The Last Pass

Marcus opened the queue at 7:14 on a Sunday evening, the way he always did, and started on the inpatient charts that had finished their concurrent CDI reviews on Friday and Saturday. The hospital's official policy was that inpatient coding could happen during normal business hours from a coding office on the fourth floor of the administration building. Marcus's unofficial policy, for nineteen years now, had been that he did the deep charts on Sunday evenings from his kitchen table, after his wife had gone to bed, with a mug of weak coffee and the small green-shaded library lamp his daughter had given him for his fiftieth birthday.

The queue tonight had thirty-six charts in it. He could see them in the sidebar of the coding application: patient initial, admission date, length-of-stay flag, primary service, the small red dot that indicated the AI had flagged its own confidence as below the autonomous-submission threshold and routed the chart to a human coder for full review. Six of the thirty-six were AI-routed. The other thirty had been autonomously coded and submitted on Friday, with the daily ten-percent random sample he was supposed to audit for accuracy.

Tomorrow morning the hospital was switching to the new policy. The autonomous system's confidence threshold was being raised from 90% to 93%, the human-routing rate was projected to drop from roughly a third of inpatient charts to roughly a fifth, and the audit-sample rate was being cut from ten percent to two percent. Marcus's name was on the internal email list that had announced the change three weeks ago. His job was not formally changing. The volume of his work was. He had run the math on his calendar app the morning the email had come out: at the new rates, his weekly chart load would drop from roughly forty-five inpatient reviews to roughly twelve. Less than a third.

His manager, Diane, who had been hired the year after Marcus and who had been promoted past him in 2014 because she had the bachelor's-in-health- information degree and he did not, had told him in his quarterly review two weeks ago that the team was "transitioning to a CDI-and-audit focused operating model" and that he should "explore the CCDS certification track" if he wanted to "remain competitive in the post-rollout environment." Diane had said all of those phrases without breaking eye contact, the way you do when you have practiced them. Marcus had nodded and said he would think about it. He had not signed up for the CCDS exam preparation course that the email had attached.

He started on the first of the AI-routed charts. A 78-year-old woman admitted Friday morning with chest pain, troponin elevation, and a documented history of three previous MIs. The cardiology workup had ruled in non-ST elevation MI. The primary diagnosis was clear. The AI had flagged the chart because of three competing secondary diagnoses in the documentation — chronic systolic heart failure that was probably acute on chronic, atrial fibrillation that might be paroxysmal or might be persistent depending on which note you read, and what looked like a new-onset acute kidney injury that could be coded as a complication of the contrast study or as an independent diagnosis depending on the timing.

Marcus read the entire chart. He had, over nineteen years, developed a reading speed for the inpatient documentation that nobody else on the team had, including Diane. He knew the cardiologists' note styles. He knew which residents tended to over-document and which under-documented. He knew that the nephrology consult had been called by Dr. Patel, and that Dr. Patel always specified whether an AKI was contrast-induced or not, because Dr. Patel had been burned in an audit in 2018 and had never let it happen again. The note was on page eighteen of the chart. The AKI was clearly identified as contrast-induced and clearly attributed to the cath-lab study. He coded the chart accordingly, typed a short note in the audit log explaining the disambiguation, and moved on.

The AI's confidence on this chart had been 87%. The system had been right to route it; the disambiguation work was real and required domain knowledge. Three years ago, an autonomous system would have guessed and gotten one of the three secondary diagnoses wrong. Two years ago, the disambiguation log Marcus had just written would have been a six-paragraph essay, because he would have had to walk Diane through why the AI had gotten it wrong and what to flag in the model for next time. Tonight, the log was three sentences. The AI knew the disambiguation logic. It just was not certain enough to commit to one of the three options without a human signing off.

Marcus did not feel particularly threatened by the AI on this chart. He felt, if anything, a small affection for it. The system had read the note carefully enough to know that the disambiguation mattered. It had not pretended to know what it did not know. This was the kind of chart that the autonomous-submission system was designed to defer. It was the kind of chart that he, Marcus, was good at. The AI had, in a sense, recognized his usefulness, and politely asked him for help.

