Every cytology laboratory in the United States runs on a number set by federal rule. Under 42 CFR 493.1274, part of the regulations implementing the Clinical Laboratory Improvement Amendments, a person who screens slides may examine no more than 100 slides in any 24-hour period, spread over no less than an eight-hour workday. The rule was written for slides read on a microscope; how digital, AI-assisted reviews count against it is set out device by device. The rule calls this an absolute maximum and forbids its use as a performance target. The technical supervisor establishes an individual maximum for each screener, based in part on a review of 10 percent of the cases that screener called negative, and the limit is reassessed at least every six months. Read plainly, the rule treats the human reader's pace as a safety variable.
On 5 November 2026 the Food and Drug Administration published a final order that carries the same concern into machine-assisted screening. The order, 91 FR 69687, codifies a new device type at 21 CFR 864.3900: the "digital cervical cytology slide imaging system with artificial intelligence algorithm." This post reads that four-page document closely, because its special controls set out what FDA will expect from any AI cytology system that follows the first one.
Practical takeaway. Since January 2024 a second AI Pap-test imaging system has been able to seek 510(k) clearance with Hologic's device as predicate, and the codified special controls now state its obligations in the Code of Federal Regulations. It will have to show reader performance with and without the algorithm against a comparator such as the optical microscope, results broken down by patient demographics, sites and scanners, a security architecture for its network and cloud services, and clinical data that support a recommended daily screening limit for the people who use it.
What Hologic asked FDA for in August 2021, and what FDA granted on 31 January 2024
Any device that was not on the US market before 28 May 1976 falls into class III by operation of law, which ordinarily means premarket approval. The De Novo route under section 513(f)(2) of the Federal Food, Drug, and Cosmetic Act lets a sponsor with a novel, moderate-risk device ask FDA for a lower class instead. According to the order, FDA received Hologic Inc.'s request for the Genius Digital Diagnostics System with the Genius Cervical AI algorithm on 27 August 2021. FDA issued its classification order to Hologic on 31 January 2024. The statute directs FDA to classify by written order within 120 days of a De Novo request; the order does not explain the 29 months between request and grant.
The Federal Register document of 5 November 2026 writes that classification into the Code of Federal Regulations, effective that day, and states that the classification has applied since 31 January 2024. Predicate status and the 510(k) route date from the January 2024 grant; the codification makes the device type and its special controls part of published federal regulation. The same issue carries a parallel order for prognostic assessment software for mild cognitive impairment, codified in 21 CFR Part 882 and applicable since 11 January 2024, so the gap between a De Novo grant and its codification is a feature of both documents.
How 21 CFR 864.3900 defines an AI cervical cytology system
The identification paragraph is precise about what the software does. The device is a prescription in vitro diagnostic system that creates, displays and manages digital images of slides prepared from Pap test specimens and conventional Pap smears. Its machine-learning algorithm selects and presents "areas of interest to facilitate interpretation by the reader," and its output is "intended to assist the user in rendering a cytology diagnosis." The diagnosis stays with the person at the review station. The labeling must say that the user should review the entire gallery of objects of interest before rendering a diagnosis, and should use the device alongside the standard of care, the patient's history and professional screening guidelines.
FDA's risk table is short. It lists two risks: false negatives and false positives caused by the device presenting abnormal cells badly in its image gallery, and the same errors caused by the user misreading the images or operating the device incorrectly. The second risk places the human reader inside the device's risk analysis. That is why so much of the order concerns training, protocols and workload.
Which data, security and validation duties the special controls impose
The design verification and validation controls run from optics to datasets. Each hardware component must be assessed, from the slide feeder and light source to the Z-focal plane acquisition that captures the full depth of cell clusters, and the system must be assessed for color reproducibility, focusing accuracy, stitching error and turnaround time. The software description must cover the training dataset, which has to represent preanalytical variables such as fixation, slide processors, difficult cases, multiple sites and patient demographics, and an independent validation dataset with clinically relevant subsets.
Two controls reach beyond the laboratory bench. The labeling must summarize performance stratified by patient demographics, medical history, user experience, sites and scanning equipment, and must describe where the device may fail, including limits that come from the data used to train, tune and test the algorithm. The software documentation must include a system architecture diagram with a matrix of communication endpoints, protocols and security protections, covering cloud services and any product in the communication pathway. The labeling must also address cybersecurity and remote review of images. In practice, a cervical screening program built on such a device inherits a network diagram as part of its regulatory file.
