For automakers, quality problems can quickly escalate. A defect that goes undetected or takes too long to diagnose can lead to unnecessary repairs, broad recalls, disgruntled customers, and mounting warranty costs. In many cases, OEMs may not become aware of quality issues until they are deluged with similar warranty claims, by which point it’s very expensive to contain the problem.
AI can help automakers escape this trap by monitoring connected-vehicle data and surfacing issues before they become costly, widespread failures. It can also bring greater speed and precision to root-cause analysis (RCA) and severity and scale estimations, allowing OEMs to focus their response on the vehicles that actually need attention. The result is lower warranty costs, improved uptime, and less disruption for customers.
Upstream, a behavioral-intelligence platform for automotive and physical AI, is putting these capabilities into practice with Proactive Quality Detection (PQD), its AI-powered, cloud-based, quality monitoring and detection platform. This system has already helped many OEMs reduce warranty costs—in one case, by 98.5 percent.
McKinsey’s Eike Reus, a partner in the Berlin office, recently spoke with Upstream Cofounder and CEO Yoav Levy about how AI is changing the way automakers approach after-sales quality, and what it takes to turn the growing volume of connected-vehicle data into faster, more effective decisions.
Catch it before it costs you
Eike Reus: What is the most significant effect that AI has had on quality management and warranties?
Yoav Levy: AI changed what “early” means. Early detection used to mean catching a problem after a few hundred claims instead of a few thousand. It was still a reactive response, just less delayed. Now you catch it before a single claim is filed, from signals alone.
But detection was never the only bottleneck. Investigation used to take weeks as engineers pieced together evidence by hand. Now, that information gets compiled automatically by AI. OEM responses used to default to recalls, which were time consuming, but now they resolve a growing share of problems over the air [OTA] then monitor the fleet to confirm that the issues don’t resurface.
Quality control is a single loop now: Detect, investigate, resolve—all ahead of the claim instead of after it. Once that loop closes, the whole economics of warranty, how much you spend, how wide a recall has to be, the impact on vehicle downtime, how much trust you preserve, also improve.
Eike Reus: How much can AI-powered solutions help companies accelerate RCA and reduce warranty and recall costs?
Yoav Levy: In our experience, RCA can happen roughly 30 to 90 percent faster, and warranty and recall costs can drop by 5 to 20 percent. The main improvement drivers are velocity and scale: cutting time to RCA and time to countermeasure, using machine learning [ML] to prioritize problems and precisely identify the affected vehicles so that you’re not replacing parts unnecessarily or running campaigns that are too broad, and making more OTA updates rather than pulling vehicles into service bays.
One more thing worth saying to a CFO: Almost all of the improvement relates to cost avoidance rather than hard savings. Since companies prevent the cost from materializing, we expect significant reduction in warranty and recall costs.
A new solution for automotive quality
Eike Reus: How does your PQD platform enable early problem detection? What signals does it read across a fleet? And how does it separate a genuine emerging defect from ordinary noise across millions of vehicles?
Yoav Levy: PQD is powered by Ocean AI, Upstream’s suite of ML, AI, and agentic workflows. It is specifically trained on automotive quality data. It monitors vehicle behavior at the vehicle, component, and fleet levels, looking at telematics, diagnostics, and quality data. It also determines how diagnostic trouble codes [DTCs] correlate with telemetry, repair orders, customer support, and even social media, rather than relying on any single signal. The system is cloud-based and agentless, with no hardware installed in the vehicle.
The platform uses two detection approaches that work in tandem. One identifies known failure patterns, and the other uses ML models to catch anomalies that engineers have not yet defined. Results from both analyses feed into the same pipeline well before a repair order or claim exists.
PQD finds true problems by looking for patterns. A fault code from a single vehicle might mean nothing, but a pattern in a group of vehicles, checked against a similar control group with no issues, is a different story. Through patterns, we can find a root cause grounded in real evidence, rather than making a guess. Quality issues with no existing fault codes are more difficult; we have to surface problems when there is no program or potential visibility.
The system clusters emerging patterns before the claim and ranks them by scale, severity, cost, and whether a fix can go over the air. It doesn’t make the final call; it puts evidence in front of an engineer, who decides on its significance. Rather than needing months of accumulating claims data to identify problems, we can surface them before the first claim is filed.
