Record of Care · Episode 7 of 7 · Medical fiction

The Finding

The conference has enough evidence to reach some conclusions. It must resist reaching others.

  • Medical fiction · Episode 7 of 7
  • With notes: An Algorithm Is Not the Standard of Care

Published · revised as medical fiction

9 min story · 14 min notes · sources in the notes checked through (Pacific)

Conceptual illustration: a findings page with five separate lines, marked as confirmed, struck through, confirmed, assigned and unresolved.
Conceptual illustration. No real patient, clinician, screen or record is depicted.

Julian Vale opened the conference the way he opened every one, by reading aloud what it was.

“This is the monthly clinical quality review. It looks at care and documentation so we can improve them. It makes findings and recommendations to the medical director. It doesn’t decide anyone’s employment or privileges. If a matter ever needed that, it would go to a different process with its own rules, and the clinician would be told so in writing.” He looked up. “I’ve read that paragraph forty times and never thought about why it’s there. I thought about it this week.”

There were seven of them. Sam Okoye, the second reviewer, sat with his arms folded. He had practiced at Marsh Road for twenty years and did not care for meetings. Dr. Pruitt joined by video from the same site. Noor had the minutes. Inez and Celia Ward were there for the items that were theirs. Lena was there for her section and would leave when it ended.

She had brought one page. Her first draft had run to five, and on Thursday night she had made herself look at it as Mara Ellison might.


“Section A,” Julian said. “The historical reviews. Mine and Sam’s.”

Noor put the new table on the wall. It replaced the single column headed Prior intervention with four dated lines for each review: when the patients were seen, when the review was entered, when the clinician was told, and the first date on which anything she did differently could show.

“All the encounters in both samples are from the same three weeks,” Julian said. “The feedback came three weeks after the last of them. The sentence in Monday’s packet, similar deficiencies persisted after feedback, came from the template comparing the feedback date to the date Sam’s review was typed in. It’s wrong, and I let it stand for three days because it matched what I expected. I’m withdrawing it.”

“I’m not withdrawing my review,” Sam said.

“Nobody’s asking you to.”

“Five visits, no reconciliation acknowledgment on any of them. That’s what I found. That’s still true.”

“It is,” Lena said. “I’d like it recorded as true.”

Sam turned his head. He had expected an argument, and without one he needed a moment.

“And the add-on visit,” he said. “Week two. I couldn’t tell from that note why you did what you did.”

“Neither could I, reading it cold. I think the care was right. The note doesn’t show it.”

Dr. Pruitt spoke from the screen. “I’d separate those in the finding. The acknowledgments are a workflow step that wasn’t completed. From the medication lists it looks as if the reconciling itself was done. That isn’t the same as care not happening. The add-on is a documentation problem. Neither one is a clinical failure on what we have, and I’d rather the summary didn’t file them under one word.”

“Agreed,” Julian said. “And five missed acknowledgments is one habit counted five times. I’d say that in the summary too, so no one reads it as five incidents.” He turned to Noor. “Finding stands, as a historical finding. No statement about response to feedback, because we have nothing that measures it. Reassessment next quarter, drawn only from visits after the feedback date. My look at her recent notes goes in as an observation. It wasn’t a scored review and I won’t dress it up as one.”

Noor typed. Lena watched the sentence that had stopped her on Monday leave the wall.

It did not feel like winning. What stayed on the wall was accurate, and some of it was not flattering.


“Section B. Ms. Ellison.”

Lena had her page. She did not read from it.

“The signed note said three things that weren’t so. It gave a diagnosis as established when I’d said possible. It said a referral was placed before I’d placed one. It said the patient refused evaluation when she’d asked for two weeks. I corrected the first draft. A second recording produced a new draft without my corrections. I signed it without rereading the plan. That’s my part of it.”

“The tool should have warned you,” Inez said.

“It should have. I still signed it.”

Sam unfolded his arms. “The signer’s the signer.”