The clinical validation control sets the comparative standard. Screening and diagnosis with the algorithm's assistance must be shown to be non-inferior to a comparator method FDA accepts, such as the optical microscope, in a study that compares intended users reading unassisted and with the device. Accuracy is measured by sensitivity, specificity, predictive value and likelihood ratios, with enough slides in each diagnostic category for confidence intervals in each subset. The readers in that study must match the intended users in training, certification and experience.
Why FDA asks for a recommended daily maximum screening workload
Paragraph (b)(2)(v)(B) of the new section is the clause with the clearest link to laboratory practice. The clinical study must collect "intended user screening time data" to "characterize and set recommended daily maximum workload screening limits for intended users," and the labeling must carry that information. The order itself does not mention the CLIA rule. FDA's decision summary for De Novo DEN210035 already makes the connection for Hologic's device. On pages 33 and 34, it counts every case diagnosed with the Genius Cervical AI algorithm as half a CLIA slide equivalent, a full manual review of the glass slide as one, and both together as one and a half, all within the cap of 100 equivalents in a workday of at least eight hours. Its first example is 200 AI-assisted case reviews equal to 100 equivalents. The technical supervisor still sets and reassesses each cytologist's individual limit. The summary also reports that, across the 12 cytotechnologists in the clinical study, review time showed minimal correlation with sensitivity and specificity.
An algorithm that preselects the most suspicious cells can make each case faster to review, and a laboratory under staffing pressure will be tempted to convert that speed into volume. For Hologic's device that conversion is now explicit: one cytologist may read up to twice as many AI-assisted cases as manual slides. For later makers, the special controls require clinical screening-time data that characterize and set recommended daily maximum workload limits for intended users, and labeling that gives intended user screening time data about daily slide screening workload limits. They do not prescribe a CLIA conversion ratio. The order keeps diagnostic interpretation with the human reader and says nothing about the allocation of legal liability. On that question the UK offers a separate comparison: the UK Jurisdiction Taskforce's July 2026 statement on liability for diagnostic AI in NHS trusts places a non-delegable duty of care on the clinical institution.
How Quentir Reads It
Quentir reads 21 CFR 864.3900 as one of the more complete US statements of what an AI diagnostic device owes to the people around it. The special controls cover the patient population through stratified results, the network through an architecture diagram, and the reader's working day through a measured workload limit. Each of those duties is concrete enough to audit, which makes the order a useful reference for regulators outside the United States who are writing conformity requirements for AI in medicine. The weakest point is timing. A request filed in August 2021 was granted in January 2024, 29 months later, and the codification followed another 33 months after the grant. The order gives no reason for either interval.
The next test is any 510(k) that cites Hologic's device as predicate. If FDA publishes a decision summary for it, that document would show how the second maker's recommended workload limits were set and how FDA applied the stratification controls. Quentir's coverage of AI liability, medical devices and diagnostics, including the posts linked here, is available in full through the All-access membership, one subscription for the archive that this analysis draws on.
Sources: Food and Drug Administration, Medical Devices; Hematology and Pathology Devices; Classification of the Digital Cervical Cytology Slide Imaging System With Artificial Intelligence Algorithm, final amendment and final order, 91 FR 69687, FR Doc. 2026-22593, Docket FDA-2026-N-12384, published 5 November 2026 (PDF), adding 21 CFR 864.3900; Food and Drug Administration, Classification of the Prognostic Assessment Software of Mild Cognitive Impairment, final order, Docket FDA-2026-N-12212, 21 CFR Part 882, same issue; Food and Drug Administration, Evaluation of Automatic Class III Designation for the Genius Digital Diagnostics System with Genius Cervical AI algorithm, decision summary DEN210035, pages 32 to 34 (cytologist workload determination); 21 U.S.C. 360c (Federal Food, Drug, and Cosmetic Act section 513); 42 CFR 493.1274, cytology workload limits under the Clinical Laboratory Improvement Amendments regulations. Links checked 5 November 2026.
Published intelligence, built to inform your own decisions. Published: November 5, 2026.