The device estate is the line missing from the business case

A field AI business case usually has a careful section on the software and the model: development, licences or inference costs, integration. The devices appear as a single line — so many tablets at so much each — and everything that happens to those devices after they come out of the box is missing.

That missing part is where field projects quietly become expensive. An estate of unmanaged devices becomes unsupportable within a year, and by then the cost is no longer in the business case. It is in the operations budget, spread across IT tickets, site visits and devices nobody can account for.

What the device line actually contains

A device that fits the environment. Not the cheapest tablet, and not necessarily the most rugged. Dust, temperature, glare, gloves, one-handed use and mounting decide it, and they differ between a cab, a loading bay and a substation.

Mounting, power and charging. Where the device lives during a shift, how it charges, what happens when the battery dies at the far end of a site. Small in price, large in whether the system gets used.

Enrolment and configuration. Every device registered, configured the same way and tied to a person or a role, so that a lost one can be found or locked.

Updates. The application, the operating system and — if inference runs locally — the model itself, delivered to devices that spend much of their time offline, without disrupting a shift.

Remote wipe and health monitoring. Knowing which devices are failing, full or out of date before a crew reports it, and being able to clear one that has gone missing.

Support and replacement. A path for faults that does not require somebody to drive to site, and a pool of spares, so that a broken device does not mean a crew back on paper for a week.

The decision that moves the cost most

Before any of that, one design decision determines which device you need at all: what runs on the device, what runs centrally, and what happens when the link is gone.

Infographic

The split decides the device

Answer needed nowCan wait for sync

Central inferenceThin device, model in the cloud; watch latency and keep a fallback.

On-device inferenceCapable hardware, local model, measured on site. The costly corner.

Thin clientThe cheapest devices. Ordinary application engineering may be enough.

Offline queueCapture locally, reconcile on return. Modest hardware, careful sync rules.

Reliable linkIntermittent link

A system that only needs to capture data and reconcile it later can run on modest hardware. One that has to answer immediately in a place with no signal needs inference on the device: more capable hardware, more battery, and a model chosen from measurements taken on site rather than from a datasheet. That is a different estate with a different cost, and it should be decided before devices are bought, not after.

Building the business case honestly

Compare with the real current process. In many field operations the competitor is paper. Paper is cheap, works without a signal and never needs an update. The case has to rest on what changes when the data arrives at the end of the shift instead of days later — fewer errors from re-keying, faster decisions, nothing lost between site and office — not on the elegance of the system.

Cost the estate over its life, not at purchase. Devices, accessories, management, support time and replacements, over the years you expect to run them.

Put a name against device management. If the plan says IT will handle it, check that IT has agreed, has the tools and has the time. An estate that belongs to nobody is the one that becomes unsupportable.

Count the pilot as part of the cost. Six to ten weeks with one crew, including the site work, is how you find out which device and which split actually work. Skipping it moves that discovery into the rollout, where it is paid for at every site.

What a good outcome looks like

A year after rollout, every device is accounted for, runs a current version and can be swapped from a spares pool without a site visit. Faults are visible centrally before crews report them. And fewer and fewer people keep a paper notebook just in case, without anybody having told them to stop.

Our part

In field deployment AI engineering, device management — enrolment, updates, remote wipe, health monitoring — is a deliverable rather than an afterthought, and the edge and cloud split is decided with measurements from the site. We do not resell hardware, so the recommendation follows the conditions rather than a supplier relationship.