One of the first conversations we have with radiology department heads evaluating any workflow tool — including ours — goes something like this: they're interested in reducing read time per study, and they're also acutely aware that shorter read time per study is exactly what gets radiologists penalized under most current productivity frameworks. The metric that matters to administrators is RVUs per shift. The metric that should matter to patients is diagnostic accuracy. These two are not the same, and the tension between them is real.
This isn't a new problem. Radiology departments have struggled with productivity measurement for decades. But the arrival of AI annotation tools makes the question more urgent, because a tool that genuinely reduces search time per study will produce a measurable change in per-study duration. Whether that change registers as a productivity gain or as an incentive to spend less time per study in ways that compromise thoroughness depends entirely on how the department has defined and measured productivity.
What RVUs Actually Measure
Relative Value Units are a billing construct, not a performance metric. They were designed to standardize physician compensation across procedure types based on estimated work time, practice expense, and malpractice cost. A chest X-ray is assigned roughly 0.18 work RVUs. A CT chest with and without contrast is around 1.3. An MRI brain with and without is somewhere around 2.0. These numbers reflect Medicare's estimate of the physician work involved in each procedure — they weren't designed to measure how carefully the radiologist read the study or whether the report was accurate.
The problem with using wRVUs as a primary productivity measure is that it encourages volume optimization at the expense of case complexity handling. A radiologist who takes 8 minutes on a complex chest CT to reconcile ambiguous ground-glass findings and carefully document Lung-RADS scoring may generate the same wRVUs as a radiologist who spends 3 minutes on the same study and signs a cursory report. The first radiologist is providing better care. The productivity metrics say they're equally productive, or possibly less efficient if the first radiologist's daily volume is lower as a result.
When we look at how workflow tools affect this, the question becomes: if an AI pre-annotation reduces the search time on that 8-minute complex chest CT to 5 minutes, does the department interpret that as the radiologist becoming more efficient, or does it create pressure to read more studies per shift and eventually cause per-study time to compress further? The answer depends on how leadership has framed the purpose of the tool.
The Case Against Velocity-Only Metrics
Some departments have begun incorporating quality metrics alongside volume metrics — miss rate tracking, discordance audits, peer review scores, and addendum frequency. These are imperfect proxies for diagnostic accuracy, but they're better than nothing. The problem is that quality metrics in radiology are expensive to generate at scale. A meaningful peer review program requires a second radiologist to actually read a sample of cases, which consumes time that doesn't generate RVUs.
The ACR has published guidance on peer learning programs — the shift from "peer review" as a punitive error-detection mechanism toward "peer learning" as a quality improvement framework. This is a more defensible approach, but it still doesn't solve the fundamental problem that the feedback loop between reading quality and measurable outcomes (missed cancers, delayed diagnoses) operates on a timeline of months to years, not shifts.
We're not saying RVU-based measurement should be abandoned. In the short term, it's the only measurement system most departments actually have infrastructure to run. What we are saying is that deploying a workflow tool while retaining an exclusively volume-based productivity framework will produce a perverse outcome: the tool reduces mechanical search time, the department interprets faster reads as additional throughput capacity, and the efficiency gain is absorbed by increased volume rather than improved thoroughness on complex cases.
What a Better Framework Might Look Like
The departments we've talked to that have thought most carefully about this tend to organize around three separate measurement dimensions, tracked independently rather than collapsed into a single productivity score:
Volume throughput. Studies completed per shift, adjusted for case mix complexity. Case mix adjustment is imperfect using wRVUs, but it's better than raw study count. Some departments use a more granular complexity score based on study type, number of series, and whether prior comparisons exist.
Report completeness. Whether reports include required structured elements for relevant findings — Lung-RADS scores for pulmonary nodules above a size threshold, Liver Imaging Reporting and Data System (LI-RADS) scores for hepatic observations, appropriate follow-up recommendations per ACR guidelines. This can be tracked with structured reporting tools and doesn't require a second reader.
Discordance and addendum rate. Tracking how often a signed report is subsequently addended, and whether the addendum represents a clinically meaningful change or a minor clarification. High addendum rates can indicate over-reporting of uncertain findings; very low rates on a high-volume service can indicate under-reporting. Neither extreme is clearly good.
None of these measures throughness directly. There is no direct measure of thoroughness in radiology that is both practical and valid at scale. The best a department can do is create incentive structures that don't actively penalize it.
What AI Tools Change About This Equation
The honest answer is that an AI pre-annotation tool changes the input to the radiologist's cognitive process, not the output measurement system. If the tool reliably surfaces 2mm pulmonary nodules that the radiologist would have found after 40 seconds of search — and the radiologist can confirm the annotation in 5 seconds instead — the read time decreases. What that freed time gets allocated to depends on the department's workflow and culture.
In departments where the senior radiologist has explicitly set the expectation that faster annotation-assisted reads should be used for more careful review of incidental findings and more detailed Lung-RADS documentation, the quality gains materialize. In departments where the scheduling model assumes a fixed number of reads per shift and faster reads simply mean finishing earlier, the quality gains don't show up in outcomes — they show up in radiologists leaving on time, which is not nothing, but it's also not the clinical value proposition.
One practical frame we've found useful: before deploying any workflow tool, the department head and lead radiologists should explicitly decide what the freed time should be used for. Write it down. Make it part of the pilot agreement. "If AI pre-annotation reduces average read time per chest CT by 2 minutes, those 2 minutes will be allocated to: [incidental finding documentation / peer consultation on ambiguous cases / worklist reorganization for urgent cases / etc.]." If you don't decide in advance, the system will decide for you, and the system usually decides on throughput.
Radiologist Buy-In and the Measurement Trust Problem
There is a second-order problem here that productivity discussions in radiology often skip: radiologists don't trust productivity measurements, and for good reason. Most have seen metrics used to justify staffing reductions rather than workload management. Most have experienced the absurdity of being penalized on a per-study metric for spending extra time on a genuinely complex case.
When a department introduces a new workflow tool alongside unchanged productivity metrics, the implicit message to radiologists is: this tool should make you faster, and faster means more studies per shift. Even if that's not what leadership intends, it's a reasonable inference from the incentive structure.
The departments that deploy workflow tools most effectively tend to involve radiologists in defining what success looks like before the tool goes live — not as a checkbox consultation, but as a genuine negotiation about what the productivity framework should measure and how any efficiency gains will be allocated. That conversation is harder than buying a tool. It's also the reason the tool actually works.