of your company, and it can change between your before and your after. Use it as color, not as a business case.

Three ways measuring changes what you measure

Pre-measurement is not a neutral act, which is the part the metric menus leave out.

The announcement changes behavior. The calendar invite tells everyone that AI use is now something leadership looks at, and usage rises before a single session runs. This is what makes the week-before pull the last defensible window rather than a good one.

The scale moves under you. George Howard’s work in Evaluation Review in 1980 established response-shift bias: training changes the standard people rate themselves against, so a traditional pre-test and a retrospective pre-test of the same period disagree in a systematic direction. One leadership-training study found self-ratings higher on the traditional pre-test than on the retrospective one across all 24 areas measured, and practitioner guidance on the post-then-pre design covers the trade-off. The buyer’s version is narrower: the design you pick moves the size of your reported gain, so pick it before you collect anything and name it in the report.

Polished work gets handed over. People volunteer their best artifact, not their median one. Collecting by date, yourself, is the countermeasure.

None of the three is a reason to skip the baseline. Each is a reason to write one line next to the number naming what it is exposed to. A stated weakness survives a finance review; a buried one does not, which is most of what our list of what a CFO should ask first is built around.

The five-day schedule

Five working days, Monday to Friday, with the session the following Monday.

Monday, scope and cost basis. Finance gives you headcount in scope, loaded hourly cost by role band, and current license spend. One meeting, one sheet. This comes first because it decides whether the program is worth measuring at this size.

Tuesday, name the deliverable. Sit with the function lead and pick one recurring artifact the team genuinely produces: a client update, a case summary, a weekly report. One, not three. Agree how its time is counted and whether review and rework count inside it. They should; a saving measured on the drafting step alone is the most common way a gain evaporates later.

Wednesday, collect the samples. Pull the most recent instance of that artifact per person, by date, from the drive, and record the date with each. Nobody is asked to submit anything.

Thursday, pull the console reports. Both numbers and both window end dates, on the record. Confirm your reporting groups clear the minimum group size before you need the breakdown.

Friday, safe-use snapshot and the failure line. Five anonymous questions on what people paste where and whether they can state the rule. Then the sponsor signs the threshold.

Teams whose output has no billable clock need a different item 1, and the substitutions are in our note on measuring gains without billable hours. The schedule does not change.

The hourly figure finance will actually use

A salary figure captures about 70% of what an hour of work costs. The missing third is a gap a CFO will find on their own if you leave it.

The Bureau of Labor Statistics Employer Costs for Employee Compensation release put private-industry total compensation at $46.89 per hour worked in June 2026: $32.82 in wages and salaries and $14.07 in benefits, with benefits at 30.0% of the total. Use the loaded figure and cite the release date, or use your own fully loaded internal rate and say which.

One detail to settle before the after-measurement lands. Starting with December 2026 data, the BLS release stops including workers’ compensation costs, so a program whose baseline sits on an earlier release and whose follow-up sits on a later one is comparing two slightly different definitions. Freeze the figure you used, in dollars, on the baseline record.

Write the failure line before anyone sits down

One sentence, dated, signed by whoever authorized the spend: the result that would mean this program did not work. It has to name a number, a date, and a consequence. “If median time on the client update has not moved by 20 November, we do not buy the second phase.” Not “if adoption is disappointing.”

This is the cheapest item on the list and the one most often skipped, because it is the only one that can make the sponsor look wrong later. That is what makes it worth something. A baseline with no pre-committed threshold is a number you can narrate either way in March, and a reader in finance knows it. The wider method this sits inside, including what the before-and-after delta can honestly claim, is in our playbook on proving capability moved.

If the training already happened

Do not reconstruct a baseline from memory. It will not survive one question, and asking it burns the credibility you need for the next budget cycle. Two options do work:

  1. Baseline a group you have not trained yet. Time and sample an untrained team this week, train that team next, and you have a real before-number from a comparable population. A genuine baseline from a different team beats an invented one from the same team.
  2. Start the clock where you are. Recover items 3 through 5, which do not care that you are late, declare today as the baseline date in writing, and measure forward. You lose the claim about the first phase and keep the ability to make one about the next.

