Last year, The New York Times published a story with a provocative title: “Your A.I. Radiologist Will Not Be With You Soon”.
The story looked back at one of the more famous predictions from the early days of artificial intelligence.
In 2016, AI researcher Geoffrey Hinton argued that people should stop training radiologists. Advances in image recognition were moving so quickly, he believed, that AI would soon outperform people at reading medical images. Radiology seemed like exactly the kind of profession that was vulnerable to AI automation: highly trained people spending much of their day looking at images, identifying patterns and finding abnormalities.
But nearly ten years later, the prediction hasn't played out the way many people expected. Radiologists remain in high demand, and at Mayo Clinic (one of the more enthusiastic adopters of AI in medicine) the radiology department has grown by roughly 55% since 2016 to more than 400 radiologists. At the same time, Mayo has more than 250 AI algorithms in various stages of development and use across its health system.
That doesn't mean AI caused Mayo to hire more radiologists. It does mean that heavy adoption of AI and a growing population of human experts have been able to productively co-exist.
And we think there is an important lesson there for permitting and plan review teams.
Part of the problem with the original prediction was that it treated radiology as if the job were simply “look at an image and find the problem.” AI has become remarkably good at pieces of that work.
At Mayo, AI can help identify abnormalities, prioritize images for review, and automate measurements that once took physicians considerable time to complete. One tool, for example, measures changes in kidney volume, a process that was previously performed manually and was both time-consuming and difficult to do consistently.
But identifying something in an image is only part of what a good radiologist does. Radiologists interpret findings in the context of a patient's history, decide which of those findings are important, communicate with other physicians, and, when the answer isn’t obvious or straightforward, apply years of experience and pattern recognition to make a determination about what to do next.
That seems to be where organizations like Mayo have landed. AI can take meaningful work off a radiologist’s plate and make parts of the job faster and more consistent, but the human expert still plays a central role in making sense of what the technology finds and deciding how to act on it.
In other words, the technology turned out to be very useful at portions of the job without being capable of replacing the whole job.
The larger workforce data points in the same direction. A 2025 study published in the Journal of the American College of Radiology projects that the U.S. radiologist workforce will continue to grow through 2055. Depending on assumptions about future residency positions, the researchers project a workforce that is between 25.7% and 40.3% larger than it was in 2023.
None of this suggests that AI has been unimportant to the field of radiology. Quite the opposite. AI is becoming deeply embedded in how radiologists do their work. The more interesting lesson is that AI has largely been used to augment expertise rather than eliminate the need for it.
That is very similar to an argument we've made before about AI in plan review: the question isn't whether AI can replace the reviewer — spoiler: we don’t think it can or will.
It’s whether AI can remove enough friction from the work to make an experienced reviewer substantially more effective.
Reviewing building plans isn't the same as reading medical images, of course.
But the work has some important similarities.
A plan reviewer works through a large amount of complicated information. Some questions have clear answers: Is a required document present? Is something the right distance from something else? Has a file been prepared correctly?
Other questions require context, interpretation and judgment. Codes vary by jurisdiction, unusual projects create unusual circumstances, and experienced reviewers may reasonably disagree about how a particular provision applies.
At the same time, many departments are struggling to maintain enough experienced people to do the work. The U.S. Bureau of Labor Statistics counts about 147,600 construction and building inspectors today and expects employment in the occupation to decline slightly through 2034. Even so, it projects about 14,800 openings every year, essentially all because existing workers will retire, leave the workforce or move to other occupations. See the Bureau of Labor Statistics data.
That changes the way we think agencies should consider the AI opportunity. In most of our conversations with local governments, the problem isn't an excess of experienced reviewers whose jobs need to be automated away. The problem is that experienced reviewers are difficult to find, difficult to replace and being asked to keep up with a growing volume of complicated work.
So, again: The right question to ask for permitting and plan review isn’t, “What can AI automate?”
It’s, “What can AI do that gives our reviewers more leverage?”
There are absolutely places where automation makes sense in permitting and plan review. If a submitted PDF needs to be flattened, for example, there is little value in having an employee discover the problem, contact the applicant and either explain how to fix it or fix it themselves. That work is repetitive, administrative and has a clear answer. Software should simply handle it.
The same principle applies to other predictable problems at the front of the process. AI can identify missing information, unreadable files and other obvious issues before they consume reviewer time. But, as we've written before, a complete submission isn't necessarily a reviewable submission. Getting the paperwork into shape and making a substantive judgment about a plan are different jobs, and they shouldn't necessarily be handed to AI in the same way.
Judgment-heavy work is where augmentation becomes more important. AI can still remove a tremendous amount of effort without taking the decision away from the reviewer. It can find information buried across a plan set, compare a resubmittal with the previous version, surface a potentially relevant code section, identify something worth a second look or draft a clearer correction comment for the reviewer to consider.
Those applications don't make the reviewer less important. They make the reviewer's time more valuable because less of it is spent searching, measuring, comparing and performing administrative work.
That distinction (automation versus augmentation) is starting to become the central framework we bring up with every agency we talk with about AI in this space. Automation makes sense when the task is predictable enough that we would be comfortable letting the technology complete it without human judgment. Augmentation makes sense when the technology can contribute information, analysis or speed but an experienced person should still make the final decision.
This matters because the biggest promise of AI in plan review may not be reducing the number of people doing the work. It may be increasing what the people already doing it are capable of accomplishing.
Radiology offers a useful glimpse of what that can look like after nearly a decade. AI became far more capable. It became deeply embedded in the workflow. It took on work that previously required human effort. And the experts didn't disappear.
Instead, the profession has been learning how to divide the work differently: let technology do what technology does particularly well, and give the human more support for the work where expertise and judgment matter.
We think plan review is headed in a similar direction. Our broader view of where AI belongs in plan review starts with the same basic idea: automate the simple stuff, assist where judgment matters, and use the data created by that work to learn what might safely be automated next.
That last part matters. AI will get better. Agencies will learn more about where their reviewers consistently agree with it and where professional judgment continues to matter. Some work that starts in the hands of a reviewer may eventually prove predictable enough to move earlier in the process.
But we don't need to assume the end state before we have the evidence.
Radiology gives us a useful example from very recent history of another path: Use AI thoughtfully. Let it take on work where it creates real leverage. Keep learning what the technology does well. And design the process around making the expert better rather than trying to make the expert disappear.
This is the approach we’re taking as we build AI into the plan review workflows we've been working on for two decades. If you’d like to see what we’re building or how it could look for your jurisdiction, get in touch today.