---
title: "AI in Legionella control: predictive maintenance"
source_url: https://legionella.io/articles/ai-in-legionella-control-predictive-maintenance/
canonical_url: https://legionella.io/articles/ai-in-legionella-control-predictive-maintenance/
pillar: "Technology & Remote Monitoring"
summary: "What AI predictive maintenance really does in UK Legionella control: catch drift and failing kit early, and what it can't tell you about safe water."
primary_keyword: "AI Legionella"
date_published: 2025-08-10
date_reviewed: 2026-06-26
author: "Legionella.io editorial team (REMOTE TECH LTD)"
reviewed_against: "HSE L8 and HSG274 guidance"
region: "United Kingdom"
license: "(c) REMOTE TECH LTD. Quote freely with attribution and a link to source_url."
---

# AI in Legionella control: predictive maintenance

No sensor on the market can count Legionella in your pipework in real time, and any product that implies otherwise is selling you the wrong promise. What "AI" can genuinely do in water safety is quieter and more useful: read weeks of monitoring data, spot a control measure slowly drifting out of shape, and flag it while there is still time to fix the cause rather than clean up the consequence. That is predictive maintenance, and it lives or dies on the quality of the data underneath it.

So the honest question is not "should we buy AI?" It is "where does pattern-spotting across our monitoring data catch something a person reviewing monthly logs would miss?" The answer turns out to be specific, and it is worth seeing it play out.

## What "predictive" actually means here

Periodic monitoring tells you the state of an outlet on the day someone checked it. Predictive maintenance works on the trend between those days. A hot return that reads in range every month can still be losing a fraction of a degree each week as a calorifier scales up or a circulating pump tires. No single reading fails. The slope does. Analytics that watch the slope — call it AI if the vendor insists — can raise a flag before any individual check would.

That is the whole mechanism. It is forecasting equipment behaviour from continuous data, not detecting bacteria. Keep that distinction sharp and the rest of the subject stays honest.

## A composite scenario: the drift nobody saw

*The following is an illustrative scenario, not a real named site or incident.*

Picture a further-education campus with a central calorifier serving a long hot-water circuit out to a teaching block added in a later refurbishment. Remote temperature sensors had been fitted at sentinel points eighteen months earlier, mostly to cut down on clipboard rounds. The monthly figures looked fine. Every manual check passed.

The analytics layer told a different story. Across one autumn term the return temperature at the far sentinel had been sliding — a little each week, never enough to fail a spot check, but a clean downward line once you plotted twelve weeks together. The system flagged the trend and ranked that circuit as the site's top developing risk.

The responsible person investigated rather than dismissing it. The cause was unglamorous: a circulating pump on the extended leg was losing head, so hot water reached the far block slower and arrived cooler. Left until the trend crossed the threshold, that circuit would have spent weeks sitting in the warm, still conditions Legionella favours — and the first hard evidence might have been a failed sample, or worse. Caught early, it was a pump replacement and a written note in the log.

Nothing about that story is dramatic. That is the point. The value was in seeing a slow signal that a snapshot regime is built to miss.

## The decisions that actually mattered

The technology did one thing: it surfaced a trend. Everything that made it useful was human, and it is worth naming what those decisions were.

The sensors were at meaningful points. A flag on the far sentinel of a known long circuit means something; a flag on a sensor stuck somewhere convenient does not. Someone competent owned the alert — the trend went to a named responsible person with the authority and the knowledge to act, not into a shared inbox nobody reads. And the alert was tied to a route: investigate, diagnose, fix, record. A prediction with no owner and no next step is just a more expensive way to be surprised later.

This is the line between a tool that earns its place and one that adds noise. If you are weighing up a system, the same logic that ran on remote temperature monitoring applies here; [Case study: remote monitoring improved Legionella compliance](https://legionella.io/articles/case-study-remote-monitoring-improved-legionella-compliance/) walks through how continuous data changed day-to-day compliance on a different site.

## What it transfers to your own site

A few lessons carry across regardless of the building or the brand of kit.

Predictive maintenance is only as good as where the sensors sit and how clean the data is. Garbage in, confident garbage out — and a confident wrong prediction is worse than none, because people act on it. Treat any model's output as a hypothesis to investigate, not a verdict. The model is reasoning about temperature, flow and usage patterns; it is not reasoning about the bacteria, which is exactly why the [question of real-time Legionella detection](https://legionella.io/articles/real-time-legionella-detection-fact-or-fiction/) deserves a sceptical read before you trust any "safe water" indicator.

And the duties do not move. L8 still expects the duty holder to keep records of precautions, monitoring and the management arrangements [1], and HSG274 still frames monitoring around what the site's risk assessment requires rather than what a platform defaults to [2]. Analytics can make that evidence richer and faster. It cannot become the competent person.

## Where this leaves the model, honestly

Predictive maintenance points at equipment, not at safety. A model can suggest a calorifier is likely to underperform next month; it cannot tell you the water is safe, and you should distrust any wording that blurs the two. Thresholds, monitoring frequency and what counts as an acceptable result are set by your site-specific risk assessment and a competent assessor, not by a vendor's preset. And because no continuous sensor identifies the organism, sampling still does the verification job on the schedule your risk assessment and HSE guidance support [3]. Use the analytics to act sooner; keep the judgement where the law puts it.

A cheap way to test all of this before you spend: pull the last twelve weeks of readings for one long or low-use circuit and plot them as a line rather than a column of pass/fail ticks. If you can see drift that your monthly checks waved through, that gap is precisely what predictive maintenance is for — and it is the test any vendor demo should have to pass on your own data, not theirs.

## FAQ

### Does AI actually detect Legionella, or just the conditions?
The conditions. Current systems infer risk from what they can measure continuously — temperature, flow, usage, equipment behaviour — and predict where a control measure is drifting. They do not count bacteria, so sampling and competent assessment still do the verification.

### Is "predictive maintenance" different from the alerts our remote monitoring already sends?
Yes, in what triggers them. A standard alert fires when a reading crosses a threshold today. Predictive maintenance acts on the trend before the threshold is reached, so it is meant to buy you lead time rather than hand you a same-day breach.

### If the analytics flag a developing fault, who is responsible for acting on it?
The duty holder and responsible person, exactly as before. A flag is an input to their decision; it does not transfer accountability to the software or the supplier [4]. The flag only helps if a named, competent person owns the follow-up.

## Related reading

- [Case study: remote monitoring improved Legionella compliance](https://legionella.io/articles/case-study-remote-monitoring-improved-legionella-compliance/)
- [Real-time Legionella detection: fact or fiction?](https://legionella.io/articles/real-time-legionella-detection-fact-or-fiction/)
- [Effective communication about water safety with occupants](https://legionella.io/articles/effective-communication-about-water-safety-with-occupants/)

## Sources

[1] HSE, "Legionnaires' disease. The control of legionella bacteria in water systems - Approved Code of Practice and guidance (L8)". https://www.hse.gov.uk/pubns/books/l8.htm
[2] HSE, "Legionnaires' disease: Technical guidance (HSG274)". https://www.hse.gov.uk/pubns/books/hsg274.htm
[3] HSE, "Testing and monitoring your water system for legionella". https://www.hse.gov.uk/legionnaires/testing-monitoring-water-system.htm
[4] HSE, "Legionnaires' disease - what you must do". https://www.hse.gov.uk/legionnaires/what-you-must-do/index.htm