Eike Reus: What sort of quality issues has your system caught early?
Yoav Levy: I’ll give you three examples, and they prove three different things. In one case, there was a high-voltage battery cell imbalance on a newly released model. The OEM identified the root cause quickly, but more than 10,000 vehicles had already shipped by the time a countermeasure was ready.
This situation would normally require a blanket recall, with battery replacements running over $10,000 per vehicle. Instead, the OEM’s own quality team built custom telematics-based detectors on our PQD platform, and it determined that only a few dozen vehicles were defective and needed a fix. That represented a 98.5 percent reduction in warranty costs, resulting in $150 million to $200 million in savings. Since the OEM’s engineers built the model, it will always be available for future analysis.
The second example relates to a fuel pump drawing too much current from debris in the assembly. Normally, it would take around 45 days to connect scattered complaints about such problems to a root cause. In this case, our PQD platform flagged the issue early and correlated it with another existing investigation. The OEM found the root cause in eight days. That’s a good example of response velocity.
Finally, there was a case when an odometer spiked and never came back down, either because of tampering or a sensor fault. Normally, you wouldn’t find out about an issue like that until a buyback claim. We caught the problem within 24 hours, before anyone had created a program to detect it. The issue wasn’t on any watch lists; it was just behavior that didn’t make sense, and our PQD platform is built to notice that.
The common thread that unites these examples is that none of our customers still wait for claims to pile up before they act. I’m very proud of that.
Transforming the OEM quality culture
Eike Reus: When using AI to shift from a reactive quality operation to a proactive one, what are the biggest changes an organization and its quality teams must make?
Yoav Levy: The biggest change is behavioral, not technological. Quality teams have long relied on institutional memory, often in the form of a senior engineer who’s seen a pattern before. That works well until that person retires, and their knowledge goes with them. Moving to a proactive quality approach means asking thousands of people to start their day differently. They will open a briefing built from live signals instead of a claims report. That habit change takes longer than implementing any model. The rule that makes it stick: AI proposes, the engineer decides.
A lot of untapped value sits in unstructured data, technician notes, repair orders, and call transcripts, but most OEMs can’t use any of it at scale. With Upstream’s Ocean AI, one agent replays a car’s full history, another surfaces patterns across a cohort, and a third pulls multivehicle evidence into a single hypothesis instead of having an engineer stitch it together by hand. Engineers can still ask questions, but the heavy lifting happened before they logged in. Ocean AI never makes the final call about how to resolve a problem, and that’s exactly why engineers trust it enough to use it.
Eike Reus: Our research found that 85 percent of OEMs are still in the early or intermediate stages of consolidating quality data, and only 18 percent use AI in production to predict issues. When you begin working with a company, how much of the work involves analytics and how much involves fixing the underlying data and getting functions and suppliers to share it?
Yoav Levy: With a new customer, our work is 80 percent foundational and 20 percent analytical in the first few months. That tracks with what your research found. Therefore, the first phase typically involves bringing existing telematics, DTCs, repair orders, and claims together. We don’t force organizations to rebuild their systems but rather build on top of existing technologies.
During the early phase, sorting out who owns what data involves a longer conversation than anything related to models. Data is fragmented because ownership is fragmented. Nobody built three incompatible systems on purpose, but three functions each built a system, focusing on the metrics that were most important to it.
Tooling is also often a challenge. Most organizations are running homegrown, siloed systems that don’t talk to each other. We connect into existing databases so that teams don’t have to check five separate systems.
My advice for beginning an AI initiative is a little counterintuitive: Don’t settle the ownership question first. The data almost always exists; you need to bring it together, even imperfectly, and pick one problem to solve. I’ve watched a single result do more to settle an ownership dispute than six months of meetings ever could.
Eike Reus: Three-quarters of OEMs say they are experimenting with AI for quality, but only about one in ten has a solution they consider satisfactory. What separates the companies that scale AI quality measures across production from those stuck in pilots?
Yoav Levy: At companies that scale quality efforts, somebody owns the decision, not the model; pilots are scored on whether their predictions were right, and production is scored on whether a decision changed. The first use case gets chosen based on its ability for replication, not just impact.