“Yes.”

He looked at her a moment longer. “I’ve done it,” he said. “Resumed a recording because the patient remembered something at the door. I’d have to go back and look at what I signed.” He did not sound pleased to have thought of it. “I just didn’t have a referral bounce.”

Julian took them through it in four parts, and Noor wrote four headings because he asked her to.

Whether the note was faithful to the visit: it was not, and the clinician’s final check was inadequate. That was a finding. The recommendation was specific. After any resumed recording, read the plan against the orders before signing. Inez would ask the vendor whether the tool could say plainly, at the moment of regeneration, that manual edits had not been carried over, and in the meantime she would put it on a single page for every clinician.

Whether the care actually planned was reasonable: it was. Labs, imaging, a specialist, time the patient had asked for.

Whether the action the note described had happened: an order had been sent to a practice that no longer took the clinic’s patients, and a setting had closed the task before anyone could learn that it was declined.

“That’s mine,” Celia said. “Not the order. The setting.” She had her tablet, but the list on it was a different one. “Tomas ran the report. Twenty-three referrals since the spring went to practices on the changed-contract list and closed themselves. He’s spoken to nineteen of those patients. He’s still calling four. The acceptance check goes to change review Tuesday with Dr. Rowan’s name on it as clinical reviewer. Two hours of his phone time move to the float on Monday. I’ll bring the count back here next month, including any of the four we can’t reach.”

“Any harm?” Dr. Pruitt asked.

“Delay,” Celia said. “I don’t know yet about more than delay. I’m not going to say none until the four are reached.”

Noor wrote that down as she had said it.

“And the fourth part,” Julian said. “How the wording got there. We know the first draft and the second differ. We know the edits weren’t carried over. We don’t know why the second draft said refused.”

“The patient thinks she may have said she wasn’t doing the X-ray that day,” Lena said. “She’s not certain. She’s certain she didn’t say never.”

“The transcript request is with privacy,” Inez said. “It may come back and settle it. It may come back and not. It may have aged out.”

“Then we say that.” Julian dictated it slowly. “Origin of the word refused: not established. Request pending. This does not change the finding on the signed note, and it is not a finding against anyone.” He stopped. “Has she been looked after? I’d like that in the record before the rest.”

“Appointment three weeks Monday, imaging the same morning,” Lena said. “The note’s corrected with the original left visible. The diagnosis field is corrected. She has a new summary. She wrote her own statement, and it went to the county clinic with the referral this morning.”

“Whose idea was the statement?”

“Hers. Tomas told her she could.”

Sam grunted, and it might have been approval.


“Section C is mine,” Sam said. “I put it in yesterday.”

He put it on the wall himself. Clinician dismissed AI-suggested diagnosis of atrial fibrillation. Patient subsequently diagnosed in the emergency department. Question: missed diagnosis.

“I’ll tell you how I got there,” he said. “The hospital sends us a notice when one of ours is seen in the ED. I opened the chart. There’s a line in the activity that says a suggested diagnosis was dismissed three weeks before. I saw the suggestion, I saw the ED result, and I wrote the question.”

“Did you open the visit?” Julian asked.

“I read the note. I’ll be honest, I read it knowing how it ended.”

Lena waited until he looked at her.

“He came in because his heart raced after coffee. His brother has atrial fibrillation and he was worried he had it too. He said so three or four times. I ran a tracing in the room, and it was normal sinus rhythm. I think the tool heard atrial fibrillation several times in that conversation and put it in the assessment. I can’t prove that’s why. I took it out because the man in front of me didn’t have it on the tracing I was holding, and the person who did have it was his brother.”

Inez nodded slowly. “The log says dismissed. It doesn’t say what the suggestion was built from. I can’t tell you she’s right about the brother. I can tell you nothing in there says she’s wrong.”

“So the suggestion wasn’t a finding,” Dr. Pruitt said. “It was a word that was in the room.”

