The strategies that helped perception teams do more with less — and what's coming next.
2025 was a year of recalibration for the autonomy industry. Tariffs, economic uncertainty, and tightening budgets forced teams to rethink how they build training data pipelines. The companies that thrived weren't the ones with the biggest budgets — they were the ones that got smarter about how they spend.
We worked with over 70 autonomy programs this year. Here's what we learned about how the best teams cut costs without cutting corners.
Across our customer base, teams implementing intelligent annotation workflows achieved an average 35% reduction in annotation costs compared to traditional manual approaches.
What changed in 2025: Teams stopped treating pre-labels as finished annotations. Instead, they used them as intelligent prompts — guiding automation features like interpolation, one-click fixes, and smart frame selection.
Zenseact's results tell the story: 11.7 MSEK saved through pre-annotations, with 80% fewer objects processed by removing already-accurate classes from human review.
Manual QC was eating engineering time. One customer told us their perception engineers were spending 40% of their week reviewing annotations instead of building models.
Budget predictability became non-negotiable. Teams demanded clear pricing tied to actual effort — not black-box per-frame rates.
Pricing scales with how much annotator work is required. Full creation costs more than adjustment. Adjustment costs more than verification. As models improve, costs automatically decrease.
This aligned incentives: better perception models directly translate to lower annotation costs.
Consolidation accelerated. Teams moved away from managing multiple annotation vendors toward unified platforms that combine human expertise with intelligent automation.
China became a battleground. Localization requirements and data sovereignty pushed companies to build region-specific annotation capabilities.
Quality standards tightened. TISAX Level 3, SOC 2 Type II, and GDPR compliance became key for enterprise programs.
Pre-SOP validation grew. More teams invested in validating perception systems before start-of-production, catching issues earlier when they're cheaper to fix.
The teams we talk to are focused on three priorities:
1. Closing the feedback loop. Connecting annotation workflows directly to model training pipelines — so improvements in one flow automatically to the other.
2. Edge case coverage at scale. Finding and annotating the rare scenarios that break perception systems, without manually reviewing millions of frames.
3. Multi-modal complexity. Camera, LiDAR, radar, and now thermal — all fused into coherent training data at production scale.
The companies that figure this out won't just cut costs. They'll ship safer autonomous systems, faster.
Ready to see what's possible?
We've helped 70+ autonomy programs optimize their annotation pipelines. Let's talk about yours.
Kognic has delivered 100+ million annotations for the world's leading autonomy programs. We help machines understand the world through collaborative intelligence — where people, process, and platform work together.