Successful organizations can also tell if their AI tools get something wrong during a particular quarter and then make adjustments to strengthen trust. Teams that aren’t able to see a tool’s shortcomings won’t trust it and will work around it.
Companies that want to scale must watch out for a pattern that deceptively looks like progress: a strong pilot from a central data team, built without input from the quality organization that would actually use it. The pilot may produce impressive initial results, but it will not promote adoption, and it doesn’t prove that the new process can scale. Overall, the biggest reason pilots fail is that they’re not wired into an engineer’s existing workflow. A pilot may yield a useful list of anomalies, but if that list simply adds to the workload of an engineer who already has a full queue, nothing will change.
Eike Reus: What problems are companies encountering when they scale their quality systems to cover entire fleets? What are the keys to success?
Yoav Levy: Companies are generally going in the direction of using agentic, closed-loop quality systems that detect problems, scope exactly which vehicles are affected, investigate the issues, and trigger a fix, all with limited human intervention at each step. Applying these systems at fleet scale requires a coherent data environment, models people trust, clear decision rights, and human oversight on anything safety-critical.
Decision rights is the factor that tends to get underestimated and that is most difficult for OEMs to address. At pilot scale, ambiguity about who signs off is survivable: There may be ten decisions a month, and everyone relevant is in the room. At fleet scale, there are thousands of people involved—and ambiguity doesn’t just slow things down, it stops them.
Focus is also essential. We prioritize quality issues by predicting their scale and severity before committing resources. Teams can thus be more confident that their efforts will have an impact instead of chasing every anomaly the same way.
And we increase the odds of scaling by building quality systems with all layers in place: data, a live digital twin, AI, and agentic execution. While everyone wants agentic execution, it only works if the three layers underneath it are real. And when selecting our initial quality focus, we begin wherever impact will be most visible to create momentum.
Lessons learned and bold predictions
Eike Reus: What’s the hardest lesson you’ve learned about getting a conservative, safety-first industry to adopt a genuinely new way of working?
Yoav Levy: Automotive is rightly a safety-first industry that adopts new methods based on proof, not promises. In one case, PQD found something that a very good team had missed, based on their own data. After that, the value was clear. That experience taught us to start with a single lighthouse win and let the result persuade.
We built the PQD platform for automotive because this sector faces multiple quality issues. There are millions of connected vehicles, and the industry is both safety-critical and heavily regulated. But the same detect-and-correlate capability applies to any connected physical system, which is where we’re headed next.
Eike Reus: As vehicles become software- and AI-defined and get updated or repaired through OTA, how will the economics of vehicle quality change? Does a defect become cheaper to fix, or does it simply get addressed earlier?
Yoav Levy: Both. With OTA, a meaningful share of fixes involve a software update rather than a physical recall, which lowers costs. But software complexity creates its own failure modes, and a bad update propagates across a fleet in hours at a speed not possible in the physical world. The risk accelerates.
With OTA, a fix may cost a few dollars instead of a few hundred dollars. When that happens, OEMs might start fixing things that they would have just monitored before. But can you tell precisely which vehicles need repairs? That’s a detection problem, which is why OTA raises the value of detection.
Eike Reus: If proactive quality measures work as you envision, where will the automotive industry be in five years, and what routine processes will disappear?
Yoav Levy: Data-driven quality will become part of a continuous process embedded across the entire life cycle, from design and manufacturing through after-sales. Quality detection and resolution will occur far upstream, ideally during product development or manufacturing, instead of waiting for the vehicle to be in customers’ hands.
The months-long, root-cause investigations that we now consider normal will disappear. We won’t have broad recalls that sweep in vehicles that were never actually affected, and customers won’t discover issues before the OEM does.
Traditionally, OEMs have started quality investigations by looking at a failed part, but that doesn’t always lead to the right root cause. As with modern healthcare and aviation, there will be a paradigm shift in which OEMs look for the needle in the haystack and the “tells” in nonintuitive data and ML models.
The greatest change will relate to viewing warranty accrual as a fixed cost of doing business. Today it is a number you budget for and hope to hold flat. If OEMs detect quality issues early, before a claim is filed, the warranty accrual cost will become a number that OEMs manage. Finance will start asking quality engineering questions that the group has never had to answer.
Source: www.mckinsey.com
