Two findings from the same C.D. Howe Institute session, published in late August:
Supply Ontario evaluated 20 AI scribe products. All 20 produced some incorrect medical information: hallucinations or incomplete documentation.
Meanwhile, when Canada Health Infoway offered primary care clinicians a funded AI scribe licence, more than 10,000 registered within two weeks. In an independent evaluation, roughly 90% called the tools valuable and nearly 70% reported less administrative burden.
Both things are true. That's the uncomfortable part.
The demand is real because the problem is real: documentation burden is eating clinical time. And the errors are real because every one of these tools can get things wrong.
The easy conclusions are the wrong ones: "AI isn't ready" or "the errors are worth it." The working group landed somewhere more useful: the risks of under-investing in healthcare AI are likely greater than the risks of adoption, provided accuracy, bias, and security safeguards keep pace with deployment.
"Provided safeguards keep pace" is doing a lot of work in that sentence. For any organization adopting AI documentation tools, that's the actual job: review the output, know the failure modes, and treat "staff find it valuable" and "it makes mistakes" as two facts to manage, not one to pick.
What this means for health organizations
Documentation tools can give real time back to staff, and they need a review step that works for busy people. Train staff on the specific kinds of errors the tool makes, such as missing or invented details, and keep a simple way to log mistakes and report them to the vendor.
