The University of Pennsylvania deployed a commercial AI auto-contouring system across its 15-site radiation oncology network starting in 2022. The system saved roughly 15,000 clinical hours over three years, but the resulting efficiency gains created downstream bottlenecks rather than reducing overall treatment planning time.
The bottleneck effect of AI efficiency
Rafe McBeth, PhD, director of AI at Penn, detailed the deployment at the Society for Imaging Informatics in Medicine annual meeting in Pittsburgh on June 12. The institution achieved near-universal clinical use, with AI-generated contours applied to 95 percent of routine cases. The system processed over 4,000 patients and returned Dice scores between 0.80 and 0.85 across hundreds of structures.
Saving about 30 minutes per patient for auto-contouring organs at risk accelerated patient arrivals. This speed-up exposed capacity limits in later stages of the workflow instead of shrinking the total treatment planning cycle.
Build versus buy decisions and vendor alignment
Commercial AI vendors are not prioritizing the contouring of tumors and targets in ways that match institutional needs. Evaluating AI for Operations requires looking past initial deployment metrics to determine if vendor roadmaps match long-term clinical goals. This misalignment often forces hospitals to reconsider whether to purchase external tools or develop proprietary solutions.
Building AI-native teams
The primary lesson from the Penn deployment is the necessity of dedicated, AI-native operational teams. Time saved by the software cannot be treated as pure labor reduction. McBeth explained that these reclaimed hours must be reinvested into building the teams required for subsequent AI rollouts.
"As we get to two, three, four, or 10 models deployed into the clinic, well, now you have a problem that looks a lot closer to your other clinical problems that require the same teams that have been kind of upscaled into AI to provide their perspective," McBeth said. This scaling challenge is a core component of AI for Healthcare, where integrating multiple models demands sustained operational oversight rather than one-off cost savings.
Why this matters for operations
Operations leaders must anticipate that automating one step in a workflow will simply push volume to the next constraint. Time savings from AI tools should be budgeted toward growing specialized support teams, not just cutting costs. Procurement decisions must weigh vendor development roadmaps against internal capacity to manage future model deployments.
Your membership also unlocks:
