A Cigna medical director sits in front of a queue. An algorithm has already flagged a set of claims where the billed procedure doesn't line up with the diagnosis code against a list of preset pairings. The director doesn't open the patient files, doesn't read the treatment notes, doesn't check the clinical history. The director signs off on the batch. The denials post, the notices go out, and the patients learn weeks later that the procedures their own doctors ordered won't be paid for. Across one two-month stretch in 2022, that sequence produced more than 300,000 denials, and the internal records put the average time a Cigna doctor spent on each one at 1.2 seconds.
That figure comes from a March 2023 investigation by ProPublica and The Capitol Forum, built on internal Cigna corporate documents and interviews with former company physicians. The system is called PxDx, short for procedure-to-diagnosis. It compares the procedure on a claim against the diagnosis codes Cigna treats as acceptable for that procedure, and when the two don't match the preset criteria, it routes the claim for denial. The medical directors then sign the denials in groups. One director's scorecard total was more than 121,000 denials over the first two months of 2022. The spreadsheet listed each physician next to a column headed "PxDx" and treated those numbers as a measure of throughput. The reporting treated them as a measure of something else.
Most of the writing about rubber-stamping starts from the assumption that the reviewer's time on a decision, the dwell time, is unrecoverable after the fact. The audit log shows that a user signed off at a given moment, and says nothing about whether the person spent twelve minutes reading a file or four seconds clearing a queue. Cigna is the rare exception. The time got recorded. It was right there in a spreadsheet as a per-physician denial count, and plain arithmetic turns that straight into seconds per claim. The number didn't protect the company; it became the first paragraph of the complaint. The usual failure in these cases is the absence of a paper trail, where nothing in the record shows what the reviewer did. Cigna's exposure runs the other way. The record existed, it was specific, and at 1.2 seconds per determination it described something that cannot be individualized medical review under any account of what that review involves.
The detail that makes the case broader than healthcare is that PxDx is not, by Cigna's own account, artificial intelligence. The company describes it as a process for accelerating payment of routine, low-cost screenings and disputes the characterization that its doctors reject claims without examining them. The dispute is beside the point. PxDx is, at its core, a deterministic matching routine: a lookup that checks two codes against a table and flags the mismatches. It is among the least sophisticated forms of automation a large company runs, and it produced one of the clearest records of substituted judgment in the entire field. That's the part worth sitting with. The liability doesn't turn on whether the automation was a neural network or a spreadsheet formula. Automated decision-making technology is defined by what it does, by whether computation replaces or substantially replaces a human decision-maker, not by how clever the computation is. The human involvement standard asks the same question of a rules engine that it asks of a model, and a routine that disposes of 300,000 medical-necessity determinations answers it the same way a model would.
The 1.2-second figure is not a sign of careless individual reviewers. It's an arithmetic property of the workflow. When an interface presents a physician with a batch of flagged denials and a single action disposes of all of them, the time spent per claim is just the length of the session divided by the number of claims in it. A director who signs sixty thousand denials in a month is not reviewing sixty thousand files quickly. The review of the individual claim isn't happening; what's happening is the disposal of a batch the algorithm assembled. This is the same mechanism that lets a recruiter dispose of a full queue of rejections in one sitting, and that let Workday's screening tools process more than a billion rejected applications. The difference is that Cigna counted the batches, so the per-claim math is on paper.
The batch signature is the mirror image of write isolation. There, the machine is kept from writing the decision into the system of record while a human ratifies it. Here, a human writes the decision into the record, by signing, while the machine made it. Both arrangements end with a person's name attached to an outcome the system produced, and neither says anything about whether the person decided it.
What turns a throughput number into legal exposure is the promise that sat above it. California's Knox-Keene Act, at Health and Safety Code Section 1367.01(e), requires that a decision to deny or modify a request for care for reasons of medical necessity be made by a licensed physician or a qualified professional competent to evaluate the specific clinical issues, not by an algorithm. Cigna's plan documents attributed medical-necessity determinations to its medical directors. The class action, Kisting-Leung v. Cigna, filed in the Eastern District of California in July 2023, alleges that the directors' role in the PxDx denials was the batch signature and nothing more. In a March 2025 ruling, Judge Dale Drozd dismissed some plaintiffs' claims for lack of standing but allowed the core theory to proceed: that if the algorithm rather than a qualified professional made the medical-necessity determination, with physicians spending negligible time per claim, that states a violation of Section 1367.01(e) and, through it, California's Unfair Competition Law. The court also held that claim survived ERISA preemption under the statute's savings clause for laws regulating insurance. Fact discovery closed in May 2026, class certification briefing runs into late 2026, and a jury trial is set for late 2027.
The economics of the arrangement rest on a number the complaint also cites: roughly 0.2% of policyholders appeal a denied claim. That figure is the quiet engine of the whole system, because a denial that's never challenged costs nothing to have gotten wrong. The batch process is profitable precisely to the extent that the people it affects don't contest it. The UnitedHealth case made the same point from the opposite direction. There, the proof of an absent reviewer came from the appeals that did happen: over 90% of the AI-driven coverage terminations were reversed once a person actually read the record. The appeal was the mechanism that surfaced the gap. Cigna's design assumes that mechanism almost never activates, and the data says it's right. But the proof of the absent review didn't have to wait for an appeal, because it was sitting in the company's own scorecard the whole time. One case shows what the appeal reveals when it finally comes. The other shows the company's own records already held the answer before anyone appealed.
Cigna's defense has a real core to it. The company argues that PxDx mostly speeds the payment of low-cost, routine claims, that a denial under the system doesn't raise an in-network patient's out-of-pocket cost, and that the process exists to free medical directors to spend their time on the complex reviews that need it. Some of that may hold up at trial. But the theory in the case isn't about the claims PxDx pays. It's about the subset it denies, where the plan and the statute both promised an individualized determination by a professional, and where the only record of that determination is a signature on a batch and a logged time too short to have read anything. The efficiency argument explains why the system was built. It doesn't explain how a 1.2-second average is consistent with the review the plan documents described.
The same question shows up wherever a reviewed determination is challenged: whether the time and the materials are consistent with the review the organization says took place. Most organizations can't answer it, because they never captured the time or the materials, and the absence cuts against them. Cigna captured the time and the absence of everything else, and that cut against it harder. There's no version of this where holding less evidence would have helped. The only thing that would have helped is a review that actually happened and a record that showed it.
Proof of Review captures what a denial count can't: whether the determination was made one claim at a time or disposed of in a batch, and whether the reviewer could have reached a different outcome. A logged 1.2 seconds is what the record looks like when the review didn't happen. The point of capturing the record is to be able to show, on the determinations that matter, that it did.