TL;DR
- Completion rate, satisfaction score and time to first task are the three metrics most likely to make a failing onboarding look healthy.
- The four worth tracking are elapsed time to working, unreported blockers found by asking, question composition, and ninety day retention read against a baseline.
- A number that rises when you add more onboarding is measuring effort, not outcome.
- Silence is not success. A quiet first week usually means problems were absorbed rather than avoided.
- Any metric that improves when people stop asking for help is pointed the wrong way.
- Measure composition rather than volume. What changed about the questions matters more than how many there were.
A company reports ninety four per cent onboarding completion, a satisfaction score of 4.6 out of 5, and an average time to first task of under two days. By every number it tracks, onboarding is working.
The same company has a third of new starters still asking, in week four, where things they were shown in week one actually live. Two of the last eight hires left within six months, and in both exit conversations somebody said a version of nobody really explained how things worked here.
None of the three numbers is wrong. They are all accurately measuring something, and none of them is measuring whether the thing worked. That gap is the whole subject of this article.
Best tools for HR Operations
The three metrics that mislead
Completion rate measures attendance. It says that an account belonging to a named person reached the end of a set of assigned items. It does not say they read, understood, retained or can apply any of it, and everybody involved knows that a meaningful share of completions are somebody clicking through at speed. The deeper problem is directional: completion rate goes up when you add more content, so a programme that doubles in length and holds its rate looks twice as successful while probably being worse.
Satisfaction scores measure politeness, especially early. Asking somebody in their second week to rate their onboarding asks them to criticise an organisation they have just joined, before they know whether criticism is welcome, through a form that may not be anonymous. The scores come back high almost everywhere, which is the clue. A metric with no variance is not measuring anything, and in this case it is measuring how new people behave rather than how onboarding performs.
Time to first task is the most actively misleading of the three. It sounds like a productivity measure and it mostly measures how trivial the first task was. A team that assigns a small, safe piece of work on day one scores well. A team that gives somebody something real in week two scores badly, and is probably onboarding better. Any metric that can be improved by making the first task smaller is not measuring what it claims.
| Metric | What it claims | What it measures | Fails because |
|---|---|---|---|
| Completion rate | Learning happened | Items were clicked | Rises when you add content |
| Satisfaction score | Experience quality | Willingness to complain | No variance, early timing |
| Time to first task | Productivity | Size of the first task | Improves by shrinking the task |
| Tickets closed quickly | Support quality | Interval after reporting | Clock starts late |
| Question volume falling | Self sufficiency | Whether people gave up | Improves when people stop asking |
The four worth tracking
Elapsed working hours to operational. From the contracted start time to the first successful use of each tool the role genuinely needs. Report the slowest tool rather than the average, because one blocker is enough and an average conceals it. This number is unambiguous, it is owned by a specific function, and it is the only one of the four that can be fixed quickly.
Two details decide whether this number is honest. The clock must start at the contracted start time rather than at first login, because starting it at first login removes the entire period where somebody could not log in, which is the failure you most want to see. And the tool list must be the ones that genuinely block work, decided in advance per role, because a list that includes everything somebody might eventually want produces a number dominated by things nobody was waiting for. Three or four tools per role is usually right, and the discipline of choosing them forces a conversation about what a person actually needs on day one that most teams have never had explicitly.
Unreported blockers found by asking. At the end of week one, ask every starter a direct question: what could you not do that you did not mention. Count the answers. In most organisations the first time anybody does this the number is uncomfortable, and that discomfort is the point. This metric is unusual in that you want it to rise initially, because the rise is your measurement improving rather than your service declining.
Question composition, before and after. Take the ten most common questions from a baseline period. Later, check how many of those ten still arrive. If seven have gone and what remains is contract specifics, culture questions and genuine edge cases, something worked. If all ten still arrive, nothing did, regardless of what any other number says.
Ninety day retention against a baseline. Not as a headline number, which is too noisy at small volumes to mean much, but as a trend with the reasons attached. The value is in the pattern across several cohorts rather than any single figure, and it is the only one of the four that connects onboarding to an outcome anybody outside the function cares about.
What these four have in common is worth naming, because it is the property that makes a metric useful here. Each of them can get worse when you do something that looks like an improvement, and that is exactly why they are trustworthy. Elapsed time rises if you add a sensible approval. Unreported blockers rise when you ask better questions. Question composition can show that a change helped one category and hurt another. Retention can fall for reasons that have nothing to do with you. A measure that only ever moves in the flattering direction is not measuring the thing it claims, and the three misleading metrics above all share that property: there is no realistic improvement to onboarding that would make a completion rate or a satisfaction score go down. Before adopting any fifth metric, ask what would have to happen for it to worsen, and if you cannot answer, leave it out.
The measures that work differently for remote starters
Distributed hiring changes what several of these numbers mean, and applying office-era thresholds to remote starters produces confident wrong readings.
Elapsed time to operational is the clearest case. In an office it is bounded by how fast somebody can walk a laptop to a desk, so a bad day is a bad morning. Remotely it is bounded by shipping, which means a bad case is a bad fortnight and no process improvement touches it. Reporting the same target for both populations makes the remote team look incompetent at something they do not control. Report them separately, and separate the shipping component explicitly so the part that is genuinely yours is visible.
Unreported blockers are both more common and harder to detect remotely, because the office version of reporting a problem is often not reporting at all. It is somebody at the next desk noticing a frustrated expression and asking. Remove that and the only route left is the formal one, which is exactly the route a new starter is least willing to use. The direct question at day five does more work in a distributed team than in a co-located one, and skipping it costs more.
Question composition shifts too. Remote starters ask fewer questions overall and a higher proportion of them in writing, which looks like better self service and frequently is not. The questions that disappear are the small ones that would have been asked verbally, and those are the ones that previously provided the informal signal about how somebody was getting on.
The practical consequence is that a distributed team needs more scheduled contact to generate the same information an office produced incidentally. That is an argument for the conversations described earlier rather than for more tooling, and it is the part most often cut when a team is busy, because it looks like overhead rather than like measurement.
Why silence is the trap
The single most common misreading in this area is treating a quiet first week as a good first week.
New starters under-report by default, and the reason is structural rather than cultural. Raising a problem in week one means declaring in writing that you cannot do your job, at the exact moment you are least sure how that reads and least able to judge whether the fault is the system or your own misunderstanding. Most people wait, ask a colleague, and work around it.
So a process that generates few tickets may be working well or may be generating exactly the same problems while making people less willing to mention them. Those two states look identical in every dashboard, and they are opposite.
This is why the second metric above exists and why it has to be a direct question rather than an open invitation. Asking whether somebody needs anything produces a polite no. Asking what specifically they could not do and did not report produces a list, reliably, from people who had no intention of raising any of it.
The organisational consequence is worth preparing for. When you start asking, reported problems rise, and somebody reading the numbers without context will see a decline. Warn them in advance, because the alternative is a conversation where the honest answer sounds like an excuse.
The metric that is pointed backwards
Falling question volume is treated almost universally as a success signal and it is ambiguous at best.
Questions can fall because people can find answers, which is good. They can also fall because people have concluded that asking is slow, unwelcome or embarrassing, and have started guessing instead. The second produces a better number and a worse company, and nothing in the volume figure distinguishes them.
The distinguishing test is composition, which is why it is on the list above. If the repetitive, documented questions have gone and the judgement questions remain, people are self serving. If both have gone, people have stopped asking, and the next evidence will be errors rather than questions.
There is a cheap confirmation available. Pick three people from the most recent cohort and ask them directly whether there was anything they wanted to ask and did not, and why. Fifteen minutes, and it settles which of the two situations you are in better than any amount of analysis.
The general principle is worth keeping: be suspicious of any metric that improves when people ask for less help. Help seeking is usually a sign of engagement, and processes that suppress it tend to look efficient on the way to failing.
What to baseline before you change anything
Most of these measures only work as comparisons, which means the single most valuable thing you can do is capture a baseline before any intervention. It costs almost nothing and it is routinely skipped, because the pressure is to improve rather than to measure.
Spend two weeks logging every question a new starter asks, in any channel, including verbally. Record the question in the words used, where it arrived, who answered, and whether a written answer already existed. That last column is the one that decides most subsequent decisions, and it must mean you can produce the document and point at the sentence, not that it is probably in the handbook somewhere.
At the same time, reconstruct the last three onboardings with timestamps. You want elapsed time per stage, not task completion, and you are looking for the two longest waits. Almost every team finds the top two account for the majority of elapsed time, which usefully narrows what needs attention.
Then ask the last five starters the direct question about unreported problems. This gives you the starting value for the metric that matters most and is the one nobody has.
Together this is perhaps a day of work. Without it, every subsequent change is evaluated against an impression of how things used to be, and impressions are generous about the past in both directions depending on who is remembering.
Reading the numbers at small volume
A practical problem that undermines most onboarding measurement: if you hire twenty people a year, almost nothing you measure is statistically meaningful on its own.
Two bad onboardings in a quarter is a ten per cent failure rate and might be two unrelated accidents. Ninety day retention moving from eighty five to seventy per cent might be one person moving house. At these volumes, treating any single period as a signal produces decisions driven by noise, and the usual outcome is a process that is changed repeatedly in response to individual events.
Three adjustments help. Use counts rather than rates, because three blockers is a fact and fifteen per cent is a construction with a denominator too small to bear weight. Pool across a longer window, usually a year, and compare like periods rather than consecutive ones. And weight qualitative input more heavily than you would at scale, because five detailed conversations genuinely do tell you more than twenty data points when twenty is all you have.
The thing not to do is adopt metrics designed for large organisations and report them at small volume with the same confidence. A dashboard showing onboarding satisfaction to one decimal place from a sample of six is presenting noise with the appearance of precision, and it will be read as the latter.
Who should own each number
Metrics without owners get reported and not acted on, and onboarding is particularly prone to this because it spans functions.
Elapsed time to operational belongs to whoever owns IT provisioning, because they control it and nobody else can. Splitting it out explicitly is also fair to them, since the alternative is being held to account for a composite number mostly determined by other people.
Question composition belongs to whoever owns documentation and internal communication, usually people operations. It is the metric that responds to content work.
Unreported blockers belongs to whoever runs the onboarding process end to end, because the fix is almost always a process change rather than a content or provisioning one.
Ninety day retention belongs to the manager population, with support. It is the one genuinely shared number and it is also the one most often used as a stick, which is why it needs the reasons attached rather than being reported bare.
The role that is usually missing is somebody accountable for the whole, with enough standing to ask another function why a handoff takes four days. Without that, each function optimises its own section, the handoffs stay untouched, and everybody can demonstrate their part is fine while the outcome stays poor.
Asking the right questions at the right time
Most of what you learn about onboarding comes from conversations rather than dashboards, and the design of those conversations determines whether they produce anything.
Timing matters more than wording. A conversation in week one produces politeness, because the person has no basis for comparison and every reason to be agreeable. A conversation at month three produces accuracy about the first week, because the person now knows what normal looks like and has enough standing to say so. The widespread practice of surveying in week two gets the worst of both: too early for perspective, late enough that the details have blurred.
Run two conversations rather than one. A short, specific check at day five aimed only at finding blockers, and a longer one at month three aimed at the experience. They have different purposes and conflating them produces a survey that does neither well.
The question design matters as much as the timing. General questions produce general answers, and general answers about onboarding are almost always positive. Replace them with questions that assume a problem exists and ask for specifics. Not whether the documentation was helpful, but what they had to ask somebody that should have been written down. Not whether they felt supported, but who they went to when they were stuck and whether that was the person they were meant to go to.
The second framing works because it is easier to answer. Naming a thing that happened requires recall. Rating an experience requires judgement about something they have no comparison for, and under that uncertainty people default to generous.
Keep the list short, four or five questions, and ask them in conversation rather than in a form where the sample is small enough to allow it. A fifteen minute conversation with three people produces more usable information than a well designed survey of twelve, at this scale, because the follow up question is where the actual content lives.
And write down what you hear in the person's words rather than coding it into categories immediately. The categories you would choose in advance are the ones you already know about, and the value of these conversations is almost entirely in the things you did not anticipate.
Metrics for the parts nobody measures
Three areas of onboarding are consistently unmeasured, and each hides a real failure mode.
The manager's preparation. Everything the manager contributes is invisible to measurement because none of it has a completion state. Nobody records whether the first week was planned, whether introductions were arranged, whether a first project existed on day one. The consequence is that the highest value contributions are the ones most easily squeezed during a busy fortnight, with no trace.
