AMA Insights  |  Agentic AI in Automotive

From SDV to AIDV

Everyone in automotive is suddenly using the term AI-defined vehicle, and most of the people using it mean different things. Before your next budget cycle, it is worth pinning down what the term means, where the technology is producing measurable results, and where it is still a slide in a vendor deck.

The one-word difference

A software-defined vehicle and an AI-defined vehicle sound like the same idea wearing different jackets. They describe different levels of maturity, and confusing them is where a lot of AI budgets go wrong.

Software-defined vehicle (SDV)

A vehicle whose functions run on software and can be updated remotely after it leaves the factory. Every decision still follows rules engineers wrote in advance. The software changes often, and its behavior in any given situation stays the same.

AI-defined vehicle (AIDV)

A vehicle that also carries AI models which interpret what is happening in and around it and choose a response in the moment, within limits the manufacturer defines and certifies.

The word that matters is interpret. An AIDV reads a situation it was never explicitly programmed for and makes a judgment call: a cabin assistant noticing a child left in the back seat, a car preparing its battery for a trip it knows is coming, a driving system recognizing an unusual road scene. Flip the switch below and watch what changes.

See the vehicle as:

Software you can update. Predictable every time.

How it decides
Follows rules written in advance
Software updates
Standardized over-the-air updates
What the data must support
Reliable collection of vehicle data
Business risk
Moderate and well understood

A vehicle that follows the rulebook. Predictable by design.

The sequence between the two is the real story. AI-defined behavior only works when the software-defined layer underneath it is dependable, because a model making judgment calls on unreliable data will make unreliable judgments. Most OEMs are still in the middle of the SDV build-out, and the market is already selling them the next layer.

Follow the money

The spending has moved faster than the definitions. The automotive AI agent market sits at $4.01 billion this year and is projected to reach $10.31 billion by 2035, while the global agentic AI market is on a much steeper climb. Bosch alone plans to put $2.9 billion into AI over the next two years, much of it in agent-based systems for vehicles.

Automotive AI agents

Market size, $ billions

Global agentic AI

Market size, $ billions

Now, the reality check

0%
of agentic initiatives that Gartner expects to be canceled by 2027
0
vendors, out of thousands claiming agentic AI, that offer genuinely autonomous systems
0%
of automakers increasing their AI budgets, while organizational maturity moves far more slowly

Who can keep paying for this?

Share of automakers able to sustain their current pace of AI investment (Gartner)

Reported gains where it works

Selected results from production deployments, %

Where it is working today

Strip away the relabeled automation and four use cases are producing measurable results in production right now. Each one has the same shape: a known signal, a defined response, and a measurable cost of delay. Tap each card for the evidence.

McKinsey documents a tier-one supplier that connected an AI workflow to its own database of historical requirements and approved tests. Work that used to take 30 minutes to 4 hours per requirement now wraps up in minutes, with the biggest gains going to less experienced engineers and the payoff landing in safety-certification work. The engineer stays responsible for whether the test proves the requirement; the agent just removes the blank page.

GM's WeldBrAIn deployment at Factory ZERO moved welding inspection from periodic sampling to continuous, real-time evaluation of the whole production run. The system sees more of the line and spots drift sooner, which gives the team time to correct the process before one bad weld becomes a batch problem, with all the scrap and containment cost that follows.

The value here is decision latency: the time between spotting a disruption and acting on it. A General Mills deployment evaluates more than 5,000 shipments a day and has generated over $20 million in savings by shortening that cycle. Deployments using this model report a 34% gain in supply chain efficiency and 19% higher ROI than traditional automation. The catch: the agent needs one consistent view of what is late, what is available, and which alternatives are allowed, and every recommendation should come with a receipt showing the data, the rule, and the approval behind it.

Predictive maintenance moves the starting point of service earlier, to the moment vehicle data first shows a failure becoming likely. The agent estimates how soon a part may fail, finds a service slot, and checks whether the part will be on hand. Salesforce found 70% of U.S. drivers are willing to trust AI agents with real-time diagnostics, and Gartner projects that by 2029 agentic AI will resolve 80% of standard customer requests on its own, cutting operating costs by 30%. For fleets, every avoided roadside failure keeps a revenue-generating vehicle on the road.

Autonomy is earned in three tiers

A process should have only as much autonomy as its data and controls can safely support. The practical way to manage that is a ladder with three rungs, and the honest answer for most automotive processes in 2026 is that they belong on the first one.

Tier 1: Watch & alert

What it does
Continuously monitors equipment, energy use, and quality, and flags issues for a person to review.
What it needs
Reliable sensor data with versioned alert thresholds.
Adoption in 2026
Highest. This is where most successful programs start, because operators keep control of the response.

How to start without becoming a statistic

The programs that survive share a sequence. The ones in Gartner's cancellation forecast usually skipped a step.

  1. Check the foundation

    Can the data move cleanly, can the systems hold steady, and can the team reverse a bad update? Name owners for every data source before anything else.

  2. Start with monitoring

    Watch-and-alert is the lowest-risk entry point. The agent observes while operators keep control of the response.

  3. Define the control model early

    Set the autonomy tiers, the human approval points, and the decision record format during design, so every later pilot inherits the same framework.

  4. Pilot one bounded workflow

    Pick the use case where the signal is reliable, the response is defined, and every hour of delay has a measurable cost.

  5. Scale against proven stability

    Grant more autonomy based on measured results from the previous stage, never on a target date.

The best first use case is one where a slow decision already has a visible cost. Map where decisions sit waiting across your operation and start where the waiting costs the most.

Take the first use case to production

AMA finds the workflow where a slow decision costs your business the most. Devox Software proves whether it is ready for production. Together, we take the first use case live and show you when it is ready to scale.

Talk to the team