There are three ways to find out a charger is broken.
A driver finds out for you — and tells you via a one-star review, a support call, or increasingly, a regulator. Your monitoring finds out — after the failure, which is better, but the site is already down. Or your platform tells you a failure is coming — and an engineer fixes it during a scheduled visit before any driver ever meets it.
Most networks live in the first two. The economics of the third are what this piece is about.
What a charger failure actually costs
The repair invoice is the smallest line. Add up the real bill for a hard failure at a busy rapid hub:
- Lost sessions. A well-used rapid connector can serve thousands of kWh a month. Every offline day is revenue that doesn't return — drivers who found an alternative often keep using it.
- Reactive callout premiums. An emergency engineer visit costs a multiple of a planned one, and the first visit frequently just diagnoses — parts arrive with visit two.
- SLA and regulatory exposure. UK rapid networks carry a 99% reliability requirement; commercial site hosts increasingly write uptime penalties into contracts. Margin for unplanned downtime is thin.
- Trust erosion. Reliability is the number one driver complaint about public charging. A network's reputation is set by its worst charger, not its average one.
Against that, the cost of attending a charger a fortnight before it fails — during a planned run, with the right part on the van — looks like a rounding error.
Why chargers are actually good candidates for prediction
"Predictive maintenance" has a buzzword history, so it's fair to ask whether it genuinely works here. Chargers turn out to be unusually good subjects, for three reasons:
They fail gradually more often than suddenly. Connector wear, cable damage, cooling degradation and payment-terminal decline all leave statistical fingerprints — rising error rates, more early-terminated sessions, delivered power sagging below rated maximum — weeks before hard failure. (Our engineering team catalogued these patterns in Why EV chargers go offline.)
They report constantly. Every OCPP charger streams status, meter values, error codes and session outcomes all day. The evidence base for prediction already exists; most operators simply never mine it.
They exist in comparable populations. A network has many chargers of the same model in similar duty. When one unit's behaviour drifts from both its own history and its peers', that divergence is signal — the kind a threshold-based alert rule can never express in advance.
What this looks like in practice
In the AmpNexus platform, Cortex continuously scores every device and site for failure risk over 7–30 day horizons, and turns raw event noise into ranked, actionable incidents:
CP-2481 · connector wear detected — failure risk 87% within 14 days. Maintenance visit recommended before Friday.
The operational shift that follows is bigger than the technology:
- Maintenance rounds get planned by risk, not by rota. Engineers visit the ten chargers most likely to fail next, with the likely cause and parts list in hand — not whichever sites are due a look.
- Alert volume goes down, not up. Thirty noisy device events collapse into one prioritised incident with context. Ops teams act on a shortlist instead of triaging a firehose.
- Firmware decisions get smarter. Risk scores inform where updates roll out first — and rollout outcomes feed back into the models.
The honest caveats
Anyone selling you a crystal ball is overselling. Prediction is probabilistic: some flagged chargers would have soldiered on, and genuinely sudden failures — vandalism, vehicle impact, board-level electronics dying instantly — will always exist. The realistic goal is to move a meaningful share of failures from unplanned to planned, and to keep improving as the models learn your fleet.
The prerequisite is data discipline: connector-level telemetry, complete session outcomes, and enough history to establish baselines. If you're evaluating platforms, ask precisely how failure risk is derived and what data feeds it — vague answers to that question usually mean a red/amber/green dashboard wearing an AI badge. (More vendor-testing questions in our tender questions guide.)
Where to start
You don't need a big-bang programme. The adoption path we see work:
- Get the telemetry flowing — connector-level status, sessions and errors into one place, across every hardware brand.
- Baseline for a few weeks — let the models learn what normal looks like for your fleet.
- Run predictions alongside existing operations — treat recommendations as advisory, and score them against what actually happens.
- Shift scheduling once trust is earned — when flagged chargers keep turning out to be genuinely failing, let risk drive the maintenance calendar.
Networks that make this shift stop discovering failures from driver complaints — and start treating reliability as something they engineer, evidence and publish, rather than hope for.
Cortex AI is part of the AmpNexus platform — predictive maintenance, demand forecasting and prioritised incidents across mixed-vendor fleets. See how it works or book a demo.