TL;DR
Device health monitoring is one of the few places where acting early is strictly cheaper than acting late, because the failure modes it catches degrade gradually rather than suddenly. The difficulty is not collecting the data. It is choosing thresholds that produce action rather than a feed nobody reads.
- Four signals carry almost all the value: battery, disk, memory, crashes.
- A threshold without an action is a feed, and feeds get ignored.
- Battery is the best single predictor of somebody about to complain.
- Offline machines are not healthy machines. Absence is a signal too.
- Aggregate by default. Per-person detail needs a stated reason.
- Tickets per machine beats any telemetry, and your help desk already has it.
The four signals device health monitoring should watch
Everything else is detail. These four account for the overwhelming majority of machines that are about to generate a complaint.
| Signal | What it predicts | Action when crossed |
|---|---|---|
| Battery health | A replacement request framed as slowness | Offer a battery, not a machine |
| Free disk space | Slowness, failed updates, crashes | Run a cleanup, do not notify |
| Sustained memory pressure | The machine is wrong for the work | Review the specification for that role |
| Crash frequency | Imminent ticket, often repeated | Investigate the individual machine |
Why battery is the one to start with
It degrades on a predictable curve, it is cheap to fix relative to a replacement, and the user experience of a failing battery is almost always reported as the laptop being slow or unreliable rather than as a battery problem.
It also fails in the other direction. A perfectly healthy battery rules out one explanation cheaply, which narrows the investigation before anybody spends an hour on it.
That misattribution is the whole opportunity. A machine that gets replaced because somebody said it was slow, when a battery would have fixed it, is an expensive misdiagnosis that happens constantly and silently.
A threshold without an action is just a feed
The commonest failure in this area is not missing data. It is generating alerts nobody has agreed to do anything about, which produces a stream people learn to ignore within a quarter.
Before setting any threshold, write the action next to it. If you cannot name one, do not set the threshold.
Disk above a stated percentage. Action: a cleanup routine runs automatically. Nobody is notified unless it fails. This removes the category rather than reporting on it.
Battery below a stated health figure. Action: a replacement is offered to that person by name, with a date. Not a notification that their battery is degrading.
Crashes above a count per week. Action: a human looks at that specific machine. This is the one that genuinely needs judgement, so it should be the only one that reaches a person routinely.
Memory pressure sustained across weeks. Action: review the specification for that role, not that machine. One person struggling is an individual case. Six people in the same role struggling is a procurement decision.
The signal everybody forgets
A machine that has stopped reporting. An absent device is not a healthy device, and in practice it is one of three things: somebody on extended leave, a machine that died, or a device that quietly left the company.
All three matter and the third is the expensive one. Control frameworks such as ISO/IEC 27001, the information security management standard expect the inventory to reflect reality, and a device that stopped checking in six weeks ago is the clearest available evidence that it does not.
What it cannot tell you, and the line to hold
Disclosure: RemoAsset, named below, is owned by the same people who publish PeopleOpsHQ. It is described with its limitations stated, on the same terms as every other supplier. This site takes no vendor money.
Two limits worth stating plainly before anybody builds a programme on this.
It does not predict most hardware failure. Boards, ports, hinges and screens fail without warning. No telemetry foreshadows a drop or a spilled drink. The answer to those is a spare machine or accidental damage cover, and monitoring is not a substitute for either.
It does not see anything off the machine. A person waiting four days for an access approval has a perfectly healthy laptop throughout, and the dashboard will say so.
The line on what gets collected
Health telemetry and activity telemetry are different things that frequently arrive in the same agent, and the second is where this stops being uncontroversial.
Battery health, disk space and crash counts describe a machine. Application usage by person, active hours and websites visited describe a person, and that is a different proposition with its own expectations. the UK Information Commissioner guidance on monitoring workers sets out that monitoring should be necessary for a stated purpose rather than collected because the agent can collect it, and under frameworks such as the consolidated text of the General Data Protection Regulation the resulting data carries obligations regardless of why it was gathered.
The workable position: collect machine health, aggregate it by default, and switch off per-person activity unless somebody can write down why it is on. the UK National Cyber Security Centre device security guidance treats the managed state of a device as the control, which is the right framing and does not require watching what the user types.
When the answer is a replacement rather than a repair, the supply side sits elsewhere. Configured device procurement and delivery handles it, and the honest limitation is the mirror of the tool’s: a logistics supplier has no idea which of your machines are failing.
