Author: Alexander Nase

Inside the gap –
Why automotive engineering struggles with AI adoption

Why AI adoption in automotive engineering is slower than the headlines suggest – and what happens when engineers gain hands‑on capabilities
The R&D floor is still quiet when the first engineers arrive. As screens wake up and coffee machines hiss, one of the lead engineers opens a dashboard, not a simulation environment. During the night, three AI agents worked through her backlog. They checked supplier data against new regulatory requirements, scanned a thermal model for unstable parameter combinations, and drafted a memo for the noon review.

She frames the questions, sets the constraints, and inspects every result. She is an engineer who can orchestrate AI with the same discipline she applies to simulation, testing, or calibration. This capability exists only in isolated pockets today, yet it illustrates what many engineering organizations aim for.
Across the industry, most AI programs start with impressive pilots. One use case shows real productivity gains. A business case is validated. Scaling is approved. But once the work moves from controlled conditions into everyday engineering, progress slows. The gap between what is possible and what is practical becomes visible.

Why the bottleneck is not only organizational
While it may be easy to blame slow organizational adoption, that doesn’t tell the whole story. Foundation models are rapidly advancing, with OpenAI’s latest benchmarks indicating near-human expert performance in mechanical and industrial engineering tasks (see graph on the next page). In this benchmark, multiple top AI models and human experts were tasked with projects like creating CAD models with experts later evaluating the results side-by-side. In software engineering, AI is already outperforming human experts, with an 86% win rate. For mechanical engineering, AI is closing the gap with a 44% win rate. Nevertheless, current models cannot consistently handle all engineering workflows. Data remains scattered, tools lack uniformity, and requirements for validation and traceability differ widely across domains.

Automotive engineering adds another layer of complexity. Thermal constraints shape electrical design. Regulatory rules influence architecture decisions. Much of this context sits in different tools and formats. Agents can support coordination across these domains, but they cannot resolve the underlying structural fragmentation or interpret ambiguous data without careful human review.

In short, the models are getting close, but the workflows are not yet ready everywhere.

The next capability: Engineers building micro‑automations themselves
As model capability increases, one limit becomes clear. No central AI team can build the hundreds of subtle, domain‑specific micro‑automations engineers rely on every day. These tasks are too numerous and too context‑dependent. The scalable path is enabling engineers themselves to build simple agents on low‑code platforms, connecting tools, checking assumptions, and structuring data flows.

Many engineers already have enough technical intuition for this. They script, model, and configure tools every day. The primary barrier is not capability, but the initial inertia. Once they take the first steps, AI becomes a teacher. Engineers use agents to refine their own agents, debug logic, restructure instructions, and add guardrails. A form of recursive self‑improvement emerges: each iteration increases confidence and capability.

This dynamic is what makes decentralized AI adoption realistic.

A live experiment in making adoption real
In 2025, FEV Consulting tested this challenge internally. Instead of solely developing a small expert hub for AI, our firm designed an experiment around a simple question: what happens when everyone learns to build, evaluate, and improve AI agents themselves?

The structure was deliberate. For the first six weeks, each consultant built an agent in Microsoft Copilot independently. No shared examples, no group sessions. Working alone reduced anchoring on early adopters and forced genuine engagement.

After the first phase, each agent received detailed individual feedback. It showed participants how to validate outputs, implement guardrails, and structure instructions so agents behaved consistently, even with partial or noisy inputs. Quality and trust began to form at the same pace.

During the second phase, the focus transitioned to collaborative efforts. Participants worked in pairs to evaluate their respective agents, integrate optimal design elements, and develop enhanced versions. These agents demonstrated notable performance improvements. Additionally, this approach promoted a culture of collaborative experimentation and shared learning.

By the end, several agents became daily tools inside the practice. More importantly, the entire team gained the confidence to build small automations themselves, the same competence engineering organizations now need at scale.

Underlying all of this was a principle that applies far beyond this challenge. In every engineering context FEV Consulting operates in, including its own consulting practice, AI does not own results. Humans do. Every output requires inspection and deliberate acceptance by an expert. AI accelerates analysis, but responsibility never shifts.

Where the industry reality stands today
A few OEMs have already started enabling engineers to build low‑code automations. Their adoption curves look different. They move faster when new AI capabilities emerge. Their toolchains become more connected. Their engineers know where to rely on agents and where to challenge them. These are early signals, not yet industry‑wide patterns, but they point toward a meaningful shift.

For most organizations, progress is slower. Data readiness varies. Integration remains difficult. Safety and compliance require rigorous human oversight, and engineering culture evolves carefully, as it should.

Yet the direction is becoming clearer. AI will not replace engineers. But engineers who can orchestrate and build small AI tools will shape how development organizations work.

The moment before momentum
Back in the R&D office, the engineer closes her overnight dashboard. Every agent output she sees is traceable and reviewed. Every recommendation is examined, as the final responsibility remains hers.

This workflow is not everywhere yet. It requires data maturity, integration, guardrails, and a workforce trained to evaluate AI output with engineering discipline. But it is beginning to emerge in organizations that invest early in hands‑on capability.

The technology is accelerating quickly. The question now is whether engineering teams are equipped to keep pace.