Predictive Maintenance for ATM Networks
I once watched an ATM in Skärholmen swallow two cards on a grey Tuesday morning. The woman in front of me — late for work, coffee in hand — just stood there. Staring at the machine like it had personally betrayed her. Which, in a way, it had.
That machine had been failing for weeks. The card reader was sticky. The screen flickered. But nobody knew. Not until it ate her card and mine.
This is how banks lose customers. Not with a bang. With a swallowed card at 8:47 AM.
The old way is broken
Traditional maintenance waits for disaster. Machine dies. Phone rings. Technician drives two hours. Problem gets diagnosed. Parts get ordered. Machine comes back to life, maybe, three days later.
Meanwhile, customers keep walking up to the corpse. They do not know it is dead until they try to use it. Every one of them walks away angrier than the last.
For a network of 5,000 machines, 2% downtime means 100 dead ATMs at any moment. That is not a statistic. That is a hundred Tuesdays in Skärholmen.
Three streams of truth
Predictive maintenance drinks from three wells. I have seen all three, and I will tell you which ones matter most.
Machine telemetry. Cash levels. Component temperatures. Motor currents. Card reader cycles. Error logs. The machine whispers before it screams. A motor drawing 15% more current than last month is a motor planning its retirement. A card reader with declining success rates is a card reader begging for attention.
Transaction patterns. A machine doing 500 transactions daily wears faster than one doing 50. Obvious, right? But here is what the spreadsheet does not show: holiday spikes that murder components, summer lulls that mask creeping failures, Monday mornings when every machine in the network gets hammered simultaneously.
Field observation. This is the one the engineers forget. Is the machine room air-conditioned or baking in a Stockholm July? Is dust creeping through the enclosure like a slow invasion? Are cables hanging loose, waiting for a janitor's mop to finish the job? Telemetry sees the machine. Field observation sees the world trying to kill it.
The prediction engine
Combine the three streams and you get a risk-ranked list. High probability: send a technician. Medium probability: watch closely. Low probability: sleep easy.
The best systems hit 70–80% accuracy. Three out of four failures prevented before they happen. The fourth one still hurts, but it hurts less than four out of four.
I have seen this work. I have also seen it fail because the bank bought the software and forgot the process. A prediction without action is just expensive anxiety.
Three things that actually matter
Flexible scheduling. Your maintenance crews need to pivot. Today. Not next quarter. If the model says machine #4,721 is going to die on Thursday, someone needs to be there Wednesday. Fixed rotations are comforting and useless.
Parts in stock. Here is a true story: a bank predicted a card reader failure 14 days out. Replacement lead time: 21 days. You do the math. The prediction was perfect. The preparation was a joke.
Field verification. Fix it, then check it. Did the intervention work? Is the machine breathing normally again? Close the loop. I have seen too many "fixed" machines fail again within a week because nobody verified.
The money
5,000 machines. $500 per day of downtime. 30–50% reduction. That is $750,000 to $1.25 million annually. Payback in 12–18 months.
But the real money is the customer who does not walk away angry. The one who gets her cash, catches her bus, and forgets the ATM even exists. That is the goal. Invisible infrastructure. Boring reliability.
The banks that get this will survive. The ones still waiting for the phone to ring will not.
Request a pilot →