The measure that works here is simple and slightly uncomfortable: ask the starter, at day five, whether their manager had a plan for their first week. The answers are more varied than anybody expects and they correlate with outcomes better than most of the process metrics.
The gap before day one. The period between offer acceptance and starting is unmeasured almost everywhere, partly because the person is not yet an employee and partly because no system owns it. It is also the period when a candidate is most likely to reconsider, and where the quality of communication varies most, since it is usually handled by individual emails rather than by any process.
A count works here: how many times was this person contacted between accepting and starting, and by whom. Zero is common and is worth knowing.
Questions asked to colleagues rather than to a system. This is the largest category in most companies and it is invisible by construction. Every measurement tool sees only what passes through it, and the majority of new starter questions never do, because asking the person at the next desk or in a direct message is cheaper than any formal channel.
There is no clean way to measure this. The usable proxy is asking recent starters who they actually went to when stuck, and counting how often the answer is a peer rather than the designated route. A high peer ratio is not a failure in itself, but it tells you that your formal channels are seeing a biased sample, which matters when you are drawing conclusions from them.
The tools, and where each one struggles
Matram is our own product, stated here rather than buried, and judged on the same criteria as everything else. It is an answering agent: it reads your documentation and answers questions in the channel people already use. Its relevance to measurement is narrow and worth stating precisely. It can show you which questions are being asked, which is genuinely useful input to the composition metric above, and it cannot tell you anything about elapsed time, provisioning or retention. Pricing is flat at $29, $69 and $199 a month with unlimited seats and a 30 day trial needing no card.
Its limits: no free tier, not an HRIS so it holds no records and runs no approval chains, and its actions reach HubSpot, Zendesk and webhooks only, so it can open a ticket but will not provision anything or change workflow state. WhatsApp is coming soon rather than available. For measurement specifically, it also cannot see questions asked outside it, which is the same blind spot every tool in the category has and the reason the direct question at week one remains necessary.
BambooHR publishes Core at $10.25 per employee per month, Pro at $17.50 and Elite at $25 above twenty five employees. It reports task completion and timeliness well, which covers the process half of measurement, and its Pro tier bundles an assistant answering from your own policies. Where it struggles for measurement is that completion is the metric it naturally produces, and completion is the first of the three misleading ones.
Rippling starts from $8 per employee per month plus a monthly base fee. Because it spans HR, devices and application access, it is one of the few places the elapsed time to operational number can be produced without manual reconstruction. Where it struggles is scope, since adopting it for measurement would be an expensive way to get a number.
Workday publishes no public rate, its pricing URL returning a 404, so any figure attributed to it is an estimate. Strong reporting at enterprise scale and implemented over quarters.
Enboarder focuses on the experience and the manager prompts, and collects structured feedback at points in the journey, which is closer to useful than a single satisfaction score. It publishes no price. Where it struggles is that feedback collected through a system still carries the early-tenure politeness problem.
Leena AI and Moveworks publish nothing. The Moveworks pricing page carries a Forrester study rather than a rate card, and the figures printed there are claimed savings rather than costs, which is worth knowing before anybody repeats one into a board paper.
| Tool | Published price | Measurement value | Where it struggles |
|---|---|---|---|
| Matram | $29, $69 or $199 flat | Question composition | Blind to elapsed time, no free tier |
| BambooHR | $17.50 per employee per month | Task completion and timeliness | Produces the misleading metric |
| Rippling | From $8 plus a base fee | Elapsed time to operational | Expensive way to get a number |
| Workday | None published | Enterprise reporting | Quarters to implement |
| Enboarder | None published | Structured journey feedback | Early tenure politeness |
| Leena AI | None published | Question volume | Procurement cycle |
| Moveworks | None published | Question volume | Only an ROI study is public |
The decision table
| Situation | Scale | Setup | Primary Pain | Recommended Starting Point |
|---|---|---|---|---|
| Numbers look good, outcome poor | Any | Metrics exist | Wrong measures | Replace completion with composition |
| No measurement at all | Under 20 hires | Informal | No baseline | Two week log plus the week one question |
| Small volume, noisy numbers | Under 20 hires | Metrics exist | Noise | Counts not rates, annual pooling |
| Nobody owns the numbers | Any | Cross functional | No action | Assign each metric a function |
| Starters never complain | Any | Any | Silence | Direct question at day five |
| Volume fell, nothing improved | Any | Tooling added | Ambiguity | Check composition, not count |
Reading retention without over-reading it
Ninety day retention is on the list of four and it is the one most easily misused, so it deserves its own treatment.
The case for it is that it connects onboarding to something the business already cares about. The case against treating it carelessly is that it is influenced by everything: hiring quality, the role itself, the manager, compensation, the market, and events in somebody's life that have nothing to do with work. Onboarding is one input among many and usually not the largest.
So read it as a trend with reasons attached, never as a score. A quarter where two people left is a quarter where two people left, and whether that says anything about onboarding depends entirely on why. One person relocating and one person receiving a substantially better offer tells you about the market. Two people independently saying they never understood what was expected tells you about onboarding, and the second is actionable in a way the number alone is not.
This means the exit conversation matters more than the metric. Ask specifically about the first month, late in the conversation when the person is relaxed, and ask what would have made a difference rather than what went wrong. The second framing produces concrete answers and the first produces diplomacy.
Watch for one specific pattern, which is departures clustered at month four to six rather than month one to two. Very early departures usually indicate a hiring or role mismatch. Departures at four to six months more often indicate somebody who never got established, kept going on determination for a quarter, and then decided. That second pattern is the one onboarding owns, and it is invisible in a ninety day measure, which is an argument for also looking at six months even though the noise is worse.
What not to do is set a retention target and manage to it, because the fastest way to improve ninety day retention is to be slower to part ways with people who are not working out, which helps nobody.
When the metrics disagree
Eventually you will have two numbers pointing in opposite directions, and how you handle that determines whether the measurement is worth anything.
The common version is that elapsed time to operational has improved substantially while question composition has not moved. That combination has a specific meaning: provisioning got better and findability did not. The starter can now log in on day one and still cannot work out how anything is done, which is progress on the half that was easier to fix.
Another version is that unreported blockers fell while question volume rose. This usually means people have become more willing to ask, which is a good outcome wearing two different disguises, and it should not be read as problems increasing.
The one that causes most argument is good process metrics alongside poor retention. The temptation is to conclude that onboarding is fine and the problem is elsewhere, and sometimes that is correct. Before accepting it, check the exit reasons against the four metrics specifically, because process measures mostly capture the first fortnight and the reasons people leave at month five usually formed in weeks three to eight, which nothing on the list covers well.
The general rule is to resist reconciling disagreeing metrics into a single verdict. They are measuring different things and the disagreement is information. A dashboard that forces them into one score is removing precisely the signal that was useful, which is why the earlier recommendation is to report them separately and accept that the picture is sometimes mixed.
The cost of getting this wrong
The direct cost of a bad metric is that effort goes to the wrong place, and that is the smaller half.
The larger cost is that a flattering metric ends the conversation. A ninety four per cent completion rate is accurate, auditable and defensible, and it makes raising a concern about onboarding an argument against evidence. The person who suspects something is wrong has an impression, and the person who thinks it is fine has a number. In most organisations the number wins, and the problem persists for another year with institutional backing.
The second cost is that metrics shape behaviour, so a measure of attendance produces more attendance. If completion is what is reported, more content gets added, because adding content is the visible way to improve the number. The programme grows, the starter experience worsens, and the metric improves throughout.
The third cost is the quiet one. Measuring satisfaction early and acting on high scores means the organisation believes new starters are happy. The people whose first weeks went badly answered politely, said nothing, and some of them left at month five. By then the departure is attributed to fit or to the market, because the onboarding data said it went well.
So the question worth asking is not whether your onboarding metrics are good. It is whether any of them could get worse if you made onboarding better, and if none of them could, you are measuring effort rather than outcome.
Presenting this to people who want one number
Whoever you report to will ask for a single figure, and the request is reasonable even though the answer should be no.
The reason it should be no is that onboarding has genuinely independent components with different owners, and the composite conceals exactly the information that makes the numbers useful. An index that moves from 72 to 76 is uninterpretable: nobody can tell whether provisioning improved, whether people became more willing to report problems, or whether one of the inputs got worse while another improved more.
