It worked in the office. The site has no signal, and the operator is wearing gloves

Field deployment is where AI projects meet reality: patchy connectivity, hardware that has to survive dust and temperature, devices used one-handed, and people who will abandon anything that takes longer than the old way. The engineering is different, and so is the definition of done.

  • Something that works where the work is. On the floor, in the cab or on site, in the conditions that actually exist rather than the ones assumed.
  • Offline tolerance. Degrading sensibly without a connection and reconciling cleanly when one returns, because intermittent is the normal state and not the exception.
  • Interfaces for the environment. Usable with gloves, in glare, one-handed, in noise. Adoption is decided here more often than by the model.
  • A supportable estate. Devices enrolled, updatable, monitorable and replaceable without a visit for every fault.
Talk to us about this

What is delivered

  • Site assessment

    Connectivity, power, mounting, lighting, noise, temperature and who actually holds the device. Done on site, because none of this survives being assumed from a meeting room.

  • Edge and cloud split

    What runs on the device, what runs centrally and what happens when the link is gone. The decision that determines whether the system is usable at all.

  • Offline-first data flow

    Local queueing, conflict handling and reconciliation, so a shift can complete without a connection and nothing is silently lost when it returns.

  • Field interface

    Large targets, high contrast, minimal typing, voice or scan where hands are busy, and an obvious state so nobody wonders whether it saved.

  • Device management

    Enrolment, updates, remote wipe and health monitoring. An estate of unmanaged devices becomes unsupportable within a year.

  • Pilot and rollout plan

    One crew or one line first, with the failures collected, before anything is rolled out to everybody.

How it runs

  1. 01

    Go and look

    A shift on site with the people who will use it. Every field project that skipped this rediscovers why it should not have.

  2. 02

    Prototype in the environment

    Tested in the actual conditions, not on a desk. Glare, gloves and dropped connections are requirements, not edge cases.

  3. 03

    Pilot with one crew

    Small, watched closely, with a route back. The complaints from this group are the specification for the rollout.

  4. 04

    Roll out and support

    Staged, with device management and a support path that does not require a site visit for every fault.

A good fit when

  • The work happens away from a desk: production, logistics, service, construction, utilities, agriculture.
  • An office-built system has been rejected by the people who were supposed to use it.
  • Connectivity is unreliable and the current answer is paper.
  • Inference has to happen locally for latency, cost or data reasons.

Not the right service when

  • Everything is in an office with reliable connectivity. Then this is ordinary application engineering.
  • You need certified safety-critical or explosion-proof equipment. That is a specialist supplier and we will say so rather than take the work.
  • The process itself is undefined. In the field this is worse than in an office, because there is no time to work it out during a shift.

Frequently asked questions

Which hardware do you use?
Whatever fits the environment and your existing estate — rugged tablets, industrial handhelds, phones in cases, or a fixed panel. We do not resell hardware, so the recommendation follows the conditions rather than a supplier relationship.
Can models run on the device?
Often yes, and increasingly it is the right answer for latency, cost and data residency. Which model runs locally is a trade-off between accuracy, battery and device cost, and it is decided with measurements from the site rather than from a datasheet.
What about safety-critical systems?
We do not build them. Anything inside a safety function falls under machinery and functional-safety obligations with assessment and change control that a business system should not claim. Our systems read and advise; they do not actuate.
How do you handle patchy connectivity?
By assuming it. Local storage, queued operations, conflict rules decided in advance, and an interface that always shows whether something is synced. Systems that treat offline as an error state fail their first bad week.
How long does a first deployment take?
A pilot with one crew is typically six to ten weeks including the site work, which is usually the part that is underestimated. Rollout depends on the size of the estate and on how much device management already exists.

Keep reading

Other services