The test car is a Jeep Grand Cherokee with a cracked windshield and chunks cut out of its side-view mirrors where the cameras go. Joydip Saha, who manages Stellantis’s hands-free driving project, calls it the most broken-up car on the lot. A yellow box on the dashboard is a kill switch: one press and the driver has the wheel back. In the parking structure under the Chrysler Technology Center in Auburn Hills, Michigan, the system beeps when it needs him, and it is confused. The garage is a diamond-shaped maze, and the car heads for the wrong exit. Saha puts his hands on the wheel. Out on the suburban roads the same software handles an intersection where two other drivers make illegal lane changes at once.

Saha drives every new software build himself, as often as twice a week. He told Fortune’s Nick Lichtenberg, who spent two days at the centre and published the account on 1 November, that his team built the Cherokee’s hands-free system in six weeks with the help of AI tools. A similar project used to take a year or even 18 months. He also has a rule: “I will not let people use AI to architect stuff because they will forget how to debug stuff.”

A four-year cycle meant to fit in two

The speed matters because Stellantis needs it. Fortune reports that the company wants to cut vehicle development from four years to 24 months, starting with two models in 2027. It runs 14 brands, its European business is losing money, and its second-quarter adjusted operating margin was 1.8%, against a target of 7% for 2030 set by chief executive Antonio Filosa. In February the company wrote down roughly 22.2 billion euros, about 26.5 billion dollars, on its electric-vehicle bet, and its shares fell 25% that day. They are down nearly 60% this year, Fortune says.

Inside the building Iacocca opened in 1991, the engineers describe specific savings. Ben Zawodni, a software engineering manager who is folding hundreds of separate electronic control units into one central computer called STLA Brain, says a proof of concept for a global observability platform took a week. Before agentic coding tools, he recalls, it would have taken about six months. Defects that took two to three weeks to fix now take two to three days.

In aerodynamics, Arturo Guzman’s team runs a trained model on a design change to a Ram pickup’s front splitter and gets a drag estimate in about three minutes. The full simulation took 14 to 24 hours. Guzman, who brought AI into the group about five years ago, takes the mystery out of it for vice presidents: “There’s no miracles. There’s no magic.” He calls the model an intern. The company’s wind tunnel, meanwhile, still tests typically one car a day, as its supervisor Steve Kubiak puts it. Only the analysis has sped up: a colleague hands an assistant the data from the last 20 runs and asks what it sees.

These figures are the engineers’ own, relayed by a reporter, and they describe single projects rather than the company’s development cycle as a whole.

Harder, not easier

The same people complicate the story. Zawodni says the agentic tools have made his work much easier, and then asks the question himself: “So why am I working so much harder?” He estimates he works at least twice as hard as before and feels “a tremendous sense of urgency.” His team, he says, is a mixed bag. “I think there was a starting gun, and that starting gun went off,” he says. “Some people heard it, and some people didn’t.” Of his engineers he adds: “Engineers are really afraid of losing their core skills.” They want to know why a semicolon does what it does. He calls that a valid curiosity and wants them to become architects instead.

Saha’s rule points the other way, and the two managers are in the same building. One wants the architects to emerge faster; the other forbids the tool to do the architecture because the engineers must still be able to answer every question in a safety-critical product.

Others describe a gain in quality rather than time. Mel Roberts, who plans infotainment features five to eight years out, says her AI research tool returns 10 to 15 sources and she has to check every one. “I don’t think it saves time,” she says, though it makes her work more robust. Amanda French, who leads platform teams, uses AI to pull together requirements and bills of materials held in regional systems that were never built to talk to each other, a legacy of the company’s many mergers. Both say human judgment stays. “The AIs aren’t going to tell us what the new platforms are,” Zawodni says.

The pattern is not unique to Detroit. On 20 October the Linux Foundation published a survey of open-source projects in which 59% of maintainers reported growing contribution volume since the rise of generative AI, and an estimated 25% of incoming pull requests were low-quality. Faster output at one end of the pipeline turned into review work at the other. At Stellantis the review work falls on the same engineers who write the code.

Redundant, and hiring

Ned Curic, the chief engineering and technology officer, who came from Amazon, Microsoft and Toyota, is selling an idea he knows is hard: “the best thing that you can do for your career is to make yourself redundant.” He says he has done it himself several times, by getting a team into good shape and moving on, and that the company must always have a new opportunity ready for the people who do. He also told Bloomberg at the Paris auto show in October, Fortune reports, that Stellantis would add “multiple hundreds” of engineers in Europe and the United States this year.

Curic is candid about where the company stands. Some teams show “phenomenal successes,” he says, but “as a whole organization, we are not there.” Anne Laliron, who runs the roughly 400-person technology research group, adds an observation about people: veteran engineers carry old failures with them, and sometimes “their brain is closing doors.” Young engineers with racing rigs in their basements took to the driving simulator at once, while test drivers with 10 or 20 years in physical prototypes were slow to follow.

What four years might settle

Alison Rahm, chief vehicle engineer for midsize and large SUVs, remembers a staff meeting in 1997 at which her boss told the room they were getting something new called email. “All of us were in a total panic,” she says. They went back to their desks and tried it. Now AI runs checks she used to do by hand, such as going down a list of about 1,000 engineering design practices, which leaves engineers free to ask whether the rules themselves are right.

She is not ready to stop double-checking. After virtual crash testing arrived, she notes, engineers kept crashing real cars into walls for years to convince themselves their assumptions were right. Asked how long it will take to trust AI the same way, she paused and answered that it would take another life cycle, “I would say four years.” That lands in 2030, the year Filosa has promised the 7% margin.

References

  • Boysel, S., & Lai, A. (2026). Navigating Open Source Software Development in the Generative AI Era: Challenges and Opportunities for Project Communities (announced by the Linux Foundation, 20 October 2026). prnewswire.com
  • Lichtenberg, N. (2026). ‘The best thing that you can do for your career is to make yourself redundant’: Inside Stellantis’ AI push after the ‘starting gun went off’. Fortune. fortune.com