Both beat a retrofitted number, because both can be checked. The case for treating capability as something you build and verify rather than something a license purchase delivers runs through our argument against buying licenses and calling it capability.

Frequently asked questions

What is an AI training baseline?

A dated record of six things, taken before an AI enablement program starts: time on one named deliverable, one work sample per person, seats and active users from the tool console, a safe-use snapshot, the cost basis from finance, and a signed failure line. Its only job is to make a later claim checkable, which means every item carries the date it was collected and the window it covers.

Is the week before too late?

It is the last defensible window, not the best one. By then the session is on calendars, so the usage numbers already include people who know training is coming. The two items that need to be earlier are the deliverable timing and the work samples, ideally collected before the invite goes out. If you are already inside the week, collect everything, note the announcement date next to the usage figure, and treat that figure as a floor.

Can the admin console give us the baseline on its own?

No. It reports seats and activity, not capability, so it cannot tell you whether anyone’s judgment improved. Its windows are also rolling and lagged: the Microsoft Copilot Dashboard shows the previous 28 days with up to a six-day delay, while the admin center uses 7, 30, 90 or 180-day windows produced within 72 hours. Two numbers for the same month that will not match. Pick one source and record which.

Should we use a self-assessment survey as the baseline?

Not as the only item. Self-ratings are cheap and loosely tied to performance, and the design changes the answer: response-shift bias means a traditional pre-test and a retrospective pre-test of the same period disagree, with traditional ratings running higher. Use a survey alongside a dated work sample, pick one design, and name it in the report. A number whose method is stated can be argued with; one whose method is unstated gets discarded.

How long a window should the baseline cover?

Match it to whatever your source already produces and write the end date down. For usage, that is the console’s own window rather than a span you choose. For deliverable timing, three recent instances per person beats one, because a single instance mixes the person’s speed with that week’s difficulty. A monthly artifact therefore means three months of lookback, which is a reason to start earlier rather than settle for one data point. Check too that each reporting group clears your tool’s minimum group size.

Who owns the baseline when there is no AI lead?

Assign each item to whoever already touches that system: finance for the cost basis, the tool admin for the console pull, the function lead for the deliverable, the policy owner for the safe-use snapshot, the sponsor for the failure line. One person coordinates and nobody takes on a second job. Four to six hours for a group of 40, spread across five people.

What hourly rate should we use to price the hours?

A fully loaded rate, not salary. The BLS Employer Costs for Employee Compensation release put private-industry total compensation at $46.89 per hour worked in June 2026, of which $14.07 was benefits, so a salary-only figure captures about 70% of the real cost. Use your own loaded internal rate if you have one, and freeze the dollar figure on the baseline record, because the published series changes definition with the December 2026 data.

What has to be written down for the after-measurement to be comparable?

For every item: the number, the source system, the window start and end dates, the collection date, and the announcement date. For the deliverable: whether review and rework were counted inside the time. For a survey: which design. Six numbers and their conditions fit on one page. A number without its conditions cannot be re-measured the same way, which makes the second measurement a new number rather than a comparison.

Key takeaways

  • An AI training baseline is six items with six owners and costs four to six hours for a group of 40. None of it needs a new role.
  • Three items recover late: the console pull, the safe-use form and the cost basis. The before-time on your named deliverable and an uncontaminated work sample do not, and the failure line stops being evidence once it is written after the result.
  • Console usage numbers describe a lagged rolling window, not the day you pulled them. Record the window’s end date and the announcement date side by side, and treat any vendor metric labeled as assisted hours as a model output rather than a measurement.
  • Price the hours on fully loaded compensation rather than salary, and freeze the figure in dollars.
  • If the training already ran, baseline an untrained team or declare today as the start date. Do not reconstruct a before-number from memory.

Working out what to measure before a program starts, and whether the one you have been quoted is worth measuring at all, is most of what a first conversation covers. Book a free AI-readiness call and we will walk through your six items against the deliverable you would want to claim.

Last Updated: Nov 13, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles