Every training metric you could be asked for belongs to one of three tiers, and the tier decides whether you can answer at all. The reason your numbers never satisfy the person who asked isn’t that you picked the wrong metrics. It’s that the question you were handed and the tier of data that could answer it are two different things, and nobody told you which tier you were in.
Your budget conversation is in two weeks. The difference between a number that holds and a number that falls apart under one follow-up question is which tier it came from. This guide sorts the six numbers your platform already has, names the four systems that hold the rest, settles the denominator argument before it starts, and says plainly which question has no software answer.
You can run the training these numbers are about on the same system that records them.
Key Takeaways
- Six numbers come out of a course platform on their own. Two of them change what you build. The other four are attendance.
- A completion rate has four possible denominators, and the same group can score 51% or 76% depending on which one you pick. Fix yours in writing before the first report.
- The number your director wants usually lives in the HRIS, the CRM or the ticket queue. Getting it needs a shared ID and an agreed date window, not a new license.
- Isolated financial return is not a report any platform produces. Agree a proxy with the person who owns the business number before the rollout, not after the question.
Every Training Metric Sits in One of Three Tiers
Before you build anything, sort the request. It takes thirty seconds and it’s free.
| Tier | What it is | What it costs you | How long it takes |
|---|---|---|---|
| One | What the course platform watches by itself | Nothing. It’s already recorded | This afternoon |
| Two | What a system you already own records, once you give it a shared ID | A conversation with whoever owns that system, and an agreed date window | Two or three weeks, most of it waiting |
| Three | What nothing records, because separating a training effect from everything else is a study-design problem | A quarter, and you still won’t have it | Not at your sample size |
A list of fifteen metrics tells you what to name. The tier tells you whether to promise it. Most published lists mix the three freely, putting completions beside “behavior change” and “return on investment” as though they were three sizes of the same thing. One is a record you already hold, one needs a second system and a person who owns it, and one needs a study you won’t run.
The tier also decides what you say in the meeting, in the room, that day:
- Tier one. “I’ll have it Thursday.”
- Tier two. “I need thirty minutes with whoever owns the HRIS, and a date we both agree on.”
- Tier three. “That one can’t be proven at our size. Here’s what I can give you instead.”
A clean no in week one costs less than a quarter of failing to answer. For the standard reports a platform ships and what each one can tell you, the companion to this piece is which reports a platform ships.
Tier One: The Six Numbers a Course Platform Gives You on Its Own
The boundary for tier one is simple: what can the course software watch? It can watch what happens inside a lesson, on its own screens. That’s six numbers, not fifteen.
| The number | What it tells you | What it doesn’t tell you |
|---|---|---|
| Completions | Who finished, who’s in progress, who’s overdue | Whether anything changed on the floor |
| Time on task | How long a lesson takes in practice, so you can plan a shift around it | Whether that time was attention or confusion |
| Question-level performance | Which question the group gets wrong, which usually means which lesson is unclear | Whether they can do the thing away from a quiz |
| Drop-off point | The lesson where people stop | Why they stopped |
| Certificate currency | Who holds a credential that hasn’t expired | Anything about the people who never started |
| Per-account rollups | Which team, site or customer is behind | Which one is at risk |
Here’s the part the metric lists leave out. Two of these six change what you build. The other four are attendance.
Question-level performance tells you which lesson is broken. If eight in ten people miss the same question, the problem is upstream of the learner. The drop-off point tells you where the course got too long or too abstract, which is the one edit that reliably raises everything else. Those two are worth acting on this week.
Completions, time on task, certificate counts and rollups tell you who showed up. Report them. Don’t mistake them for evidence that the training worked.
Time on task needs a warning of its own. A long time on a lesson reads equally well as careful attention or as total confusion, and both readings are defensible in the room. That’s what makes it useless on a slide and useful beside the question data: long time plus a failed question means the lesson is unclear.
Mini Course Generator’s own reporting sits in this tier: learning analytics and learner tracking, certificates and micro-credentialing, and account-level rollups on the larger plans. That’s a reporting surface, not an analytics platform, and nothing in tier one is.
Fix the denominator before you report a completion rate
A completion rate is a fraction, and the argument is always about the bottom half. Four denominators are in common use, and they’re all defensible:
- Assigned: everyone the course was assigned to.
- Started: everyone who opened it at least once.
- Eligible: everyone in scope, whether or not the assignment reached them.
- Headcount on the day: everyone on the payroll when you pulled the report.
An example of the spread. Say 400 people work at the site, 350 are in scope, 312 got the assignment, 268 opened it, and 204 finished. Same people, same week:
| Denominator | The bottom half | Completion rate |
|---|---|---|
| Started | 268 opened it | 76% |
| Assigned | 312 got the assignment | 65% |
| Eligible | 350 in scope | 58% |
| Headcount on the day | 400 on the payroll | 51% |
Twenty-five points apart, and not one of those rows is wrong.
The one your director remembers is whichever you reported first. So pick the denominator before the first report and put it in the footer of every report after that: “Completion is finished divided by assigned, measured on the due date.” When someone challenges the number six months later, and they will, that footer is the whole defense.
For compliance, report currency instead of completions
For anything with a legal consumer, “how many people we trained this year” is the wrong number. An auditor doesn’t want a total. They want to know who, today, holds a credential that hasn’t expired.
The two move in opposite directions. Your completion total only ever goes up. Currency falls every month on its own as certificates age out, which is how a program that looks healthy on completions can be failing an audit at the same time.
Report the currency figure with the date it was true. If that’s the whole reason you were asked for numbers, compliance training records covers the delivery side.
When Your Course Runs in Someone Else’s System, the File Format Sets the Ceiling
If you export your course as a package and load it into a corporate LMS, the format you exported decides what can be reported back. Not the vendor. Not the dashboard.
A SCORM package reports the small set the standard carries: completion, pass or fail, a score, time. That’s most of tier one, and for a lot of training it’s enough. xAPI exists because that set was too small for anything that happens outside a lesson page. It’s an IEEE standard, IEEE 9274.1.1-2023, published in October 2023, and it defines a JSON statement model plus a web interface to a Learning Record Store, which is a system built to hold that stream of statements.
So “can your platform report X?” is often a question about the file, and asking the vendor to fix it is asking the wrong party. If the package you exported can’t carry the field, no amount of pressure will produce it, and the fix is a different standard plus somewhere to put the data.
That’s a project, not a support ticket. Check which side of that line your request falls on before you escalate. The differences are covered in which standard carries which data.
Mini Course Generator sits on the SCORM side of this. It exports SCORM, and its SCORM Upload Block accepts a package you already have from another system and tracks how learners interact with it inside the course.
Tier Two: The Number You Want Is Already in a System You Pay For
Here’s the good news about tier two. The outcome number your director asked for usually exists already, in a system your company pays for. What’s missing is a shared identifier and a date window that both sides agree on.
| The system | The outcome number it holds | The join key | Who owns it |
|---|---|---|---|
| HRIS | Time to productivity, retention in role, internal moves | Employee ID | HR operations |
| CRM | Ramp to first closed deal, attach rate, discount depth | The rep who owns the opportunity | Sales operations |
| Ticket queue | Tickets per account, first-contact resolution, reopens | The requester’s account | The support lead |
| QA scorecard | The score on the behavior you trained | Agent ID | QA or the team lead |
A join key is a field both systems store for the same person or account, so two spreadsheets line up without guessing. Email addresses work until someone changes their name. Employee IDs work until a contractor has none.
That table is your meeting agenda. Take it to the owners, ask which key they trust and what they’d need from you, and within a week you’ll know whether the number is three weeks away or not available at all. No new license appears in that sentence, which is the point.
Agree the window before you pull the data
Decide the before-and-after window in advance: thirty, sixty or ninety days after the rollout, and the same length before it.
Picking the window after you’ve seen the result is the single thing that makes a training number unbelievable to a finance reader. They’ve done it themselves, they know what it looks like, and one raised eyebrow at that table costs you more credibility than a mediocre number ever would.
