Only five percent of automakers are expected to maintain ambitious AI investment programs through 2029, building the engineering foundations now that will separate them from competitors. Such companies are at the vanguard of the industry, pursuing disruptive value, as they have laid the groundwork needed to sustain it.
Software-Defined Vehicles (SDVs), where vehicle functionality is delivered primarily through software rather than fixed hardware, are at the center of this ambition. However, for many engineering programs, the strategic question is how to build the engineering workflow infrastructure that can make AI-driven software development actually deliver.
In established Original Equipment Manufacturing (OEM) programs, the development lifecycle still depends on manual handoffs between teams and tools. Engineering data lives in fragments across separate systems, with requirements in one tool and design decisions in another. Code, tests, and compliance evidence live somewhere else. When a change arrives, teams must manually piece together what is connected, what is impacted, and what validation evidence is required before release. Without live traceability, that reconstruction demands significant engineering effort. Late changes force teams to regenerate validation deliverables across multiple testing environments, introducing delays and inconsistency risk.
Legacy programs carry an additional constraint. Most conventional OEMs manage ICE, hybrid, and electric vehicle lines simultaneously, each with its own software platforms. The overhead compounds. Each product line adds its own coordination burden, making consistent tooling across the portfolio difficult to maintain.
An agentic AI engineering workbench addresses both problems. The lifecycle becomes a continuous workflow. Traceability updates automatically. Compliance is enforced at each step rather than saved for gate reviews.
Build a live traceability layer, not a reporting artifact
A live traceability graph maintains relationships between requirements, design decisions, code versions, and test outcomes in real time as engineering work proceeds. A change arrives. The graph shows what is affected and what evidence already exists. Any relationship in that graph can be queried directly, not reconstructed by hand.
Most OEM programs reconstruct this chain at gate reviews rather than maintain it continuously. Functional safety and process compliance frameworks both require it to be auditable and current. A live graph delivers this as work proceeds, not as a pre-release exercise. The same graph extends across ICE, hybrid, and electric vehicle lines, so a shared component’s traceability does not have to be rebuilt separately for each platform.
Govern AI agents across the workflow
An agentic engineering workbench orchestrates multiple AI agents across the development lifecycle, each with defined scope and standards references, all sharing one traceability backbone. Human engineers review and approve proposals at defined decision points. Human-in-the-loop (HITL) is an architectural feature. Engineers actively review each agent action rather than receive notifications after the fact. Standards are enforced, not advised. Through Retrieval-Augmented Generation (RAG), the applicable standards are retrieved directly at agent invocation time and applied as hard gates. Exceptions require named human approval. This is how governance becomes enforcement rather than guidance.
Make governance the control plane
The case for governance-by-design in agentic AI is no longer theoretical. In automotive specifically, the concern is acute: UNECE R155 requires manufacturers to run a certified cybersecurity management system across the vehicle lifecycle, and ISO 26262 requires software rigor calibrated to safety risk. Uncontrolled automation is incompatible with both. In these environments, governance is the procurement gate, not a productivity checkbox.
The principle is evidence before autonomy. Transparency and coordination improvements come first. Automation in higher-risk functions is phased in, as trust is established through outcomes, not assumptions. An agent working on a function classified at a higher Automotive Safety Integrity Level (ASIL) tier operates with less autonomy and more mandatory checkpoints than one working on a lower-risk function. Governance is the control plane the architecture is built around. By 2030, as per recent research, half of all AI agent deployment failures will result from insufficient governance runtime enforcement. For automotive programs, governance built in from the start is a procurement requirement. When it is embedded from the design stage rather than added later, the productivity outcomes follow. Infosys’s own SDV toolchain work already shows this: one forklift manufacturer cut build and deployment turnaround by roughly 40 percent and saved close to 670,000 dollars a year after standardizing its toolchain.
SDV programs succeed or fail on engineering workflow discipline, not tool access. The time-to-market divide in automotive software development is real and is not closing on its own. Most established OEM programs are still running development processes designed for hardware-defined vehicles. Adapting those processes to SDV cadences requires more than just deploying AI tools.
Going forward, successful programs will embed traceability and governance infrastructure into engineering workflows. The workbench that governs engineering is the infrastructure on which competitive release cadence is built. Infrastructure is the long game in a sector where release speed determines market position.
