TL;DR
Workplace analytics is useful exactly to the degree that it stays about the organisation rather than about the people in it. The same dataset answers both kinds of question, which is why the discipline has to come from the design rather than from the tool.
- Ask organisational questions. The data answers individual ones too, so be deliberate.
- Aggregate thresholds, enforced in the tool, not promised in a policy.
- Meeting load and queue time are the two findings that repay the effort.
- Never join it to performance data. That is the decision you cannot reverse.
- Publish the findings internally. Secret analytics becomes surveillance by default.
- Delete on a schedule or the dataset becomes the liability.
The workplace analytics questions worth asking
Four that repay the effort, and they share a shape: each is about the organisation, each has a decision attached, and none needs a name.
Where does elapsed time actually go? A request takes nine days. Which step held it? This is the highest-value question in the category and the one most often skipped in favour of something flashier.
Best tools for AI HR Tools
What is the real meeting load, by team? Not an average across the company, which hides everything. Per team, including the teams nobody worries about.
Which teams never interact that probably should? Useful for structure decisions, and the one place a network view earns its keep.
Did the change we made actually do anything? Requires a before number, which is the main reason to start measuring before you need to.
| Question | Decision it supports | Needs names? |
|---|---|---|
| Where elapsed time goes | Fix the slowest step | No |
| Meeting load by team | Change a norm, protect focus time | No |
| Teams that never interact | Structure or rituals | No |
| Before and after a change | Keep it or revert it | No |
Keeping it aggregate by design, not by promise
A policy saying individual data will not be viewed is worth very little, because the data is there and the pressure to look arrives eventually, usually during something difficult.
Build the constraint into the system instead.
A minimum group size, enforced by the tool. No view returns a result for fewer than a set number of people. Ten is a common floor. This single setting prevents most of the drift.
No drill-through to individuals. Not greyed out. Absent.
Access logged and reviewed. Who ran what, visible to someone other than the person running it.
Retention measured in months. Analytics answers questions about now. A three-year archive serves no analytical purpose and is purely a liability, and under frameworks such as the consolidated text of the General Data Protection Regulation holding personal data longer than the purpose requires is exactly what the storage limitation principle is about.
The one join to refuse
Do not connect this dataset to performance ratings, pay or promotion decisions. Ever, and the reason is practical as much as ethical.
Once that link exists, every number becomes a thing people manage rather than a thing that describes reality, and the analytics stop working as analytics. You also cannot un-ring it: the belief that the link exists does the damage whether or not anyone uses it.
the UK Information Commissioner guidance on monitoring workers makes the related point about purpose, and the practical version is that a dataset gathered to find slow processes should not quietly acquire a second use.
The two findings that justify the whole exercise
Disclosure: RemoAsset, named below, is owned by the same people who publish HROpsLab. It is described with its limitations stated, on the same terms as every other supplier. This site takes no vendor money.
Most programmes produce a lot of charts and two useful results. Knowing which two in advance lets you run a much smaller exercise.
Meeting load concentrated somewhere nobody expected. Reliably, one team or one level is carrying far more than anyone realised, usually middle managers or a function that sits between two others. That is actionable the week you find it.
One step holding most of the elapsed time. Equally reliably, a single approval, handoff or queue accounts for the bulk of a process that everybody describes as generally slow. Specific, fixable, and invisible without measurement.
What usually turns out not to matter
Communication volume, after-hours activity in aggregate, and most network visualisations. They are interesting, they look impressive, and in practice they rarely change a decision.
That is the test worth applying to every chart produced: did anything change because of this? A programme that cannot name a changed decision after two quarters is a reporting exercise, and the honest action is to stop it rather than expand it.
Expansion is the usual response, because a programme that has not produced a result is easy to read as one that needs more data. Occasionally that is true. More often the question was wrong, and more data answers the wrong question more thoroughly, which is the most expensive way to stay stuck.
The finding that is not in the data at all
Worth saying because it comes up every time: a team that looks slow sometimes has old hardware rather than a process problem. Analytics will show you the symptom perfectly and will never name that cause, because machine age is not in the dataset.
Check it separately before concluding anything about a team. Where a refresh is the answer, configured device procurement and delivery handles the supply, and the limitation is the same as always: it cannot tell you which machines are due. That comes from your own register.
