The AI Promise in Public Transit Maintenance: Why the Journey Matters More Than the Destination

There’s a version of the AI story that goes like this: you buy the software, plug in your data, and suddenly your maintenance operation runs itself. Buses don’t break down unexpectedly. Parts are always on hand. Your team is always in the right place at the right time.

That version is a fantasy. And the sooner the transit maintenance community talks honestly about that, the sooner we can get to what AI actually looks like in practice — which, when done right, is genuinely transformative.

McKinsey estimates that AI-driven predictive maintenance could reduce maintenance costs by 10–40% and cut unplanned downtime by up to 50%. A single unplanned bus breakdown, when you factor in emergency labor, towing, substitute transportation, and lost service hours, averages around $8,500. The math for getting this right is compelling. So is the cost of getting it wrong.

AI Is a Journey, Not a Switch You Flip

Every maintenance manager who has sat through a vendor demo has seen the gleaming dashboards — the predictive alerts, the recommended actions populated by an algorithm trained on millions of service records. It looks effortless. It almost never is.

The 2026 Verizon Connect Fleet Technology Trends Report — now in its sixth year, drawing on insights from nearly 900 fleet professionals — offers a useful window into where fleets actually are on this journey. GPS fleet tracking has reached 80% adoption. Video telematics is at 46%, up 10 percentage points since 2023. And yet fleets using GPS tracking report average maintenance cost reductions of just 16%. Solid — but nowhere near the ceiling that AI-enabled predictive maintenance promises. The gap between where most fleets are and where the technology can take them remains enormous.

The agencies that have started to close that gap share a common trait: they didn’t treat AI as a technology decision. They treated it as an organizational transformation. Singapore’s SBS Transit didn’t start with a sophisticated algorithm — it started with a commitment to disciplined data capture across a 1,000-bus pilot fleet, partnering with Stratio to monitor over 200 vehicle parameters in real time. According to Stratio, the result was a 20% drop in bus breakdowns, with the program now expanding to cover more than 3,000 vehicles — 62% of Singapore’s entire public bus fleet. Transport for London took a similar approach on the London Underground’s Central line, using predictive analytics to get ahead of motor failures. The expected savings: approximately £3 million per year, according to TfL data scientist Akis Tsiotsios — a figure significant enough that the agency had previously been spending heavily on scheduled maintenance just trying to prevent the same failures reactively.

These outcomes aren’t magic. They’re the compounded result of years of foundational work.

The Dirty Secret: Your Data Is the Real Variable

In the old world of analytics, there was a simple warning: garbage in, garbage out. Feed a system bad data, get bad outputs.

With modern AI, the stakes are dramatically higher. These systems don’t just return a bad number — they learn from bad data, encode it into their models, and amplify it at scale. It’s not garbage in, garbage out anymore. It’s garbage in, garbage cubed out.

Think about what that means in a maintenance context. Research on CMMS systems shows that as much as 4% of manually entered maintenance data contains errors — and over 39% of agencies are missing more than 10% of their data due to gaps in vehicle assignment and service history records. In a fleet of 300 buses with thousands of work orders annually, that’s not a rounding error. That’s a systematic corruption of the training data your AI model depends on.

The failure modes are maddeningly specific:

Inconsistent fault codes. Two technicians write up the same HVAC failure four different ways. The AI sees four different events with no pattern to learn from.
Blank root cause fields. Work orders get closed without a documented cause — common under time pressure, catastrophic for model training.
Manual odometer entries. A technician miskeys 47,823 miles as 478,23. The AI now schedules a major service inspection 430,000 miles too early — or too late.
Missing vehicle assignments. A repair is logged against the wrong asset ID. The vehicle with the actual fault history looks pristine. The one that never needed the work looks like a maintenance nightmare.

An AI system trained on data like this doesn’t just produce unreliable predictions. It produces confidently wrong predictions — delivered with the false authority of an algorithm.

The path forward isn’t to lower your expectations of AI. It’s to raise your standards for data. That means auditing what you’re actually capturing today, standardizing fault codes across your team, building validation rules into your CMMS so incomplete records can’t be closed, and integrating your telematics feed directly rather than relying on manual entry. Data discipline isn’t a prerequisite you knock out once and move on from. It’s an ongoing practice. The agencies winning with AI treat it that way.

What the Journey Actually Looks Like

A realistic, successful AI adoption journey in transit maintenance tends to follow a recognizable arc.

Stage 1: Getting the foundation right
This is about data hygiene, system integration, and establishing baselines. You’re not doing AI yet. You’re doing the work that makes AI possible — auditing asset records, standardizing how work orders are documented, connecting your CMMS to your telematics rather than bridging the gap with spreadsheets and manual entry. This stage is unglamorous and chronically undervalued, which is exactly why so many agencies skip it and then wonder why their AI pilot failed to deliver.

Stage 2: Starting with targeted use cases
The agencies that build momentum pick one failure mode where they have reasonable historical data and a clear operational impact. Copenhagen Metro is instructive here: already maintaining service availability above 99% — one of the highest rates of any metro system in the world — they’re not resting on that record. As Executive Director Søren Boysen put it in 2024: “We aim to keep that level by continuing to explore new ways to conduct maintenance.” Their answer is a partnership with Hitachi Rail to shift from time-based to condition-based maintenance, using sensor monitoring to detect patterns in train components before they cause failures. TfL took the same focused approach on the Central line: one failure mode, one model, measurable results before expanding. That’s not a lack of ambition. That’s what sustainable AI adoption actually looks like.

Stage 3: Scaling with organizational trust
AI recommendations are only valuable if your team acts on them. That requires trust — and trust requires transparency. Your technicians need to understand why the system is flagging an asset, not just that it is. The best tools surface the signals behind their predictions: “This bus has shown elevated coolant temperature variance over the past 340 service cycles, consistent with early-stage water pump failure.” When your team can see the reasoning, they can interrogate it, validate it against their own experience, and eventually rely on it. That human-machine feedback loop is where the real gains compound over time.

The Leaders Who Will Define the Next Decade

The maintenance directors who will look back on this period as a turning point — not a missed opportunity — are making deliberate choices right now. They’re not chasing every new AI announcement or letting vendors shortcut them past the foundational work. They’re building the data culture, the technical infrastructure, and the organizational habits that make AI actually work.

That’s harder than buying software. It takes longer. It requires convincing leadership to invest in data governance and system integration that won’t show up as a line item in a vendor ROI calculator.

But consider what’s on the other side of that investment. Deloitte’s research on predictive maintenance in rail found that Trenitalia — one of Europe’s largest rail operators — reduced downtime by 5–8% and cut annual maintenance spend by 8–10% through sensor-based predictive systems across 1,500 locomotives, translating to roughly $100 million in yearly savings. For a transit agency operating on tighter margins than a national rail operator, even a fraction of that impact is material. Every repair logged cleanly, every inspection completed on time, every fault code entered with discipline — it all becomes training data. Your maintenance operation gets smarter every day, because your data gets richer every day.

The agencies that get this right won’t just have better AI. They’ll have a compounding operational advantage that their peers won’t be able to buy their way into overnight.

The journey is the point. The journey is where you build what can’t be bought.
Start there.

Value solutions

Value solutions