Guide · for MSPs and IT teams
What is a device health score?
One number that says how a machine is really doing and, when it's built properly, why. Here's how health scores work, what feeds them, and how to tell a trustworthy score from a black box.
The definition
A device health score condenses everything a machine is telling you — the signals listed below — into a single number or band that a human can act on. Instead of reading twelve graphs per laptop, you read one score per fleet and drill into the machines that need attention.
MSPs use health scores to decide three things: which devices need attention now, which belong in next quarter's refresh budget, and which are fine, and to prove all three to a client who is about to be asked for money.
What signals actually matter
The signals that predict real cost live below the operating system's surface, and most of them never reach an RMM dashboard:
- Storage wear and uncorrected errors — the strongest hardware signal there is. A drive logging uncorrected errors is warning you; act before it takes data with it.
- Battery capacity and cycle count — a laptop at 60% design capacity is a productivity problem long before it's a failure.
- Thermal throttling and power-on hours — the difference between "slow laptop" and "cooked laptop".
- Per-application crash counts and patch age — because very often the device is fine and one application is the problem.
- Network quality and workload pressure — a dead Wi-Fi zone and three VMs on an 8 GB machine both masquerade as "failing hardware".
What makes a score trustworthy
It shows its working. A score you can't explain to a client is a liability in a QBR. Every point traces to a named rule, shown inside the product next to the score: storage uncorrected errors in the last 7 days → −22 points is an argument; "the AI thinks it's bad" is not.
It's deterministic. Same inputs, same score, every time. If the number can drift without the device changing, nobody will budget against it.
It admits what it can't see. A device with missing signals should say Not assessed — never be shown as a zero or quietly averaged in. A wrong score in front of a client is worse than no score.
It doesn't pretend to predict failures. Google's study of over 100,000 drives found that a large share of failures showed no SMART warning at all, while healthy drives showed the same "warning" signals, and Backblaze's published drive stats keep confirming it. Treat any score with a "failure probability: 92%" badge with suspicion; a condition band with evidence is what the data actually supports.
It diagnoses before it recommends. A useful score distinguishes a degrading SSD (hardware) from a crashing app (software), a Wi-Fi dead zone (network), and an overloaded spec (usage) — because three of those four don't need a new machine.
A worked example
Here is how a rule-based score actually lands on a number. One warehouse laptop, five signals, each tied to a named rule:
| What was found | Reading | Points |
|---|---|---|
| Starts at full health | — | 100 |
| Storage wearing out, with errors | 92% used, errors this week | −40 |
| Battery getting tired | 61% capacity, 612 cycles | −11 |
| Noticeably slower to start up | 40% slower over 90 days | −9 |
| A few smaller things | patching, minor items | −6 |
| Final score | Replace — the hardware is the problem | 34 |
An illustration, not the real scoring. The point is the shape: every drop has a plain reason and a reading you can check — never a mystery number.
Device health score vs device age
Most refresh lists are still ranked by age: everything over four years gets replaced, everything under it stays. Age is cheap to measure and weakly correlated with the thing you care about. A four-year-old desktop that has spent its life in an air-conditioned office browsing spreadsheets is usually fine; a two-year-old laptop with a cooked battery, worn SSD and three VMs on an 8 GB spec is a ticket factory.
A measured health score reorders that list. In practice it does two commercially interesting things at once: it rescues devices an age policy would have replaced (the biggest hidden saving), and it catches the young machines an age policy would have missed — the ones that generate the angriest tickets. Age still matters as one input; it just stops being the whole policy.
Common questions
Is a device health score the same as an experience (DEX) score?
Can an RMM produce a device health score?
What should a score do when data is missing?
How many points should a single signal be worth?
How Lilo does it
Lilo runs a lightweight agent on Windows, macOS and Linux, feeds a fixed, published set of signals into a deterministic rules engine, and produces a banded 0–100 score with every point traceable to a named rule, plus the four-category diagnosis, the refresh budget, and a white-label client report. Free for 5 devices, with published pricing after that.
See a real score, with the work shown.
Lilo is in early access — see the score, the reasoning, and the report on your own fleet.