AI Does Not Practice Medicine": Darren Bullock on Building Trust Into Radiology's AI Boom

Darren Bullock, COO and CIO of Envision Sally Jobe Radiology Imaging Associates, draws a hard line under every AI tool his radiology practice deploys: no matter how much of the diagnostic workflow AI touches, the radiologist still signs the report and owns what happens next. That line, he told Sarah Richardson on The 229 Podcast, is what lets the practice move fast on AI without losing the trust of the physicians using it.

The Job Nobody Designed

Bullock didn't set out to run both operations and IT. He was already COO when the practice's dedicated CIO left under "less than ideal circumstances," and Bullock stepped in to stabilize a team he describes as genuinely talented while the organization figured out what came next. The interim label never came off.

At roughly 500 people, Bullock said, Envision Sally Jobe simply isn't large enough to justify splitting the two chairs. Holding both lets him evaluate a technology investment through both an operational and a technology lens at the same time, without a handoff between two executives who might weigh the tradeoffs differently. Asked how often he drives home arguing with himself between the two domains, he told Sarah Richardson on The 229 Podcast that it's "fortunately never." The reason, he said, comes back to that old "three legs of the stool for project management," referring to people, process, and technology, which in his view should always stay aligned whether one person holds both jobs or two.

More AI Than Any Other Specialty

At RSNA in December, he said, there were roughly 1,200 FDA-approved AI algorithms across all of healthcare, and about 800 of those were in radiology alone.

Some of that is already routine. Bullock described physicians using large language models to take a dictated report and generate a summarizing impression, with the tool learning and adapting to each radiologist's individual dictation patterns over time. It's a small workflow assist on its face, but it's also a preview of how much of the specialty's day-to-day is being touched by AI, whether or not leadership has fully caught up to it.

A Risk Score Built for Everyone

The sharpest example Bullock brought to the conversation was a breast cancer risk-scoring tool the practice deployed last month, and the bias problem it was built to fix. The existing standard, the Tyrer-Cuzick score, is a questionnaire-based lifetime risk assessment trained on a largely homogenous dataset of European white women. It works, but it wasn't built with everyone in mind.

The new tool takes a different approach entirely: it analyzes the DICOM image itself rather than a patient questionnaire, generating a nearer-term five-year risk score trained on a diverse population that includes Black and Asian women. That shift lets patients act on real, near-term information and lets the practice track whether a woman's risk score moves year over year as her lifestyle changes.

Envision Sally Jobe was the first outpatient imaging center to deploy it, under a brand-new de novo FDA clearance with no prior device to model against. Bullock said nobody had implemented this specific technology before, "not only in our community, but internationally," which meant the practice and its vendor were inventing the workflow together from scratch. They had to work out how images would route to the right system, how results would come back, and how those results would populate radiologist and technologist queues, none of which existed on paper before they built it.

Where the Line Sits

Bullock is unambiguous about where that accountability boundary sits. "AI does not practice medicine," he said. The technology can generate a recommendation, draft a report, or flag an area that helps a radiologist focus faster, but the radiologist signs that report and carries the responsibility for it.

The practice's board plays a bigger role in that equation than a typical governance body. Bullock described a highly engaged board that evaluates and validates every AI tool before it reaches a radiologist's desk and builds the change-management structure that lets peers adopt it with confidence. The biggest win, in his telling, has been simply having candid conversations with that board early, before a purchase decision gets made, so the organization avoids "chasing shiny objects." Every vendor claims its algorithm does everything. In practice, the real capabilities are often narrower, and the tools can carry bias, error rates, or outright glitches that don't show up in a sales pitch.

That candor pays off downstream, too. When board members talk to peers at other organizations, Bullock said, it's clear the decision was clinician-led rather than technology-led, which is exactly the reputation he wants the practice to have.

There Is No One Metric

Bullock pushed back on the instinct to judge every AI tool by the same yardstick. The breast cancer risk-scoring tool isn't about speed. It actually slows the workflow slightly, since it's one more data point for radiologists and technologists to review, but its value is entirely in what patients can do with the information: catch cancer earlier, adjust their screening plan, act.

A second tool the practice deployed, MRI acceleration, is judged by an entirely different measure. The algorithm predicts and subtracts the distortion that comes from speeding up the magnet, producing an image as clear or clearer while cutting scan time, which helps patients who are claustrophobic or can't hold still and helps a practice running short on machines relative to demand. But speeding up the scanner just shifts the bottleneck to the technologist, which meant adding a lower-licensed staff member to help with patient prep. Applying one efficiency metric across every AI use case, Bullock said, is where a lot of organizations get it wrong.

Outrunning the Shortage

He's candid that not every call has gone his way in hindsight. The practice held onto an aging radiology information system longer than he now thinks it should have, wary of the disruption a replacement would cause for staff and patients. Looking back, he believes he should have pushed to replace it sooner, since the newer system unlocked workflows the old one simply couldn't support. Moving faster is one of the things he'd do differently across the board, he said, "and I think with AI, that's one of the key takeaways is it goes so fast that even if we think we're moving fast, it's really hard to keep up with it."

That lesson matters more now because the runway is shorter than it used to be. A shortage of radiologists and a parallel shortage of technologists are colliding with rising demand for imaging exams, and neither shortage is close to resolving. "That trust piece is gonna be really critical to be able to be one of those practices that thrives versus one that's being left behind," he said, describing the accuracy and trustworthiness of a practice's AI tools as the real dividing line ahead, not just how many of them it has deployed.

Darren Bullock is COO and CIO of Envision Sally Jobe Radiology Imaging Associates, a roughly 500-person, multi-state radiology practice serving hospitals, health systems, and outpatient imaging sites.