LLM-Powered Workflow Optimization: From 5 Hours to 7 Minutes at Volvo
The most compelling evidence for LLM-powered workflow automation often comes from the least likely places. A new engineering paper from Volvo Group documents what happened when a multidisciplinary automotive software team replaced manual coordination steps with LLM agents — not through a big-bang rewrite, but through a graph-based approach that mapped existing workflow steps and replaced them incrementally. The target was spapi, a production in-vehicle API system spanning 192 endpoints, 420 properties, and 776 CAN signals across six functional domains.
The core problem the team was solving is one most engineering organizations will recognize: domain experts and developers work in fundamentally different formalisms. CAN signals, functional specs, and API definitions don't translate cleanly between disciplines, so every handoff spawns clarification rounds and error-prone manual reconciliation. The graph representation Volvo developed identifies exactly which coordination steps are ripe for LLM replacement without disrupting the processes that still require human judgment. The result: a 93.7% F1 score on automated outputs, per-API development time dropping from roughly five hours to under seven minutes, and an estimated 979 engineering hours saved — with high developer satisfaction scores.
What makes this paper particularly valuable isn't just the numbers, striking as they are. It's the incremental adoption model itself: a practical, low-disruption template for any engineering team in a domain-heavy environment who wants to automate workflow coordination without betting the entire process on a single LLM rollout. The pattern transfers directly to anyone building across incompatible technical formalisms — which, in complex software organizations, is nearly everyone.