Predictions about artificial intelligence in medicine tend to fail in a specific direction. They are usually right about what the technology will be able to do and badly wrong about when it will actually be doing it.
Mark Bouzyk, a biotechnology executive and scientist whose career spans more than 14 years in academia and more than 23 years in global pharmaceuticals and biotechnology, has been through this cycle before with genomics.
He argues that the gap between capability and adoption is a structural feature of medicine. Understanding it is the difference between a realistic roadmap and an expensive disappointment.
Capability Is Not Adoption
A tool becomes technically capable on the day it works in controlled conditions. It becomes adopted on the day a clinician uses it.
Those two dates are rarely in the same year.
Everything in between is where adoption timelines are actually decided. Integration with existing record systems. Validation on local patient populations in addition to the ones the tool was trained on. Training for staff who are already stretched. Committee review.
Liability questions that need to be answered before anyone signs off. Billing and coverage pathways that determine whether the tool is financially survivable. Often, this work takes longer than building the model itself.
Mark Bouzyk Has Watched This Timeline Before
The genomics era offers a close analog. The cost of sequencing a human genome collapsed over roughly two decades, faster than almost anyone forecast. The scientific case for using genetic information in routine care was largely settled well before that information reached routine care.
Bouzyk was making predictions about genomics in 2006 that proved substantially correct on substance while the timeline stretched considerably further than the technology alone would have suggested.
The delay was caused by everything the science had to travel through.
Laboratories needed accreditation. Results needed interpretation frameworks that clinicians could apply without a genetics fellowship. Reimbursement had to catch up. Consent, privacy, and data governance all needed working out. Each of those was reasonable on its own, and yet, collectively, they added years.
That progression, traced in his account of the path from genomics to global impact, is the pattern Bouzyk expects to repeat with AI, with one difference.
AI tools are being deployed in clinical settings faster than genomic testing was, which compresses the technology timeline while leaving the governance timeline largely unchanged.
The gap between those two clocks is where most of the current friction sits.
Three Clocks Running at Once
Bouzyk describes adoption as governed by three clocks that run at different speeds.
The technology clock is the fastest. Capability improves continuously, and the version available this quarter is meaningfully different from the one evaluated last quarter.
This sounds positive. In fact, it can be. However, it also complicates every downstream process.
The evidence clock runs slower. Demonstrating that a tool improves outcomes rather than merely producing plausible output requires study designs, comparison groups, and follow-up periods that cannot be compressed by better engineering.
A model that performs well on retrospective data has not yet shown that it changes what happens to patients.
The trust clock is slowest of all, and it is the one most often left out of planning. Clinicians extend confidence to a tool based on their own experience with it and the experience of colleagues they respect.
That accumulates at human speed. It can also be lost far more quickly than it was gained, which is why a single high-profile error can reset a timeline that took years to build.
Why Administrative Work Moves First
The clearest early adoption is happening in administrative and documentation. Bouzyk considers that entirely predictable.
Administrative work has properties that suit early deployment. It is high volume, largely rules-based, and the cost of an error is recoverable.
A misrouted claim can be corrected. The feedback loop is fast enough that problems surface within weeks. Time saved is measurable in hours, which makes the case for continued investment straightforward to argue internally.
Diagnostic and treatment applications have the opposite profile. Errors can be irreversible, the evidence bar is higher for good reason, and the feedback loop can run to months or years.
Expecting both categories to move at the same pace mistakes a difference in consequence for a difference in enthusiasm.
This is closely related to a point Bouzyk has made repeatedly: precision health is about more than data.
The volume of information a system can process says little about whether it changes a clinical decision, and less still about whether that change was beneficial.
What a Realistic Timeline Looks Like
Bouzyk’s framing avoids both of the usual failure modes, the forecast that a technology will transform medicine within eighteen months and the assumption that healthcare will simply resist indefinitely.
A realistic sequence looks roughly like this. Administrative and documentation tools reach genuine routine use first, because the economics and the risk profile both support it.
Decision support spreads next in narrow, well-bounded applications where performance can be measured against a clear standard.
Broader clinical integration follows only after monitoring infrastructure, liability frameworks, and reimbursement pathways have matured. That maturation is measured in years rather than quarters.
Institutions that plan against that sequence tend to make better decisions than those planning against a promised timeline.
They invest in data infrastructure and governance capacity early, because those are the constraints that actually bind.
They select a small number of applications and evaluate them properly rather than piloting many and measuring none. They budget for the monitoring that comes after deployment, which is routinely underfunded because it is invisible when everything is working.
Speaking about the direction of the field in a recent interview, Bouzyk returned to a point that has shaped much of his career.
The technologies that lasted in medicine were rarely the ones that arrived fastest. They were the ones that were still trusted after the initial enthusiasm had worn off, and earning that kind of trust has never been a matter of engineering alone.
The Slow Part Is the Point
There is a tendency to treat the lag between what AI can do and what medicine does with it as pure waste.
Bouzyk’s reading is different.
A significant share of that lag is spent testing claims where the cost of being wrong falls on a patient.
Not all of it is useful. Fragmented data, duplicated regulatory committee review, and procurement processes that predate the technology add delay without adding safety, and those are worth attacking directly.
But the portion of the timeline devoted to establishing that a tool works for the population in front of you and that the people using it understand its limits is not friction to be removed.
As Dr. Mark Bouzyk has put it in his writing on precision medicine, the speed that matters is not how quickly a technology arrives but how quickly it becomes something a clinician can rely on. Those are different measurements. Only one of them shows up in a patient outcome.
