The radiologist's day in an organised programme is a queue of batches, and most of what is in them is normal. Vara puts that share at around 97% [5], and European screening reads every mammogram twice [6], so 1,000 screens carry roughly 970 normal examinations and as many as 1,940 reads spent confirming nothing is there [1]. Autonomous triage takes both reads off the subset the AI is prepared to call clearly normal. How large that subset is, the announcement does not say [22], so nobody outside Vara can size the staffing case yet.
What a programme buys here looks more like a governance instrument than a classifier. Vara credits the certification less to the model than to ATMON, its Autonomous-Triage Monitoring layer [8], which sets each site's operating point, tracks mammography hardware changes, system health and daily performance signals, watches cancer detection and recall rates, and reverts the site to full radiologist reading when those numbers leave defined limits [9]. Chief technology officer Stefan Bunk rests the case on "seven years of continuous monitoring of how our AI performs in the real world, on every single case" [10]. Vara is opening ATMON to every customer, including those whose national guidelines bar autonomous triage and those with no intention of adopting it [11]. That makes the monitoring record the thing a site can start collecting on Monday, with the autonomous mode as what a long enough record eventually buys. Vara also presents ATMON as the human oversight the EU AI Act demands of high-risk systems, and its two documents put that in different tenses [20].
PRAIM, published in Nature Medicine, screened 463,094 women at 12 German sites between July 2021 and February 2023 [12], a window of about 20 months [5], with 119 radiologists reading 260,739 of those cases with AI support [13], which is 56% of the cohort [2]. Detection came out at 6.7 cancers per 1,000 with AI against 5.7 in the control group, a relative increase of 17.6% [14], or one extra cancer per 1,000 women screened in absolute terms [3]. The paper records that participating radiologists chose for themselves whether to use the AI, which leaves the two groups short of randomised [16]. Vara's release describes the study as nationwide and puts the cohort at 461,818, a figure the paper does not carry [17] and 1,276 women below the published number [4].
The two-by-two worth drawing here separates what a programme would actually run from what it actually holds. One axis is the mode: AI-assisted double reading, or autonomous closure. The other is the evidence: a vendor trial of assisted reading, or a site's own monitored detection and recall series at its own operating point. Three of those cells describe practice that already exists somewhere in Europe. The fourth, autonomous closure defended by a trial of a different mode [15], is the cell where a missed cancer arrives and no reader's name sits on the case file. The evidentiary path that follows is to run ATMON in assisted mode and treat the reversion log as the asset acquired, moving to closure only once the local series is long enough to defend in front of a review panel. The price of that patience is paying for monitoring without collecting the read savings, possibly for a year or more, while the shrinking radiologist workforce and widening screening ages [7] press in the other direction.
