From Perception to Decision: Deploying AI Inside a Live Claims Workflow [RSVP REQUIRED]

This event requires an RSVP; to attend, submit form here

Every day, adjusters open tire and brake photos and make a call: approve or deny. At 300,000+ claims a year, the problem isn't whether any single adjuster gets it right. It's that a significant chunk of their day is spent on tasks that don't require their expertise: manually validating photo quality, chasing missing gauges, calling repair facilities to coordinate resubmissions.


We built Screech to fix that. It's a two-phase AI system embedded directly into our warranty claims workflow at SilverRock. Phase one validates incoming photos. Is the gauge visible? Is the image usable? If not, it triggers the resubmission request automatically. Phase two takes the validated photos and runs them through an ML model that generates an approve or deny recommendation. Claims that meet defined happy-path criteria are auto-decisioned. Everything else goes to a human adjuster, recommendation in hand, who makes the final call.


What surprised us wasn't the technology. It was the math. We're spending cents per claim to govern decisions worth thousands in severity, while freeing skilled adjusters to focus on work that actually requires them. At our volume, those shifts compound into a multi-million dollar annual impact. That's the case for embedding AI at scale: not dramatic transformation, but small, disciplined automations applied where the numbers are big enough to matter.


We'll walk through how we built and validated the system, where we drew the line between autonomous decisions and human oversight, and what it took to get a model making real decisions inside a live production workflow. Then we'll open the floor, because the hard questions around trust, thresholds, and when AI earns the right to decide are worth discussing out loud.