The decision table
| Situation | Scale | Setup | Primary pain | Recommended starting point |
|---|---|---|---|---|
| Process slow, cause unknown | Any | Multi-step | No visibility | Elapsed time by step, aggregate |
| Meeting overload suspected | 100 plus | Calendar data available | Averages hide it | Load by team, not company-wide |
| Asked to identify individuals | Any | Pressure from above | Changes the dataset’s nature | Refuse, offer the aggregate answer |
| Works council or EU staff | EU | Consultation needed | Purpose and consent | Agree scope before any pilot |
| Two quarters, no decisions changed | Any | Running already | Reporting, not analytics | Stop it rather than expand it |
Publish the findings internally
The single practice that separates analytics from surveillance, and it costs nothing.
Share what you measured, what you found and what you changed, with everyone, routinely. Not a summary for leadership while the rest of the company wonders what is being collected.
Three reasons it works.
It constrains the questions. A question you would be embarrassed to publish the answer to is a question not worth asking, and knowing the finding will be shared filters the programme before it starts.
It converts suspicion into interest. People are far more comfortable with measurement they can see the output of, and markedly less comfortable with measurement they only know exists.
It generates better questions. Teams who see the meeting load chart ask about their own, and those requests are the ones most likely to lead to a change, because the team asking already wants to act.
What to publish, concretely
A short note each quarter: the question asked, the aggregate finding, the decision taken or the reason none was. Two pages. No individual data, no team league tables, which recreate the individual problem one level up.
Include the questions you decided not to answer
The most useful paragraph in the note, and the one nobody writes.
Somebody will have asked for a breakdown by individual, or a ranking, or a view of after-hours activity by name. Recording that it was asked and declined, with the reason, does two things at once. It tells the company where the boundary actually sits rather than where the policy claims it sits, and it protects whoever declined from being asked again every quarter by somebody new.
It also makes the boundary a shared decision rather than one analyst’s judgement, which is the only version that survives a change of leadership.
Keep the list even when nothing was asked that quarter. An empty entry is itself a finding, and it is the one that tells you the boundary is understood rather than merely written down.
Designing the measurement around how people actually experience the result rather than around what the tool can emit is the substance of ISO 9241-210, the standard on human-centred design for interactive systems, and publishing is the cheapest way to keep that honest. For the adjacent question of what activity data cannot show, see our piece on employee productivity software.
Rule out the hardware before blaming the process at configured device procurement and delivery, or read what productivity software cannot see.
Frequently asked questions
What questions is workplace analytics good for?
Four. Where elapsed time actually goes in a slow process, which is the highest-value question and the most often skipped. Real meeting load by team rather than a company average. Which teams never interact that probably should. And whether a change you made did anything, which needs a before number.
How do you keep it from becoming surveillance?
Build the constraint into the tool rather than the policy. A minimum group size enforced by the system, commonly ten, so no view returns a result for fewer people. No drill-through to individuals, absent rather than greyed out. Access logged and reviewed by someone other than the person running queries.
What should never be joined to this data?
Performance ratings, pay and promotion decisions. The reason is practical as well as ethical: once the link exists, every number becomes something people manage rather than something that describes reality, and the analytics stop working. The belief that the link exists does the damage whether or not anyone uses it.
How long should the data be kept?
Months, not years. Analytics answers questions about now, so a long archive serves no analytical purpose and is purely a liability. Holding personal data longer than the purpose requires is precisely what the storage limitation principle addresses, and unconfigured deletion is the commonest failure in this category.
Which findings actually repay the effort?
Two, reliably. Meeting load concentrated somewhere nobody expected, usually middle managers or a function sitting between two others. And one step holding most of the elapsed time in a process everyone describes as generally slow. Communication volume and network visualisations look impressive and rarely change a decision.
Why publish the findings internally?
Because it constrains the questions, since one you would be embarrassed to publish is one not worth asking. It converts suspicion into interest, as people are far more comfortable with measurement whose output they can see. And it generates better questions, because a team that asks about its own data already wants to act.
Could a slow-looking team just have old laptops?
Often, and analytics will never name that cause because machine age is not in the dataset. It will show the symptom perfectly. Check hardware age separately before concluding anything about a team, since the cheapest explanation is frequently the correct one and it sits entirely outside the tool.