$7.4 Million in Six Months: How Heavy Industry Got AI Customer Experience Right

Originally posted at eGlobalis: https://www.eglobalis.com/7-4-million-in-six-months-how-heavy-industry-got-ai-customer-experience-right/

I have spent most of my working life around companies that sell services and experiences. Oftentimes in heavy industry you cannot return the machine. Trains, crushers, harbour cranes, ship engines, machine tools that run three shifts a day. These companies rarely use the words customer experience. They almost never appear at CX conferences. And over the past three years they have quietly done a better job of putting AI near their customers than most of the software industry has managed.

One number made me look harder. At a mining site in South America, six months of Metso’s data-driven service reported capturing 94% of identified risks, avoiding more than 215 hours of potential downtime, reducing the spare parts needed, and producing a total captured value of USD 7.4 million.

Two direct caveats before we go on. That figure is reported by Metso itself, and the customer is not named. So read it as a company statement rather than an audited result.

The size of the number is not the point. Where it was counted is. Nobody measured what Metso saved. They measured what the customer would have lost — hours of production, parts not bought, failures that never happened.

That is the whole difference. Most companies bought AI and then went looking for something it could improve, usually inside their own cost line. Heavy industry made a promise to the customer first, and bought AI to keep it.

1. The promise came first, the AI came second

In Bucharest, Alstom has maintained the Metrorex metro fleet since 2004. Fleet availability went from 56% to 100%, and in 2022 the contract was extended for another fifteen years. That result belongs to a full maintenance contract, not to an algorithm, and that is exactly the point. Long before any AI ran in that depot, the company had already agreed what it owed the customer, in a single number both sides could read.

Alstom runs more than eighty maintenance contracts lasting twenty years or longer. A twenty-year availability commitment rewires a company. You cannot run customer experience as a campaign. You cannot hand the relationship to a tool and hope. Someone signed for a result that will still be measured after today’s executives have retired.

In Gold Coast, Australia, Alstom took over tram maintenance in 2014. It delivered 98% availability across 70 million passenger journeys, a 25% increase in fleet availability, and the work converted into a fifteen-year contract. The availability was the experience. Nothing else needed saying.

Compare that with how AI arrived in most service organisations. Someone bought a tool, pointed it at the contact centre, then looked for a number to justify it — usually containment rate or cost per contact. Nobody asked the customer what outcome they were buying. Skip that step and the AI has nothing to aim at, so it optimises the only thing left: your own costs.

2. The alert goes to an expert, not to the customer

Wärtsilä, the Finnish marine and energy group, runs a service called Expert Insight on live vessel data. It Expert Insight combines AI with rule-based diagnostics, learns the behaviour patterns that come before a failure, and flags the symptom to an expert who then recommends the intervention.

Metso does the same in mining. Its analysis combines instrumentation, advanced analytics and AI models with failure-mode knowledge, handled by experienced people in the company’s performance centres. Hitachi Construction Machinery built ConSite Mine to ConSite Mine to monitor mining machines around the clock and use machine learning to predict expensive failures before they happen, such as cracks forming in an excavator boom or arm, with the alerts going to customers and to dealers.

TRUMPF, in Germany, has gone furthest. In one business model the machine sits at the customer’s site making parts, but a TRUMPF team in Neukirch, Saxony, operates it across three shifts, including at night, with cameras recording a short clip either side of every fault so the AI learns to avoid it. The manufacturer took operational responsibility instead of selling a box and a warranty.

Notice what is missing: no chatbot. Not one of these companies hands the customer raw AI output and leaves them to interpret it. The machine talks to the people who can act. The customer meets the result.

3. Measured in the customer’s economics

This is where the industrial numbers get uncomfortable for the rest of us. Metso now has 800 connected units, over double the number when it launched its data-driven services in September 2025. It reports detecting 80% to 90% of issues, availability gains of up to 6%, and CO₂ reductions of up to 20% per tonne.

The South American site I opened with is one instance of that. At launch the company also reported resolution times cut by up to 50%, a potential risk found roughly every 1,400 operating hours, and over eight hours of production saved per case solved. Every one of those units is time on the customer’s clock.

Alstom publishes the same kind of arithmetic. HealthHub delivers up to 20% in material cost savings, cuts train downtime by up to 30% and recurring faults by up to 50%. On the Kazakhstan Railways contract, failures fell by 40%. On Dubai’s tram system, performance rose by 130%.

The small details prove the point. Alstom monitors train toilets, because a failed toilet pulls a train from service and can trigger penalties of up to €10,000 per instance. It monitors brake pad wear because that maintenance can reach €600,000 a year. Nobody measured a satisfaction score. They measured the operator’s exposure.

This is what I meant when I wrote that certainty rather than delight. Heavy industry never had the option of delight. It only had certainty to sell.

4. Governance before scale

Komatsu is doing the sequence in the order almost nobody else managed. It a central AI function in April 2026 and is working toward a Global AI Center of Excellence from 2027, built on know-how gathered through regional workshops and hands-on implementation.

The governance function comes first, the global scale-up after. Most companies I meet did the reverse: they scaled, discovered how many tools were already running across the business, and are now writing rules retroactively while customers are already exposed to the output.

5. Where the model falls short

These companies have not earned a prize. Here are the limits.

Almost everything above is aimed at an asset, not a relationship. I found little published on the human side: the commercial conversation, the renewal negotiation, the complaint that has nothing to do with a machine. A crane that never fails is not the same as a customer who feels well treated when something goes wrong.

None of these companies publishes retention or renewal data tied to AI. The closest evidence is indirect; in the length of the contracts they keep winning. That is real, but it is not proof, and I will not pretend otherwise.

And some of what is announced right now is marketing rather than measurement. Doosan Bobcat presented Jobsite Companion at CES 2026, SANY showed a service assistant at CONEXPO 2026, and Emerson says its Ovation AI agents resolve complex issues in a fraction of the time. These are company statements about products, not results from customers. Treat them as intent.

One more gap. When alerts travel through a dealer network, as they do at Hitachi, the customer’s experience depends on a partner you do not employ. Hitachi is solving the technical half, extending remote diagnostics into countries with strict data-residency rules. The commercial half is harder and mostly unsolved, in their industry and in ours.

What travels to your company

You do not need cranes to use any of this. Eight things travel.

1. Write down the outcome before you buy the tool. If you cannot state what you owe the customer in a number they would recognise, your AI has nothing to serve and will default to serving your cost line. That is how deflection got mistaken for service.
2. Send the output to whoever can act, not to the customer. The industrial pattern is consistent: the model finds the pattern, a human decides, the customer receives a resolved situation. Most consumer AI reversed this and handed the customer a machine to argue with.
3. Measure in the customer’s currency. Hours of production, penalties avoided, availability. Not tickets closed. Metso can say USD 7.4 million because it counted what the customer would have lost, not what Metso saved.
4. Build the rules before the rollout. Komatsu’s order of operations costs more at the start and far less afterwards.
5. Hunt the failures that cost more than they look like they should. Alstom instruments toilets and brake pads, not because they are interesting, but because their financial consequence is out of proportion to their size. Every industry has these. Almost nobody has listed them.
6. Turn every alert into an assigned action. A detection that does not become a scheduled job, with a part number and a person’s name against it, is noise you are paying to generate. If nobody can be held to an alert, stop producing it.
7. Decide who owns the customer when a partner delivers. If your dealer performs the work, you made the promise and somebody else keeps it. Either share the economics with them or serve those accounts directly. Ambiguity here is the most common way a good service fails in front of a customer.
8. Say plainly what you have not proven. Metso reports USD 7.4 million and does not name the customer. Stating that openly costs a manufacturer nothing and buys it credibility for every other number it publishes. The reverse is also true, and it is expensive.
9. There is evidence the market rewards this. Epiroc reported in its first quarter of 2026 that customers valued dependability and high parts availability, with order intake up 11%, or 23% organically. Konecranes describes a customer in India, Bothra Shipping Services, that carried spare parts for failures that never materialised, and moved to predictive maintenance to plan ahead. These are the plainest customer experience wins I have seen this year, and none of them was designed by a CX team.
10. The lesson is not that heavy industry is smarter. It is that heavy industry never had the luxury of vagueness. When failure costs a plant its shift, you find out quickly whether your technology serves the customer or serves your own reporting. Everyone else has been able to avoid that test. AI is about to end that, in every industry, and the companies that already know what they owe their customers will be the ones ready for it.

👉 Stay ahead of CX, AI, and innovation — Subscribe to my weekly LinkedIn Newsletter “CX Insights by Ricardo S. Gulko.

Share it if it’s useful — and let’s connect for more. Ricardo Saltz Gulko

My columns in several respected CX publications.

• My recent articles on Eglobalis: https://www.eglobalis.com/blog/
• My recent articles on the European CX Organisations ECXO.org : https://ecxo.org/blog/
• My recent articles on CMSWire: https://www.cmswire.com/author/ricardo-saltz-gulko/
• My German articles on CMM360: https://www.cmm360.ch/author/ricardo/
• My old articles on CustomerThink: https://customerthink.com/author/rgulko/