The decision table
| Situation | Scale | Setup | Primary pain | Recommended starting point |
|---|---|---|---|---|
| Replacing on age | 200 plus | Fixed cycle | Wrong machines replaced | Battery health first, then crashes |
| Alerts already ignored | Any | Tooling in place | No action attached | Delete every alert with no action |
| Under 100 machines | Small | One platform | Extra tooling overhead | Use what the MDM already reports |
| Devices stop checking in | Any | Remote fleet | Register drifts | Treat absence as its own alert |
| Asked for per-person detail | Any | Pressure from above | Different proposition | Aggregate, and ask for the purpose |
The best signal is not in the telemetry
Tickets raised per machine. Your help desk already has it, no agent produces it, and it correlates better with somebody actually struggling than anything an endpoint reports.
The reason is simple: a ticket is generated by a human deciding the situation was bad enough to ask for help. Telemetry measures the machine. A ticket measures the experience.
Join the two and the picture gets much sharper.
- Bad telemetry, no tickets. Somebody is coping silently. Worth a conversation, because these are often the people who later leave citing frustration nobody knew about.
- Good telemetry, several tickets. The machine is fine and the work does not fit it. A specification question, not a fault.
- Bad telemetry and tickets. Obvious, and the easiest to action.
Why the first row matters most
Because it is invisible to both systems in isolation. The dashboard shows a degraded machine with no complaints and concludes it is fine. The ticket queue shows nothing and concludes the same.
Act on that row by asking rather than by replacing. Somebody coping with a degraded machine has usually decided the request is not worth making, and the useful information is why they concluded that. Often it is a refusal they or a colleague received eighteen months ago.
People cope with bad equipment for a long time, particularly new joiners who do not yet know what is normal, and particularly anybody who has had a request refused before. Designing around what people actually experience rather than what the instrumentation emits is the substance of ISO 9241-210, the standard on human-centred design for interactive systems, and this is the clearest practical example of the gap.
Worth checking in the same pass whether the self-service route people would use to report a problem is usable by everyone. the W3C Web Accessibility Initiative introduction to accessibility covers that ground. For what endpoint tooling covers more broadly, see our piece on DEX tools.
See the supply side of a replacement decision at configured device procurement and delivery, or read what DEX tools miss.
Frequently asked questions
Which signals are worth monitoring?
Four carry almost all the value: battery health, free disk space, sustained memory pressure and crash frequency. Everything else is detail. Battery is the one to start with, because it degrades predictably, it is cheap to fix relative to a replacement, and users almost always report it as the laptop being slow rather than as a battery problem.
Why does the battery misdiagnosis matter?
Because a machine replaced on the grounds that somebody said it was slow, when a battery would have fixed it, is an expensive misdiagnosis. It happens constantly and silently, and it is the single clearest case where early action is strictly cheaper than late action.
What makes an alert useful?
An action written next to it before the threshold is set. If you cannot name the action, do not set the threshold. Disk above a percentage should trigger a cleanup that runs rather than a notification. Battery below a health figure should trigger a replacement offered by name with a date.
Which alert should actually reach a person?
Crash frequency above a count per week, because that is the one genuinely needing judgement about a specific machine. Everything else should either run automatically or feed a periodic review. Sustained memory pressure across several people in one role is a procurement decision rather than an individual alert.
What about machines that stop reporting?
Absence is a signal. A device that has not checked in is one of three things: somebody on extended leave, a machine that died, or a device that quietly left the company. The third is the expensive one, and a device that stopped reporting six weeks ago is clear evidence the asset register no longer reflects reality.
What can it not tell you?
Most hardware failure, because boards, ports, hinges and screens fail without warning and no telemetry foreshadows a drop or a spill. The answer there is a spare machine or accidental damage cover. It also sees nothing off the machine, so somebody waiting four days for an approval has a perfectly healthy laptop throughout.
What is the best signal of all?
Tickets raised per machine, which your help desk already has and no agent produces. A ticket is generated by a human deciding things were bad enough to ask for help, so it measures the experience rather than the hardware. Joined to telemetry it reveals the people coping silently with a degraded machine, which neither system sees alone.
PeopleOpsHQ covers the operational side of distributed work. We describe suppliers with their limitations stated, and we take no vendor money.