Human review of AI output before it enters the record

An AI-drafted output (a clinical note, a radiology pre-read, a variant classification) is presented to a named clinician as a draft, with the means to check it against its source, before anything is filed, signed or acted on. The pattern is the review step itself: what is shown, what can be verified, and what the clinician has to do to accept it.

Generative and analytical AI now draft things that used to be written by the person responsible for them. The review step is where responsibility is kept with that person, and where it is quietly lost if the step is a formality. A draft accepted unread because it is usually right is the failure mode regulators name: automation bias.

The FDA's clinical decision support guidance makes the ability to independently review the basis for a recommendation the line between software it regulates and software it does not, and the EU's AI Act requires that high-risk AI systems can be effectively overseen by people, with awareness of automation bias. The design of the review step is therefore a compliance question and a safety question at once.

Review has to be possible, not merely permitted. For an ambient note that means being able to trace a sentence to what was actually said: Abridge describes Linked Evidence, which highlights the transcript passage behind any part of the note and plays the original audio. For an imaging pre-read it means the AI's output is presented as assistance ahead of the human read, which is how Qure.ai describes qXR. For a genomic classification it means the pipeline's pre-classification is provisional until a person has interpreted it, as SOPHiA DDM's workflow steps describe.

Time is the other constraint. The FDA notes that where a decision is critical and time-sensitive, a clinician is unlikely to have enough time to review the basis of a suggestion, which argues for review steps designed around the minutes actually available rather than an ideal reader. And the record should show what was drafted, what was changed and who accepted it.

Review interfaces are dense, split-screen and often keyboard-driven. Visible focus (WCAG 2.2, 2.4.7) is essential when a clinician is moving between a draft and its evidence, and the link between a highlighted sentence and its source should be conveyed in the accessibility tree, not only by a shared colour (Use of Color, 1.4.1). A “confirm and sign” action commits a legal record, so it needs the checking and reversal that Error Prevention (3.3.4) asks for.

AI-flagged findings on an image need a text equivalent. And the review step should not depend on a time limit that discards the draft, or accepts it, when the limit expires (Timing Adjustable, 2.2.1).

The three products in the library that describe this pattern show three different distances between the AI output and the record. Abridge's is the shortest: the draft note is in the clinician's EHR workflow, the verification tool is built into the editor, and its product page describes actionable outputs such as orders being captured “for clinician review”. Qure.ai describes qXR as “pre-read assistance” producing pre-filled reports and AI overlays, which places the human read after the AI but leaves the shape of that review to the site that deploys it. SOPHiA DDM's workflow lists annotation and pre-classification, then prioritisation, then “exploration and interpretation”, then reporting: a sequence in which interpretation is explicitly the person's step.

In every case the company's own description keeps the clinician before the record. What varies, and what public pages do not show, is how hard the interface makes it to accept a draft without reading it.