Chapter 2 — The Signature

The Signature

It was 6:47 in the evening when she opened the documentation queue.

Fourteen notes pending. The clinic had run late; it usually did on Tuesdays. She had seen twenty-three patients. She was a primary care physician at a mid-sized outpatient practice that had adopted an AI documentation tool eight months earlier. The sales pitch had been straightforward and accurate: the tool would generate structured clinical notes from ambient audio recordings of appointments, reducing the time physicians spent on administrative documentation. Before the tool, she had been finishing notes at home, often past nine o'clock. Before the tool, she had sometimes spent ninety minutes after the last appointment just transcribing the day's encounters into the chart system. The tool had changed that. She was still in the office at 6:47, but she was finishing, not beginning.

She opened the first note.

The history of present illness read cleanly. The patient—a 54-year-old man who had come in for a blood pressure recheck and mentioned, toward the end of the visit, that he had been more short of breath than usual on the stairs—was described accurately in the broad strokes. Chief complaint: hypertension follow-up. The review of systems was present. The physical exam section looked right. The assessment offered a reasonable framing. The billing codes were suggested at the bottom.

She read it through once, moving at the pace the queue encouraged—enough to catch an obvious error, not enough to reconstruct the encounter from first principles. Near the end of the assessment section, she noticed that the note described the shortness of breath as "exertional dyspnea, likely deconditioned," which was one plausible reading of what the patient had said. It was also the reading that fit most cleanly into a routine follow-up note. She remembered thinking, in the room, that the shortness of breath was worth following. She clicked into the assessment and added two sentences—a note that this symptom warranted cardiopulmonary evaluation and a return visit in two weeks.

She signed the note. 6:51 PM.

She opened the second note. It was 6:52 PM when she began reading.

---

The signature takes approximately one second. What it transfers takes longer to name.

A physician's signature on a medical record is not merely administrative closure. It is a formal declaration that the record is accurate, that it correctly represents the encounter, and that the named clinician takes responsibility for what the chart will become: the foundation on which the next physician, the consulting specialist, the insurance reviewer, the pharmacist filling the refill, and the patient herself will make decisions. The chart becomes institutional memory. Every person who encounters it afterward will treat it as a careful account of what happened, produced under conditions that made care possible.

What the signature does not transmit is whether those conditions were present. It does not carry information about the queue that contained thirteen other notes. It does not record that the physician read the note once, quickly, because the alternative was to stay past seven and there was a school pickup at seven-thirty. It records only the name and the timestamp.

---

A July 2026 Government Accountability Office technology assessment found that AI tools for medical notes and coding may reduce administrative burden while remaining difficult to verify, with uncertain effects on spending and oversight.¹ The same assessment noted that U.S. clinicians average a fifty-seven-hour workweek, including roughly seven hours of administrative work per week.

The promise embedded in that second figure is real. Seven hours of administrative work per week, if genuinely reduced, is meaningful time. Physicians who spend less of their working hours on documentation could spend more of them with patients, in clinical teaching, in the slower consultations that complex cases require. Documentation toil is not a form of medical care worth defending. Some of what physicians spend time recording is genuinely rote, genuinely transferable to a system that can do it reliably, genuinely without loss to the patient who never knew the note was being typed.

The relocated work is also real. A generated note creates a new task, and the new task is different in kind from the one it replaced. Writing a note from scratch is slow, but it keeps the clinician inside the encounter: what was said, what mattered, what is uncertain, what the patient seemed to understand and not understand. It is cognitively expensive partly because it is comprehensive. Verification is a different cognitive posture. It asks: does what appears on this page match what I remember of the encounter? And it asks that question under conditions—end of a twenty-three-patient day, fourteen notes in the queue, school pickup at seven-thirty—that may not be well matched to careful verification.

The GAO assessment noted that it was difficult to verify the accuracy and completeness of AI-generated notes. That difficulty is not incidental. It is a property of what the notes are: fluent summaries generated from the encounter and whatever record context the system can access, using learned statistical patterns rather than a human memory of the particular person who was in the room. The reorganization may be accurate. It may also have softened an ambiguity, dropped a detail that seemed minor at the scale of the full encounter but turns out to be clinically relevant three months later, or carried forward an assumption from an earlier visit that this patient has been trying to correct for a year.

---

Verification is not the same work as drafting. It is harder.

Drafting asks what to say. Even a careful writer who struggles with documentation is still inside the encounter while writing: retrieving, organizing, making choices about what to include. That process is slow, but it is also a second pass through the clinical material. The act of writing a note is partly a check on what happened—an opportunity to notice, in the act of describing, that something doesn't add up, that a detail was left ambiguous, that a follow-up was promised and not entered.

Verification asks something different: whether what was said is true enough for the decisions that will lean on it. And it asks that question from outside the encounter, looking at a finished document, without the benefit of the drafting process as a second pass. The verifier has to hold the encounter in memory and compare it to the text, looking for discrepancies. Under time pressure, with fatigue, with thirteen other notes remaining, that comparison is harder to do carefully than it appears.

