The Use of AI in Deployments
How AI can be used to improve the overall delivery of connected devices.
A large AMI rollout generates an enormous amount of operational data — install photos, crew routes, commissioning reads, field notes — most of which used to be reviewed, if at all, well after the fact. AI's most practical role in deployment isn't some autonomous future capability; it's using that data while the programme is still running, to catch problems earlier and route resources better.
None of this replaces the judgement of a programme team or the hands of a field technician. What it does is compress the time between something going wrong — a bad firmware batch, a mis-installed meter, an inefficient route — and someone finding out, from weeks or months down to days or hours.
This briefing sets out where AI is genuinely useful across the deployment lifecycle, and the technical risks worth understanding before leaning on it too heavily.
Where AI fits in a rollout
The same five stages exist in every deployment programme — AI's contribution looks different at each one.
Figure 1. AI's contribution shifts from optimisation early in the lifecycle to detection and feedback as the estate goes live and matures.
What AI actually does well here
What changes for the programme
The clearest benefit shows up in how early a problem is caught. A firmware defect that would once have been noticed after several thousand installs — usually through a spike in support calls — can surface after a few dozen, once anomaly detection is watching commissioning data as it arrives.
The estate's data quality foundation, covered elsewhere in this series, does most of the work here: an AI model trained on inconsistent or incomplete commissioning data will happily learn the wrong pattern with just as much confidence as the right one. The programmes getting the most from AI in deployment are the ones that treated clean data as a prerequisite, not a nice-to-have.
Six risk points worth naming
Training data reflects past deployments
A model tuned on last year's estate can miss new device types, regions or failure modes.
False confidence in automated QA
A computer-vision check that misses a subtle defect can be trusted more than the manual process it replaced.
Explainability gap for field decisions
A technician overridden by an AI-driven schedule needs to understand why, not just comply.
Data quality dependency
AI models inherit every gap in the estate data they're built on.
Model drift over a multi-year rollout
A model trained early in a programme can degrade in accuracy as the estate and conditions change.
Over-automation at the wrong point
Safety, access issues and customer interactions still need a person empowered to deviate from the model.
Adopting AI responsibly
Start with narrow, well-defined use cases
Route optimisation and photo QA before anything predictive or autonomous.
Keep a human in the loop for exceptions
AI should flag and prioritise, not silently decide.
Monitor model performance continuously
Not just at launch — accuracy drifts as the programme and estate evolve.
Feed field outcomes back into the model
A deployment programme generates exactly the labelled data AI needs to improve.
Treat data quality as a prerequisite, not a parallel workstream
A model is only as good as the estate data it learns from.
AI doesn't change what a deployment programme has to get right — good planning, quality installation, clean data. What it changes is how quickly a programme finds out when one of those things is going wrong, which, across a rollout spanning years and hundreds of thousands of devices, is often the difference between a contained issue and an expensive one.
How a read moves across the estate
HOVER OR TAP A STAGE TO EXPAND TECHNICAL SPECS
Where AI earns its keep in a rollout
The clearest measure of AI's value in a deployment isn't the sophistication of the model — it's the time between something going wrong and someone finding out. A bad firmware batch that once surfaced through a support spike after several thousand installs can be flagged after a few dozen once anomaly detection watches commissioning data as it arrives.
That compression only holds if the estate's data foundation is clean: a model trained on inconsistent commissioning data learns the wrong pattern with as much confidence as the right one. The programmes getting the most from AI treated data quality as a prerequisite, not a parallel workstream.
- 01
- 02
- 03Smart Meter Implementation Programme — DESNZ (UK)
- 04AI in the Energy System — Energy UK
Chris Gales
Chris Gales leads CEHG. He converts complex technical and commercial choices into clear decisions and dependable outcomes for utilities running connected device estates at scale.
He writes these field notes from the practice — on bearer migration, smart metering architecture, procurement and operations — to help estates start counting down early.