He worked through the other five AI-routed charts over the next two hours. Two were routine disambiguations like the first one. Two were documentation-quality flags that he forwarded to the CDI team for clinician follow-up. The last was a complex polytrauma admission with eleven ICD-10-PCS procedure codes and a partial DRG mismatch between the AI's recommendation and what Marcus thought the chart supported. He spent forty minutes on it, reread the operative reports twice, made his recommendation, and submitted it. The AI's confidence on the polytrauma chart had been 71%. He was sure his coding was correct. He was less sure, the way he had become less sure in the past year, that his confidence in his own correctness had not itself been calibrated by the AI.

At 10:30 he started on the audit sample. Three of the thirty autonomously-submitted charts had been randomly selected for his review. He read all three. All three were correctly coded. The first was a straightforward COPD exacerbation with appropriate secondary diagnoses. The second was an elective cholecystectomy with the correct procedure codes. The third was a more complex cardiac admission with multiple secondary diagnoses, including an AKI that the AI had correctly identified as not contrast-induced (the chart had no cath study) and an atrial fibrillation flagged as paroxysmal, which Marcus would have coded the same way.

The third chart bothered him for a reason he did not immediately understand. He sat with it for a few minutes, the small green lamp warm on his hand, before he realized what he was looking at. The chart had been autonomously coded by the AI on Friday afternoon. The AI had made the same disambiguation decisions on the secondary diagnoses that Marcus would have made. Not similar decisions. The same decisions. The pattern of which acute-on-chronic call to make, the way to handle the AKI versus AKI-on-CKD distinction, the choice of which co-morbidity to elevate to a Major CC for DRG impact — all of these were exactly what Marcus would have done.

The AI had been trained on his charts. Of course it had. The hospital's vendor had used the back-coded charts from 2018 through 2024 as training data. Marcus had coded a substantial fraction of the inpatient volume in that period. The AI had learned how he made disambiguation calls, what notes he flagged for CDI, how he handled the edge cases. The system was, in a meaningful sense, a model of Marcus.

He finished the audit, closed the application at 11:18, and sat at the kitchen table for a long time, the coffee cold in the mug, the green lamp the only light in the room. The thing he had been practicing not thinking about since the email three weeks ago was that this was not really an automation story. The AI was not an algorithm replacing him. The AI was an instance of him, running at hospital scale, no longer needing him to maintain it.

He thought about the CCDS exam preparation course. He thought about his daughter, who was twenty-four and working as a nurse in Cleveland. He thought about Diane, who had not been wrong about the post-rollout environment. He thought about whether nineteen years of careful work was a thing you could grieve.

He thought, eventually, that the chart he was most proud of in his career had been a complicated polytrauma admission in 2014 — a fifteen-year-old boy who had been hit by a car on the way to a soccer game, who had survived against the odds, whose chart had included thirty-one procedure codes and a DRG that had needed three rounds of CDI work to land on the right code. Marcus had spent weeks on that chart. He had, at the time, felt quietly heroic about it. The boy had recovered fully. The hospital had been paid correctly. The audit had cleared.

The AI, he thought, would have coded that chart in 47 seconds. It would have gotten the DRG right. It would not have known how the parents had looked when they came down to the coding office to thank the team for "being the people who made sure the insurance covered all of it." The AI would not have known that that thank-you had been the small thing that had kept Marcus in the field for another decade.

He stood up, washed the mug, turned off the green lamp. The new policy started in nine hours. He would go in tomorrow morning at 8:00, the way he always did, and he would do the twelve charts the new system routed to him, and he would do the smaller audit sample, and at 4:30 he would go home. The work would be smaller. He did not yet know whether it would feel smaller.

He thought, as he turned off the kitchen light, that he should probably sign up for the CCDS course in the morning. The CDI work sounded honest. He would still be useful. He would just be useful in a different shape.

He went to bed.