The gap between what verification requires and what is available to do it is easy to underestimate because the note reads as finished work. Fluency disguises uncertainty. A human draft often carries the marks of its production: gaps, hedges, incomplete sentences, inconsistencies that invite a second look. A generated draft can arrive already polished, which removes the social cue that something still needs checking. The verifier faces harder epistemic work—is this specific sentence true of this specific patient?—while the interface presents the task as nearly complete.

Fluency can also conceal a category of error that is difficult to catch through verification alone. A documented allergy omitted because the patient mentioned it conversationally and the model treated it as background. A symptom softened into the language that fit a billing code better than the patient's actual description. A prior diagnosis carried forward because it sounded consistent with this presentation and nothing in the note required departing from it. A follow-up ordered in the previous visit and not resolved, not mentioned in this one because the note was generated from this encounter's audio and nothing in the template asked about continuity. None of these failures require malice or even inattention. They require only a fluent surface and a reader whose attention is divided between the document and the memory of an encounter that happened seven hours ago.

Once signed, the note becomes the record. Nurses, consultants, insurers, and future physicians will inherit it as if a careful human had authored every sentence under conditions that made full attention possible. The correction to the assessment section—the two sentences about cardiopulmonary evaluation—will be visible. The original framing of the shortness of breath as likely deconditioning, before the correction, will not show that it came from a model that had never been in the room.

---

The question of saved time is not merely whether time was saved, but what happened to it.

A reduction in documentation hours can become patient attention. It can also become more patients scheduled into the same day, which absorbs the freed time before the clinician experiences it as margin. It can become more notes to sign under the same fatigue, because the generation is faster and the queue refills. It can become inbox management, prior authorization appeals, administrative tasks that were always present and are now reachable because the documentation work moved earlier in the workflow. The tool's framing is assistance. The workflow's logic may be more demanding: not that the physician works less, but that the physician works differently, with administrative tasks redistributed rather than reduced.

The distinction between assistance and autonomous administrative action is not settled by marketing copy. It is settled by whether the institution still allocates time and explicit authority for the clinician to refuse, revise, or stop—to say "this note does not represent the encounter and I am not signing it as written," and to have that refusal be a supported option rather than a personal sacrifice against the queue.

A fifty-seven-hour workweek does not leave generous margin for close reading of every paragraph a model produced. Seven administrative hours, if reduced, may return to the schedule as available time for more encounters before they return to the physician as felt relief. The queue does not shrink because generation became faster. It adjusts to the new production rate.

---

The formal structure of responsibility stays intact through all of this. A human name remains on the chart. Formal responsibility attaches. Regulators and payers see a responsible party. The chart shows that a physician reviewed and signed each note. What the chart does not show is whether the conditions for that review were present.

Responsibility without time is not safety. It is a transfer of risk onto the person least positioned to refuse the queue. The physician's signature establishes formal responsibility and may carry legal consequences. It does not establish that the workflow provided the conditions that responsibility assumes.

This is not an argument that clinicians should write every note by hand forever. Handwriting did not guarantee accuracy. Dictation did not guarantee attention. Templates did not guarantee truth. The older bottleneck of documentation consumed time and attention that could have gone to patients, and that was a real cost with real consequences. The concern is more specific. When generation accelerates and verification is the remaining human task, the scarce work becomes the quality of that verification. If the organization measures only notes completed, signatures obtained, and minutes saved, it will not see whether the verification was adequate to the stakes—whether anyone had enough context, rest, and time to catch the sentence that did not belong.

An institution that wants the benefits of generated documentation without degrading care has to staff the pause the signature represents. Not only the model that fills the page above it. Not only the platform that delivers the note to the queue. The pause itself: the time, the margin, the permission to say that this one needs another look.

The queue was at thirteen when she closed the second note at 6:58. She had corrected one thing in the second note—a date that was off by a day—and signed it.

She opened the third. This one was harder to read quickly. The patient was a sixty-eight-year-old woman with multiple chronic conditions: diabetes managed over many years, hypertension, a history of cardiac disease. The note described the encounter accurately in broad strokes, but one sentence in the medication reconciliation section listed a dosage that didn't match what the physician remembered from the visit. She checked the patient's current medication list in the chart, verified that the dosage the note recorded was the old dosage—changed at the last visit, three months ago, a change the model had apparently not registered as superseding the prior record. She corrected it. She made a note to herself to check whether the chart had been updated properly after the previous visit.

It was 7:09 when she signed the third note. The queue was at twelve.

The error in the third note was the kind that a careful clinician drafting the note would likely have caught, because drafting requires pulling the medication list to populate the reconciliation section. Verification, done quickly, may not. The correction took eleven minutes. She had eleven notes remaining.

Footnotes

1. U.S. Government Accountability Office, "Science & Tech Spotlight: AI for Medical Notes and Coding," GAO-26-109116, July 16, 2026, https://www.gao.gov/products/gao-26-109116. Finds that AI tools may reduce documentation burden while remaining difficult to verify for accuracy and completeness, with uncertain effects on spending and oversight. The fifty-seven-hour average workweek and roughly seven administrative hours per week are drawn from the assessment's characterization of clinician time burden. The particular errors in this chapter's composite scene are illustrative; the broader difficulty of verifying AI-generated clinical documentation is documented in the cited assessment. ↩