Ranked Before the Diggers Arrive: How AI Decides Which Street Gets Dug Up

There's a damp patch on a pavement near you. Maybe it's been there since August. You've reported it twice. And somewhere in a water company's systems, your street already has a score — one that decides whether the diggers arrive next Tuesday or next spring.

This isn't a conspiracy. It's a spreadsheet with ambition. Across England and Wales, water companies have spent the past few years wiring up pipes with sensors, hiring data teams and buying in models that promise to find leaks before they surface. The trouble is that the ranking those models produce is almost never visible to the people living above the pipes.

What the software is actually looking at

Strip away the vendor slides and most leak-prediction systems are doing a few fairly humble things at once.

First, they watch district metered areas — chunks of the network, often a few hundred homes each, where the company knows exactly how much water goes in. Compare that to what's billed out and you get a rough leakage figure. Traditional practice measured this overnight, when nobody's showering. Machine learning mainly speeds up the pattern-spotting: which DMAs drift out of character on a Tuesday, which ones spike after a cold snap, which ones creep upwards for months before anyone notices.

Second, there are the sensors themselves. Acoustic loggers listen for the hiss of escaping water. Pressure transmitters pick up the small thumps that suggest a joint is failing. Some companies use satellite radar to catch ground movement, and a growing number mine smart meter data for the signature of a leaking loo or a burst supply pipe on private property.

Third, there's you. Customer calls, web reports and social media complaints get fed in as data points, which is why two reports of a damp patch can sometimes move a job up the queue faster than a sensor that's been grumbling for a fortnight.

The ranking isn't really about fairness

Here's the bit that rarely makes it into the press release. These models are usually not asked "which leak is worst?" They're asked something closer to "where does a repair buy us the most?" — and "most" is defined by the company's regulatory incentives.

Ofwat sets outcome targets covering leakage, supply interruptions and pollution incidents. Miss them and the company can be penalised; beat them and it can earn a reward. A burst on a trunk main serving four thousand homes is a different commercial animal from a slow seep under a cul-de-sac in a village. Both annoy people. Only one shows up in the numbers the regulator scores.

So if your road is a quiet little DMA with 1970s pipework and a handful of patient residents, the algorithm may be perfectly correct in ranking you low. It has read the incentives. It just hasn't read the fact that the damp patch is outside a care home.

What the models tend to miss

Vulnerability, mostly. Unless a household is on the company's Priority Services Register — the list for people who need extra help with supply interruptions — the model has no idea it exists. Dialysis at home, a baby, a wheelchair, a shop that loses a day's takings when the water goes off: none of that is in the training data unless someone put it there.

Local knowledge goes missing too. The foreman who knows that bit of cast iron has been patched four times since 2011, the parish council that has a folder of photos going back years — neither feeds the score.

And "AI" is doing a lot of heavy lifting in some of these announcements. Plenty of systems in use are closer to rules and statistics than to anything you'd call a neural network. That's not automatically bad. It does mean the marketing and the maths don't always match.

So why did your street get skipped?

The honest answer is that nobody will tell you directly, because there's no statutory letter that says "your road ranked 47th". Water companies in England are private businesses. Freedom of Information doesn't reach them the way it reaches your council, which is worth knowing before you fire off a request into the void.

What you can do:

• Ask the company, in writing, for the reason for the prioritisation of a specific reported leak. Not the algorithm — the reason. Firms are usually willing to explain whether a job was held for permit, traffic management, parts or budget.
• Check Ofwat's published performance data for your company's leakage and supply-interruption record. If your area feels neglected, the annual figures either back that up or they don't.
• Go to the Consumer Council for Water if a complaint stalls. They handle disputes about service and can press for an explanation.
• Ask about the Priority Services Register — both to join it, and to ask whether it feeds into repair prioritisation. Increasingly, it does.
• Read the company's leakage strategy and board papers, which are usually online and often mention where modelling is being used. It's dry, but it's the closest thing to seeing the machine from outside.
• Raise it with your MP or council's scrutiny committee if a whole neighbourhood is affected. Councils have no direct power over water companies, but they can make a fuss in public, which has a way of loosening statements.

The question worth asking

None of this means the models are wrong to exist. A network with tens of thousands of kilometres of pipe and a finite number of crews has to be triaged somehow, and a data-led ranking is probably better than a clipboard and a hunch.

But triage is a policy choice dressed as a technical one. The model is told what matters. If nobody outside the company ever sees what it was told, the street with the damp patch is arguing with a black box.

So ask the boring question. Not "is the AI accurate?" but "what is it optimising for, and who decided that?" It's the same question you'd ask of any decision-maker with a digger.