Shop-floor telemetry feed rendered in the Fleet dashboard.
The pitch goes like this: connect all your equipment to the cloud, feed the data into an AI model, and the model will tell you when a machine is about to fail so you can fix it before it breaks.
A lot of that pitch is true. Some of it is aspirational. A specific part of it is bluntly a fiction that gets us calls from customers who tried the glossy version first and want something that actually ships.
Here’s what our experience says is real.
What actually moves the needle
1. Continuous CMH tracking
Cumulative machine hours, accrued continuously from real telemetry — not estimated, not entered by the operator on a clipboard. This one thing changes the maintenance conversation more than anything else. Suddenly you can schedule the 500-hour service at 490 hours instead of “when someone notices.”
This is not AI. It’s addition. It works.
2. Threshold alerts
A skid-steer that was running at 85% utilization for six months and then drops to 40% is telling you something. Same with a generator whose runtime hours suddenly spike. You don’t need a neural net to catch that — you need a simple rule that says “if this week’s number differs from the rolling average by more than N standard deviations, page ops.”
Most of what gets called “anomaly detection” in the IoT marketing material is this, dressed up. It’s fine, it’s useful, and it runs in a hundred lines of code on a single Postgres row.
3. Geofence + status pairings
Equipment that leaves its expected site at 2am tells you something. Equipment that logs offline for 48 hours tells you something else. Pairing a simple location rule with a simple connectivity rule gives you a surprisingly complete “something’s wrong” signal without any ML.
4. Human being looking at the dashboard
We’ve yet to find a customer whose preventive maintenance program improved more from AI than from the ops lead finally having a dashboard they actually open. Software that makes the data legible beats software that makes predictions the human ignores.
What doesn’t ship
- Unsupervised “failure prediction” models trained on equipment telemetry typically don’t generalize across machines, let alone across customers. Every fleet has its own quirks. A model trained on generators doesn’t predict skid-steer failures, and a model that worked at site A needs to be retrained for site B.
- Energy consumption as a leading indicator of failure sounds great but in our data, energy consumption is a lagging indicator 80% of the time. It changes after the machine is already struggling.
- Marketing-grade “root cause analysis” — a dashboard that tells you why a machine broke is not AI, it’s a PDF the service tech writes after the fact.
What we ship instead
Fleet has no ML model. It has:
- Continuous CMH accrual from live telemetry
- Rule-based alerts with trend comparisons
- A live dashboard the ops lead can walk through in three minutes
- A report scheduler that emails the right numbers to the right people on the right cadence
That’s the whole product. Customers tell us it moves the needle. We’ll take that over a model every time.
The honest version of the pitch
Connect your equipment to a system that measures what’s happening, shows it to the right person at the right time, and pages someone when the numbers look wrong.
No “AI.” No “revolutionizing.” Just the part that actually changes outcomes.