“He did have it, though,” Sam said. “Three weeks later.”

“He had it three weeks later,” Dr. Pruitt said. “It comes and goes. That’s what paroxysmal means. A normal tracing that day doesn’t tell us he didn’t have it that week. It also doesn’t tell us she missed something that was there to see. The question I’d ask is the one I’d ask without any software. First visit, palpitations with a clear trigger, normal tracing, a family history. Was it reasonable not to order a monitor that day?”

No one spoke at first.

“I’d have gone either way,” Sam said at last, and it was plain he disliked saying it. “I’ve done both.”

“So have I,” Dr. Pruitt said.

“Then it isn’t a missed diagnosis.” Sam reached over and struck his own heading. He did not ask Noor to do it. “But I’m not done. Your note says return if it recurs. It doesn’t say how soon, or what to do if it happens at two in the morning. He went to the ED at two in the morning, which was right, and he didn’t get that from you.”

Lena started to answer and stopped. He was correct.

“No,” she said. “He didn’t.”

“That’s a recommendation, not a finding,” Julian said. “Sam?”

“Recommendation.” He sat back. “And I’d like it noted I wrote that item from what the machine said and what happened afterward. I didn’t start with what she knew that day. I’d mark a resident down for that.”

“It’s noted,” said Noor, who had already typed it.


There was one more piece of paper. Lena passed it down the table.

“This is from Ava Chen. It isn’t about my section. The review queue shows PA supervision items, quality-sample items and consult requests with the same label and the same button. She’s proposing three labels and a required line saying what the reviewer did. Eli Navarro has signed it. So have I.”

Julian read it through.

“I’ll take it to Dr. Amari with the summary,” he said. “And I’ll go with it.”

When her section closed, Lena stood. No one said anything to mark it. Sam was already reading the next clinician’s packet with his arms folded again. Celia caught her eye and held up four fingers, the patients still to be reached, and then turned back to her list.

Lena had one new-patient visit that afternoon. Celia had put it into a Friday slot that was truly open, and they had agreed on it together. At the end of the visit she reconciled the outside medication list against the bottles the man had brought in a grocery bag. Then she moved the cursor to the small control at the bottom of the screen and clicked it. She had read the plan first.


On the Monday of her appointment, Mara Ellison sat on the end of an examination table at the county clinic with her X-rays already taken and her jacket folded on her knee.

She had rehearsed it in the car. She would say first that she had not refused anything. She would say it before anyone could ask.

The rheumatologist came in with a tablet under her arm, shook Mara’s hand gently, as people in that specialty learn to, and sat down on the stool.

“I’ve read what Dr. Rowan sent,” she said. “Possible inflammatory arthritis, blood tests that don’t show the usual markers, and you needed a couple of weeks to arrange things around the bakery. And I’ve read what you wrote.” She set the tablet on the counter, face down. “That’s what other people think. Tell me about your hands.”

Mara found that she did not need the sentence she had rehearsed.

She held out her hands.

The questions behind this episode

Clinical and governance notesAn Algorithm Is Not the Standard of CareSoftware can suggest an answer. A reviewer must still establish the clinical obligation.

Everything below is commentary on real clinical, technical and legal questions. It refers to the story only as an illustration. Characters named here are fictional.

Evidence at a glance

Established by the cited sources
Medical Board of California expert-reviewer guidelines frame the standard around a reasonably careful physician in similar circumstances; North Carolina’s 2026 position statement addresses responses to AI recommendations but is not California law; FDA’s CDS guidance treats independent review of a recommendation’s basis as one nondevice criterion; research describes automation bias as over-reliance on automation.
Not established, and not claimed
That any real reviewer used AI output as a standard of care, or that any physician’s removal of a generated diagnosis was correct or incorrect. The encounter is hypothetical.
The analysis in these notes
A generated recommendation can raise a question. Only clinical evidence, contemporaneous facts and explained reasoning can establish the obligation against which a physician is judged.

