Alberta Health Services built its own AI clinical scribe rather than buying one, and the architecture decision turned out to be a data sovereignty argument dressed in medical software. The tool, Jenkins, runs inside AHS’s own cloud environment, has logged over 80,000 patient sessions across 12 specialties, and delivered roughly 20% productivity gains per shift during an emergency department pilot. The team then open-sourced the underlying architecture as Berta under Apache 2.0, so any organization can run it on its own infrastructure. Canada’s new AI procurement framework<\/a> now points to this model as the template to scale nationally.<\/p>\n The two questions Ross Mitchell wants patients asking their doctors, where does the data live and is it training the vendor’s model, are the same questions a CIO should be walking into every AI vendor renewal with right now. If your organization can’t answer both from the contract, not from a sales call, you have the same exposure the clinics have. The divide isn’t between healthcare and other industries; it’s between organizations that built data control into the architecture and those that outsourced that decision to a vendor agreement they haven’t fully read.<\/p>\n Ontario’s auditor general found that all 20 provincially approved AI scribes had accuracy problems in testing, nine fabricated information outright, and 17 missed critical clinical details. The CIO lesson isn’t specific to scribes. Commercial AI tools fail in ways the vendor’s demo never surfaces, and when they do, the accountability lands on the buyer, not the vendor. Mitchell’s team sidesteps that by treating hallucinations, omissions, and factual errors as three distinct failure modes with different causes, each requiring physician sign-off before output enters the record. Most enterprise AI deployments haven’t gotten that granular about their own failure taxonomy.<\/p>\n The open-source release reframes what “build vs. buy” actually means at this moment. Berta isn’t a research artifact; it’s a deployable reference architecture that any organization with an AWS account can run, at under $30 per seat per month by Mitchell’s estimate. For a CIO defending an AI deployment to a board or a regulator, the difference between “we run this on our own infrastructure and here’s the audit log” and “we trust the vendor’s data handling” is the difference between a governance posture and a hope. I’d revise this read if the open-source path proved too operationally demanding for organizations without AHS-scale technical teams, but the cost and sovereignty case is strong enough that the burden of proof has shifted to the buy side.<\/p>\nWhat this means for your business<\/h2>\n