Maintenance discipline
Fix first, predict later: Five predictive maintenance lessons from AMNA 2026
Lucid, Tenneco, Gestamp, Autoliv and Ford discussed where predictive maintenance pays back and where it stalls. The consensus was that prediction is the last step in a plant's digital journey, not the first.
Unplanned downtime has become far more expensive for carmakers. Siemens’ True Cost of Downtime 2024 report puts an hour of lost production in a large automotive plant at $2.3mn. That is a 113% increase between 2019 and 2023, against US inflation of 19% over the same period. Large plants still lose an average of 27 hours a month to unplanned downtime.
With capital budgets tight and experienced maintenance staff retiring, plants need to get more out of existing equipment, and predictive maintenance is an obvious place to start.
At Automotive Manufacturing North America (AMNA) 2026 in Detroit, experts from Lucid Motors, Tenneco, Gestamp North America, L2L, Autoliv, Kistler, Ford and Continental discussed where predictive maintenance pays back, where it stalls, and what plants need in place before investing in it. Here are five lessons from those discussions.
Avoiding downtime is a cheaper lever than new capex
Speakers agreed that North American plants face a utilisation problem, not a capacity problem. Arun Prakash Mantena, global manager for advanced manufacturing engineering (stamping) at Lucid Motors, said manufacturing leaders should look first at single-minute exchange of die (SMED), changeover times and predictive maintenance.
In his view, expensive unplanned downtime carries more risk than deferring new capex.
“The predictive maintenance is the biggest thing that we are trying to focus on currently.”
Marla Stern, manager of manufacturing excellence at Tenneco, treated maintenance as a precondition for extending asset life. “It’s not worth it for us to modernise an asset in a plant that doesn’t have good maintenance,” she said. Without good preventive maintenance, she warned, a company can spend twice the price of an asset bringing it back.
Tenneco’s AI budget this year is going into small modules, preventive maintenance among them, rather than large bets with returns it cannot predict.
Stabilise before you predict
Mohan Swaminathan, operations director for the US region at Gestamp North America, set out a different order:
Pick a bottleneck or critical asset.
Stabilise it, standardise it, and digitise it, storing its data securely in the cloud.
Connect people, process and technology.
Only then is a line ready for machine vision or predictive maintenance. “For us, this is a very sequential process.” Swaminathan said.
Gestamp’s intermediate step delivers value before any prediction. Operators once copied PLC parameters by hand at the start of each shift. Connected machines are now monitored around the clock.
A parameter out of tolerance raises a work order for the responsible engineer, and a critical deviation shuts the machine down. That is condition monitoring rather than prediction, but it builds the history a model needs.
Eric Whitley, senior manufacturing solutions engineer at L2L, said manufacturers are “data rich and action poor”. L2L’s failure-analysis software flags PMs completed late, or a spare part that has failed three times but is missing from the PM schedule.
Why predictive models stall
Nitin Sethi, vice-president and head of IT at Autoliv Americas, cited a study finding that 88% of AI proofs of concept never reach wide deployment. He described two routes to predictive maintenance:
Building a Pareto of likely failure causes from historical data.
Reading live machine data to detect that a failure is coming.
The harder problem is calibration. Scheduled maintenance is itself costly, so a model that triggers unnecessary jobs undermines its own business case. Deciding which process parameters genuinely warrant intervention needs maintenance, quality and production to tune the model together.
“AI is not a magic bullet which is going to be right every time.”
Data readiness is the other test. Sethi said a plant is not ready if it plans a model on outdated machines that produce unreliable data, with no way to ingest that data and no MES connection. Waiting for perfect data is also a mistake, he added, because AI can now help curate it.
The return is in material, not labour
Sethi’s strongest example linked maintenance to quality. Autoliv’s computer-vision inspection project, which cost almost $1mn, detects defects. The next step records the machine’s process parameters at the moment each defect appears. Once the system learns which drift comes before defects, it signals maintenance to plan an intervention before bad parts are made.
In his account, the payback comes from parts not scrapped rather than inspectors replaced. That also changes how projects get funded: a request for money “for AI” tends to be rejected, while the same sum tied to a stated saving gets approved.
Other speakers took a similar line. Hiroyuki Natsume, sales director for industrial production at Kistler Group, said condition monitoring is easier to adopt than more ambitious projects, because it costs less and disrupts plant routines less.
Rajeev Kalamdani, director of manufacturing and IIoT analytics at Ford, was asked which single project he would fund. He chose predictive quality because it is “closest to the customer”.
Ownership decides adoption
An accurate model still fails if nobody has to use it. Sethi said an Autoliv bot for finding the root causes of OEE losses was highly accurate but reached only 20–30% adoption, because it sat outside standard operating procedure.
An early vision-inspection model stalled at 30–40% accuracy while IT owned the data labelling, and improved once operators labelled data through a simple interface. His conclusion: the output must trigger a workflow, and a plant manager, not IT, must own the result.
Knowledge is the other gap. Knowledge transfer ranks among manufacturing leaders' top concerns for the next five years, with much of it held by retiring maintenance staff. Continental stores plant incidents, fixes and PM documentation in a global database that employees have been able to query with an internal AI tool since 2025. Paired with predictive alerts, Sethi argued, that knowledge lets AI suggest a remedy rather than just flag a risk.
Lazola Tutani, production department manager at Continental in South Carolina, said Continental has made its global database of quality incidents and maintenance documents searchable with AI.