What works instead is a short standing format rather than a single number. Four lines, each with the figure, the direction and one sentence of cause. Elapsed hours to operational, down from eleven to four, because two tools moved into the role template. Unreported blockers, up from two to six, because we started asking. Seven of the top ten questions have stopped arriving. Ninety day retention unchanged, with both departures citing relocation.
That fits on a slide, takes thirty seconds to read, and is actionable in a way an index is not. It also protects you in the specific case that matters, which is the quarter when a number moves in the wrong direction for a good reason, and the sentence of cause is the difference between a conversation and an inquisition.
Set the expectation early that one of these numbers is designed to rise. Unreported blockers going up is the system working, and explaining that after the fact sounds like an excuse while explaining it in advance sounds like competence. This is a small piece of political work that determines whether honest measurement survives its first bad quarter.
And be willing to report that a quarter was uninformative. At low hiring volumes some periods genuinely contain no signal, and saying so is more useful than constructing a narrative from three data points. Organisations that tolerate that answer get better measurement than organisations that require a story every quarter, because the requirement for a story is what produces metrics that always have one.
Starting with four numbers and nothing else
If the current measurement is not working, the replacement is smaller than people expect and can be running inside a fortnight.
Pick the four. Elapsed working hours to operational, reporting the slowest tool. Unreported blockers found by direct questioning at day five. Question composition against a ten question baseline. Ninety day retention with reasons, pooled across a year. That is the whole set and there is no fifth.
Capture the baseline before changing anything else, which is the two week log and the conversations with recent starters. This is the step that gets skipped and the step that makes everything afterwards interpretable.
Report them separately, with separate owners named, and resist combining them into a score. Composite onboarding indices are popular because they produce one number for a slide, and they conceal exactly the variation that makes the individual numbers useful. A single figure that went from 72 to 76 tells nobody anything actionable.
Then leave them alone for two quarters. Onboarding metrics at normal hiring volumes need time to say anything, and the most common failure after adopting better measures is reacting to the first period as though it were a trend. The four above will tell you something real by the third quarter, and very little before then, and knowing that in advance is what stops the measurement becoming another source of noise.
Frequently asked questions
What are the most important onboarding metrics?
Four cover almost everything useful: elapsed working hours until a starter can operate every tool their role needs, the number of blockers found by asking directly rather than waiting for reports, how the composition of questions changes against a baseline, and ninety day retention read as a trend with reasons attached. Everything else tends to measure effort rather than outcome.
Why is completion rate a bad measure?
Because it records that items were clicked rather than that anything was learned, and because it moves in the wrong direction when you change the programme. A programme that doubles in length and holds its completion rate looks twice as effective while probably being worse, since front loading more content into week one reduces retention rather than increasing it.
Should we survey new starters about onboarding?
Ask specific questions late rather than general questions early. A satisfaction score in week two asks somebody to criticise an organisation they have just joined, before they know whether criticism is welcome, and the scores come back uniformly high everywhere. A direct question at day five about what they could not do produces a list instead of a number.
Why did our reported problems increase after we improved things?
Almost certainly because you started asking. New starters under-report by default, since raising a problem means declaring you cannot do your job during the period you are least sure how that reads. When you ask directly, previously invisible problems become visible, and the rise is your measurement improving rather than your service declining.
Is falling question volume a good sign?
It is ambiguous and should never be read alone. Questions fall when people can find answers, and they also fall when people conclude that asking is slow or unwelcome and start guessing instead. Look at composition rather than count: if the repetitive questions went and the judgement ones remain, that is self service, and if both went, people stopped asking.
How do we measure onboarding at low hiring volume?
Use counts rather than rates, pool across a year rather than comparing consecutive quarters, and weight conversations more heavily than you would at scale. At twenty hires a year almost no single period is meaningful, and reporting percentages to one decimal place from a sample of six presents noise with the appearance of precision.
Who should own onboarding metrics?
Split them by who can act. Elapsed time to operational belongs to IT provisioning, question composition to whoever owns documentation, unreported blockers to whoever owns the process end to end, and retention to managers with support. Someone also needs to be accountable for the whole, or each function will improve its own section while the handoffs between them stay untouched.