Ask what event the platform emits, not what dashboard it draws
Put this question to any vendor, including us. A dashboard you can’t export is a screenshot. An event that lands in the HRIS or the CRM the moment it fires is the join in the table above, and it’s the difference between a number you produce every quarter and a number you rebuild by hand every quarter.
Ask it in those words: when someone finishes a course, what do you send, where can you send it, and what’s in it? Mini Course Generator sends webhooks and connects through Zapier and Make, with a REST API and Headless LMS on the top plan. If the answer is “we have a great dashboard”, that’s a system which will never feed tier two. The wiring is covered in connecting training to the rest of your stack.
If your audience is customers rather than employees, the outcome numbers change shape: activation, time to value and ticket deflection instead of time to productivity, and they’re covered in the customer-side version of these numbers.
Tier Three: Isolated Return Is Not a Reporting Feature
Now the honest part. When someone asks you to prove that the training caused a business result, they’re asking for a study, not a report.
Think about what else moved in the same quarter. A new manager started. Pay bands changed. It’s the busy season. A product release shifted what the job involves. Three experienced people left and two joined. Your sample is one team and your control group is the same team three months ago, and every one of those things moves the number you’re being asked to pin on a course.
More data doesn’t fix that. A bigger dashboard doesn’t either. A different study design would, with a real control group and a sample you don’t have, and you won’t be running one before the budget meeting. The fuller version of that argument, and why the newest tools don’t change it, is here: why more data doesn’t fix attribution.
Saying this out loud in week one is cheaper than a quarter spent failing to answer, and it moves the conversation somewhere useful. The question behind “what’s the training ROI” is usually “should I keep funding this”, and that one has an answer.
The agreed proxy, in four parts
A proxy is a business number you and one other person agree to watch together, knowing it isn’t proof. It survives a follow-up question when it has four parts, and only when it had all four before the rollout:
- The business number. One number, named, that already exists in a system somebody owns.
- The window. Fixed in advance, same length before and after.
- The owner. The person whose number it is, who agreed to watch it with you.
- The written line. One sentence in the report saying this is a directional read, not an attribution.
An example, written down: a support team is about to get a new product course, the agreed proxy is tickets per account in the ninety days after rollout against the ninety before, the support lead owns that number and signed up in advance, and the report carries a line saying other things changed in the same period.
Timing is what makes it work. A proxy agreed before the rollout is a plan. The identical proxy produced afterwards is a defense, and everyone in the room can tell which one they’re looking at.
Nobody reports a skills gap from a learning platform
This one gets asked a lot, so here it is plainly. No learning platform reports a skills gap. Ours included.
A skills-gap dashboard needs three things: a skills taxonomy, a competency map that says which role needs which skill at which level, and workforce data about who currently holds what. Mini Course Generator has no skills taxonomy, no competency mapping and no workforce-planning data. A course platform that shows you a skills-gap chart is displaying a spreadsheet somebody on your side maintains by hand, and the moment that person stops maintaining it, the chart goes wrong without telling you.
So if the request is a skills gap, the honest answer is that it’s a tier two job at best and usually tier three, and the data lives in the HR system and in your managers’ heads, not in the training system. Say that in the meeting. It’s a much better position than producing a chart you can’t stand behind in six months.
Translating the Model You Were Handed
Somebody has probably handed you the four-level evaluation model: reaction, learning, behavior, results. It’s a vocabulary for talking to executives. It is not a measurement plan, and the tiers are what turn it into one.
The mapping is one line. Reaction and learning are tier one, and they’re cheap. Behavior is tier two, and it costs you a join key and an agreed window. Results is tier three, and at your sample size it isn’t obtainable, whatever the model implies.
That’s why “we need level four this quarter” is a request to translate, not a deadline to accept. The translation is one sentence: “Level four means an isolated business result, which we can’t separate from everything else that happened. What I can give you is an agreed proxy on a number you already own, watched over ninety days, with both of us signed up to it now.” That gets accepted more often than it gets refused, because what they wanted was a way to talk about whether the money was worth spending.
One more request to decline while you’re here. Popular rules of thumb about how much learning happens on the job rather than in a classroom get quoted as if they were measurements. They were never measured the way they’re quoted, and their most common use is to explain why the numbers are missing. Don’t build a report around one, and don’t accept one as an explanation when you ask where the evidence went.
