Real-time dental image verification with Amazon SageMaker AI at Henry Schein One

Dental offices have a quiet, expensive problem: up to a fifth of insurance claims get denied at first pass because the X-ray attached to them was blurry, cropped wrong, or just missing something. Nobody finds out until the claim bounces back, days or weeks after the patient has already left the chair. Henry Schein One decided to fix the timing of that feedback loop rather than the paperwork around it, and the result is a system called Image Verify that grades an X-ray's quality on a 1-to-5 scale the moment it's captured.

What's striking here isn't the concept, quality checks aren't new, it's the speed of the rollout and the engineering underneath it. The product existed only as an idea at the end of Q3 2025. By spring 2026 it was running in over 10,000 practices, up from 250 at launch, a 43x jump in a matter of weeks. It's now chewing through 1.5 million X-rays a week and has processed more than 20 million total, all while returning a verdict to the technician in a median of 1.4 seconds. That's fast enough that a tech can just retake a bad shot on the spot instead of calling the patient back in.

The technical story is really about finding the right bottleneck. Henry Schein One runs the whole thing on Amazon SageMaker AI, with a multi-model pipeline that first classifies the image type, then runs specialized quality checks for sharpness and alignment, then merges the scores. Early on, engineers assumed they'd need more GPUs to keep up with growth. Profiling showed the opposite: preprocessing, decoding and resizing images, was hogging CPU while GPUs sat half-idle. Once they moved that work onto the GPU and switched to async inference with queue-depth-based autoscaling, they cut the fleet from 15 instances to 10 while latency actually improved. Swapping ml.g6e.4xlarge boxes for the newer ml.g7e.4xlarge shaved median latency down from 1.69 to 1.43 seconds too.

None of this required regulatory sign-off, which is part of why it moved so fast. Image Verify doesn't diagnose anything; it just answers a narrower, less fraught question about whether an image is usable. That let the team ship changes through A/B-tested, zero-downtime deployments multiple times a day, sometimes working through a backlog of more than 60 identified optimizations one item at a time. The GPU fleet now runs around 70% utilization, and the architecture is already validated to handle Henry Schein One's stated target of 40,000 locations across the US, Europe, Canada, and Asia Pacific, meaning the current footprint is roughly a quarter of where they're headed.

There's also a small behavioral hook baked in: technicians can see their own quality scores, which turned into an informal gamification layer that seems to be driving adoption without anyone mandating it. That's the kind of detail that tends to matter more for real-world uptake than any benchmark number.