How to actually deploy AI in healthcare

The admin burden

• Visit notes, which need a specific structure.
• Billing codes, with intricate rules. Ontario's rulebook alone is about 1,000 pages.
• The inbox: lab results, prescription renewals, reports and duplicates, hundreds a day.
• Referrals: finding a specialist who will take the case, then handling their reply.

Doctors pay for it with evenings spent charting ("pyjama time"), and patients get less of their attention.

Why it's solvable now

This isn't AI diagnosing rare cancers. The work is manual and repetitive, and only sometimes needs a little judgement. Just inside what AI can already do, these are engineering problems: we know how to build reliable systems out of unreliable parts.

Why it's hard

• Yak shaving: delivering the vision means a long chain of infrastructure work first.
• Integrations: the EMR is the legal system of record, with different rules in every region. A tool that takes three clicks and 20 seconds to log into hasn't solved anything.
• Real-world audio: telling speakers apart in a noisy exam room is much harder than in a video call (speaker diarization).
• People: healthcare is a sociotechnical system. The term comes from a 1951 study of coal mining, where new machinery reorganised the work and made miners miserable. People also have a limited capacity for change.

Whose behaviour has to change?

The question his team asks before building any feature. Software engineers were among the first whose own work AI agents changed, so they can design with lessons they've lived through. And reasoning about systems that aren't perfectly predictable isn't new: manufacturing and aviation have done it for decades. Rolling AI into healthcare is one big migration: hard, but solvable.