Device management looks straightforward at pilot scale because a hundred devices forgive almost any design decision. The problems that decide whether a product is viable appear later: provisioning without manual steps, updating safely over bad connections, knowing a device is unhealthy before a customer does, and retiring hardware that will outlive the software team.
- 01A pilot fleet forgives everything. Design for the fleet size that makes the business work, not the one in the lab.
- 02Over the air update is the single capability that decides whether a product can be supported, and it has to fail safely.
- 03If you cannot tell a broken device from an idle one, support costs scale with the fleet.
- 04Provisioning that needs a human step does not survive volume.
- 05Devices outlive software teams. Whoever maintains this in year five is a design input, not an afterthought.
What does IoT device management cover?
Getting a device securely onto the network, knowing it is there and working, changing its software safely, and retiring it at end of life. Four things, all unremarkable at small scale and all decisive at large scale.
The reason it is underestimated is that a pilot works without any of it done properly. Devices can be configured by hand, watched informally and updated by someone visiting them.
None of that survives a fleet that earns money, and the gap between the two is where most connected product plans lose their margin.
Why does what worked at a hundred units break at ten thousand?
Because manual steps that cost minutes become the dominant cost, and because rare failures stop being rare. A fault affecting one device in five hundred is invisible in a pilot and is twenty support cases a week at scale.
Connectivity assumptions break too. In the lab devices are on good networks; in the field they are intermittent, behind unpredictable firewalls, or asleep when you need them.
And the distribution of device states widens. At scale you always have some devices on old firmware, some half provisioned, some that have been offline for months and will reappear expecting the world to be as they left it.
What makes over the air updates safe?
The ability to fail without bricking. An update interrupted by power loss or a dropped connection has to leave the device bootable, which usually means dual partitions and a verified rollback rather than writing over the running image.
Staged rollout, so a bad build reaches a handful of devices rather than the fleet. The value of that is proportional to how expensive a site visit is, and for devices in the field it is usually the single highest return design decision.
And signing, so a device will only accept software you actually produced. Retrofitting that later is considerably harder than including it.
How do you know a device is unhealthy?
By deciding in advance what healthy means and having devices report it, rather than inferring from silence. Silence is ambiguous: it means broken, asleep, out of coverage or decommissioned, and the four need different responses.
A heartbeat with a small amount of state is usually enough, and far cheaper in bandwidth than streaming telemetry nobody reads.
The commercial consequence of getting this wrong is that support cost scales with fleet size instead of flattening, because every question requires a human to investigate a specific device.
Should you buy a platform or build one?
Buy, for most teams, most of the time. Provisioning, update and monitoring are solved problems with mature offerings, and building them is months of work that does not differentiate your product.
Build or extend where your device does something the platform has no concept of, where regulation requires control a vendor will not give you, or where the per device economics of a platform break your unit cost at target volume. That last one is worth modelling at the fleet size that makes the business work rather than at current volume.
Either way, price the exit. Device fleets are long lived, platform relationships are sticky, and the cost of moving is much easier to estimate before signing than after.
What gets forgotten until it is expensive?
End of life. Devices outlive the teams that built them, and a fleet with no decommissioning path accumulates hardware that still connects, still consumes certificates, and still appears in support queues years after anyone intended to support it.
Certificate and credential rotation, which arrives suddenly as an expiry date and is painful to do for the first time under pressure.
And documentation of why things are the way they are. The engineer who knew why a timeout is set to that particular value is the most important dependency in most device fleets, and that dependency is rarely written down.
We take technical ideas through design, prototyping and validation to something that can be built, tested and manufactured.