Cigna PxDx
The system that reviewed 50 claims a second and read none of them
What They Said
Cigna positioned PxDx, short for “procedure-to-diagnosis,” as a routine medical-review tool. Internally, executives described it as a way to handle high-volume, low-complexity claims efficiently — the kinds of cases where the procedure code and the diagnosis code obviously did not match. Cigna told regulators and investors that licensed physicians remained in the decision loop on every denial.
The company reported tens of millions of dollars in annual claims-review savings from automation programs across its medical-management portfolio. PxDx was one of the workhorses. Cigna repeatedly defended the system in public statements as a “simple matching tool” that did not constitute the practice of medicine.
What Actually Happened
In March 2023, ProPublica and The Capitol Forum published an investigation, “How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them,” showing that PxDx allowed Cigna medical directors to deny claims in batches at a rate of roughly 1.2 seconds per case. Internal data reviewed by reporters showed Cigna doctors signing off on more than 300,000 denials over a two-month period in 2022 — an average of about 50 per second across the program.
The medical directors were not actually reviewing the underlying records. The PxDx system pre-flagged claims as mismatches, surfaced a pre-written denial reason, and routed them in batches for a single click-through signature. Doctors interviewed by ProPublica said they were given monthly handle-time targets that made individual review mathematically impossible.
A federal class action was filed in California in July 2023 alleging the system violated the state’s requirement that claim denials involve a physician’s professional review of the actual case. California’s insurance regulator opened an investigation that same year. Cigna disputed ProPublica’s framing but did not contest the throughput numbers.
The Root Cause
Cigna optimized for the legal artifact of physician review, not the act of it. The system was built to produce a defensible paper trail — a doctor’s name, a denial code, a timestamp — without producing the underlying clinical judgment that the paper trail was supposed to represent. PxDx did not fail technically. It succeeded at exactly what it was designed to do, which is what made it a scandal.
The second failure was that no internal control measured the gap between the audit artifact and reality. There was no metric tracking the average time a medical director spent on each denial, no flag when handle times dropped below physically plausible thresholds, no review of whether the diagnosis-procedure match logic was producing clinically reasonable outputs. The KPIs measured throughput and savings. They did not measure whether the work was real.
The Pattern to Watch For
If your automation generates a compliance artifact (a signature, an approval, a “human-in-the-loop” stamp) faster than a human could plausibly do the work, the artifact is fiction. Regulators, plaintiffs’ lawyers, and journalists will all eventually compute the per-decision time. So should you. Any process where the human-review timestamp is generated by the system rather than the human is a defective control.
What You Should Steal
Instrument the human side of the loop. For every decision your AI surfaces for human review, log the active time the reviewer spent on the case, not just the timestamp of the click. If the median active-review time is under what the task realistically requires, redesign the workflow before someone redesigns it for you. Cigna’s exposure is not that PxDx existed. It is that no one inside the company asked how long 50-claims-per-second meant per claim.