Radiology AI Can Draft the Report. Does It Fit the Reading Room?

Featuring insights from Dr. Jeff Chang, Co-Founder and President of Rad AI, in a conversation with our CEO, Saul Marquez on the Outcomes Rocket podcast.

A radiologist reaches the impression section of a report. Instead of dictating the conclusion and follow-up language from scratch, the text is already there in a style that resembles how that radiologist writes. No separate prompt or additional action is required to make it appear.

That's how Dr. Jeff Chang, Co-Founder and President of Rad AI, describes the company's early reporting product. In his Outcomes Rocket conversation, he explains that the team designed the software to work at the exact point in the reporting process when a radiologist needed it. He says the resulting reduction in dictation was roughly one-third, with an estimated hour saved during a nine-hour shift. Those are Jeff's reported product figures, not an independent measurement of every deployment.

The design choice offers a useful test for healthcare AI buyers and developers. Does the software save time inside the clinician's normal work, or does it ask the clinician to adopt another task before the benefit appears? With imaging volumes putting pressure on radiology departments, that distinction matters as much as the sophistication of the model.

Start With the Task Radiologists Already Have

Jeff spent years working overnight radiology shifts before helping start Rad AI in 2018. That experience shaped his view of the problem. Radiologists weren't waiting for more screens to manage. They needed support with work that already consumed significant portions of their day.

The company's first use case focused on the impression portion of a radiology report, including the summary, conclusions and follow-up recommendations. Its approach was to generate text at the appropriate moment, customized to the individual radiologist's language.

Jeff calls the original experience “zero click.” The software didn't require the radiologist to stop, open a separate tool or ask for the result. He says users could become familiar with it quickly because it mirrored their existing workflow.

That doesn't mean every automated draft is clinically ready to sign. The radiologist still needs to review the report and retain responsibility for its clinical content. The important implementation lesson is narrower: the system was designed to remove repetitive work without first adding a new step.

For a radiology leader evaluating similar products, the demonstration should start in the reporting environment staff already use. Watch what the clinician has to do before the AI appears, what must be checked afterward and whether the finished report can proceed through its usual process. Time saved generating text is only useful if it isn't consumed by moving between systems or repairing an unsuitable draft.

A Better Report Isn't the End of the Patient's Journey

Rad AI later expanded beyond impression generation. Jeff describes a patient follow-up solution intended to help make sure patients return for recommended imaging, such as monitoring a possible cancer finding or a growing aneurysm. He also discusses a more comprehensive AI reporting product that can generate a report from a few dictated words or lines using the radiologist's format and style.

These products address different parts of a connected process. Reporting tools aim to reduce the burden of producing clinical documentation. Follow-up tools address what happens after a finding and recommendation have been recorded. Faster documentation alone doesn't guarantee the recommended next exam occurs.

For health systems, that separation is important. A vendor might show an impressive report-generation demo while the organization still struggles to determine who tracks an outstanding follow-up recommendation. The clinical report can be complete even when the next step in care isn't.

A useful review therefore asks two questions. Does the reporting tool reduce the work required of the radiologist? If the report identifies a follow-up need, who is responsible for making sure the recommendation reaches the next appropriate person or process? Efficiency inside the reporting task and follow-through after it are separate outcomes, and both deserve named owners.

New Capabilities Still Need to Fit Clinical Responsibility

Jeff describes radiology departments facing rising imaging volumes without a comparable increase in the number of new radiologists. In that environment, adding an AI product that feels unfamiliar or interrupts established steps is a difficult proposition, even when the technology can do something impressive.

His team has worked with radiologists using different reporting workflows, which he identifies as a source of product insight. He argues that AI has to accommodate existing features and habits before teams can benefit from new capabilities. The deeper the tool reaches into critical clinical work, the more important that fit becomes.

The interview also turns to the possibility of AI-generated preliminary reports directly from images. Jeff discusses this as a future development and acknowledges uncertainty around regulatory clearance and the longer-term role of radiologists. It shouldn't be confused with the reporting assistance his company already describes offering.

That difference should shape a buyer's requirements. A tool that drafts prose from the radiologist's inputs, a system that organizes follow-up recommendations and a system that interprets images to draft a preliminary clinical conclusion do not carry the same workflow or oversight questions. Each needs its own review of accuracy, responsibility and where the clinician remains involved.

Test the Real Shift, Not Just the Demo

Before approving a radiology AI expansion, put the proposed workflow in front of practicing radiologists and trace one representative study from interpretation to final report and any required follow-up. Count the extra actions, observe the corrections the draft needs and confirm how the tool fits the existing reporting system. Ask what happens when the workflow differs from the vendor's prepared example.

The deployment decision should rest on the work removed, the work retained for necessary clinical judgment and the team's ability to complete the task without unnecessary detours. Compare those findings with the results the vendor reports, rather than treating a single time-saving estimate as a promise for every practice.

A radiologist doesn't need another place to work. They need useful assistance where the work already happens. Jeff's account of building Rad AI is a reminder that, for clinical AI, the surrounding workflow is part of the product.

About the Conversation

Dr. Jeff Chang's conversation with Saul Marquez on Outcomes Rocket covers reporting automation, patient follow-up, radiologist workload and the importance of designing AI around established clinical workflows. Learn more at Rad AI.

‍