Cost Per Learner Moves for Reasons That Are Not Learning
Cost per learner is the one number in this guide a finance reader takes without translation. It’s also the one most likely to be misread, because it moves for reasons that have nothing to do with training.
Two pricing mechanics dominate, and they create opposite illusions. Per-seat pricing makes your cost per learner fall steadily as headcount grows, then spike the month a true-up lands. Nothing about the training changed in either direction. A yearly learner allowance makes it fall through the year as more people take a course against a fixed fee, then reset in January. Again, nothing changed.
Mini Course Generator uses the allowance model, described this way on the pricing page: “A learner is anyone who takes your courses. Each plan includes a yearly learner allowance, so occasional visitors don’t force you into a bigger plan.” Whichever model you’re on, the reporting rule is the same.
So put the mechanic in the same line as the number. “Cost per learner fell by a third, because headcount grew against a fixed annual fee” is a sentence that can’t be misread. The bare number invites congratulations in a good quarter and blame in a bad one, and you’ll be handed both for something you didn’t cause. For what each plan includes, see the plans and what’s in them. For the rest of the bill, including the parts that aren’t the subscription, see what a learning platform costs to run.
The One-Page Report
Here’s the document. One row per program, six columns, and the last column is why you get asked back next quarter.
| Program | Who it was for | Denominator | Tier-one number | Agreed proxy | What we changed because of a number |
|---|---|---|---|---|---|
| Example: New hire product course | 38 support hires, Q3 | Finished / assigned, on the due date | 82% complete, 4 overdue | Tickets per account, 90 days, owned by the support lead, directional | Split lesson 4, where 40% stopped. The failed pricing question fell by half |
That last column is the one that changes the meeting. It’s proof you read the numbers and did something, which is what your director is trying to find out. A page of percentages with no action attached reads as activity. One line of “we saw this, so we changed that” reads as management.
Keep the shape fixed: the same six columns every quarter, for every program, even when a cell says “none agreed” or “no change yet”. A report whose shape changes can’t be compared with itself, and comparison is most of its value by the third quarter.
The Short Version
Sort the request into a tier before you do any work. Tier one is six numbers your platform already holds, two of them worth acting on. Tier two is a number a system you already pay for already has, waiting on a shared ID and an agreed window. Tier three is a study you won’t run, and saying so early is a stronger position than a chart you can’t defend.
Do this first, for the next program that launches: write down the denominator and the agreed proxy before it goes live, and send both to the person who’ll ask you for the numbers. Two sentences in an email, and it’s the only part of this that has to happen before the rollout rather than after.
Once the denominator is written down, you can assign the training and let the platform keep the record instead of rebuilding it by hand each quarter.
Frequently Asked Questions
What should I report if all I have is completion data?
Report it with the denominator stated in the footer, add the overdue count, and add one line naming the proxy you’re going to agree for next quarter. Completion plus a named denominator plus a forward commitment is a credible report. Completion alone invites the follow-up question you can’t answer.
How do I prove training reduced turnover?
You don’t prove it. Retention moves with pay, managers, the job market and the season, and your sample is too small to separate a course from any of that. Go to whoever owns the retention number, agree on a window before the rollout, and report it together as a directional read.
Which training KPIs should go on a quarterly slide?
The six columns from the one-page report: program, audience, denominator, the tier-one number, the agreed proxy, and what you changed because of a number. The list is short on purpose. A slide with fifteen metrics on it gets read as noise, and none of them stick.
How do I get training data into our HRIS?
Ask your platform what event it emits when someone finishes, and ask HR operations which identifier they’ll match on, usually employee ID. If the platform sends a webhook and HR accepts an employee ID, the connection is a small piece of plumbing rather than a project. If the platform only draws a dashboard, that’s your answer.
Sources
- IEEE 9274.1.1-2023, the xAPI standard, published October 2023, for the statement model and the Learning Record Store
- Mini Course Generator plans and what each one includes, for the learner-allowance definition and the reporting included with every plan