In the final episode a reviewer sees that a generated diagnosis was dismissed and that the patient was later given that diagnosis elsewhere. He writes missed diagnosis and, to his credit, takes it back when the conference reconstructs what was known at the visit. The same conference lets a real criticism of the physician stand.

The general form of the problem is this. An AI assistant produces a note containing a diagnosis the physician has not established. The physician removes it, documents the uncertainty and records the actual plan. Months later a reviewer notices the deletion. Perhaps the physician corrected an unsupported inference. Perhaps the physician overlooked a condition that should have been recognized. Perhaps the software converted a possibility discussed during the visit into a definitive statement. Perhaps the reviewer is seeing a label without the clinical context that gave it meaning. The deletion alone cannot settle the question.

A dangerous shortcut is available: treat the generated answer as the expected answer, and treat the physician’s departure from it as the event requiring justification. Once that shortcut enters a review process, a tool intended to assist clinical judgment begins to function as its examiner. An algorithm can draw attention to evidence, organize information and suggest a useful question. Its output still requires a clinical foundation. A standard of care cannot be established merely by pointing to what appeared on a screen.

A diagnosis is more than a sentence

The hypothetical is plausible because modern clinical software can do more than transcribe. Microsoft’s documentation for Dragon Copilot, for example, describes diagnosis generation from ambient encounters and reconciliation with existing encounter diagnoses in an Epic workflow. That describes a product capability; it does not establish which features any particular organization has enabled or whether a particular generated diagnosis is correct.[1]

A reviewer therefore needs to distinguish several things that can look deceptively similar in a finished chart: a patient reporting a previous diagnosis, a clinician considering a possibility, a system extracting a term, and a clinician adopting a current assessment.

Those are different claims about what is known.

Imagine a patient mentioning that a relative has a particular condition. Now imagine the same condition appearing in a generated assessment. A physician who removes it may be restoring the boundary between family history and the patient’s own health. A count of accepted and rejected suggestions would miss the central issue: whose condition was being described?

The same problem arises when uncertainty disappears. “We discussed whether this could explain the symptoms” and “The patient has this condition” express different levels of confidence. Good review must examine that difference before treating the shorter, more definite statement as superior documentation.

This is a proposed discipline for evaluating the record: classify the statement before judging its inclusion or removal. Was it a report, an observation, an inference, a working assessment, or a confirmed conclusion? What evidence supports it? Who adopted it? What did the patient and clinician actually decide?

Without those distinctions, completeness can become a demand to preserve assertions that should never have become facts.

The expert must identify the standard

The Medical Board of California’s currently linked Expert Reviewer Guidelines offer a useful foundation. The guidelines, revised in November 2023, frame the standard around reasonably careful physicians in similar circumstances at the time of care. They direct experts to identify each medical issue, describe the applicable standard, apply the facts, and explain the conclusion. They also caution reviewers against substituting their own practice preferences for the standard.[2]

The Board’s expert program describes independent, impartial review and expects reviewers to explain the basis of their opinions.[3] These are existing professional-review principles, not an AI-specific Board ruling.

Applied to AI, the implication is straightforward: the reviewer must establish why a diagnosis, investigation, treatment, or follow-up action was clinically required. Showing that software suggested it does not complete that analysis.

A useful recommendation may lead the reviewer to strong evidence. The evidence may reveal a real failure. But the recommendation should not become a substitute for the explanation connecting the patient’s circumstances to the professional obligation.

This protects the integrity of an adverse finding as much as it protects the physician. A conclusion grounded in clinical evidence can survive questions about the tool that first raised the issue. A conclusion grounded only in the tool’s output has nowhere further to go.

Three disagreements that need different evidence

An AI-assisted record can generate several superficially similar complaints. They should be investigated separately.

First, a reviewer may question why the physician removed a suggested diagnosis. The relevant inquiry is whether the diagnosis was supported and whether the resulting assessment accurately represented the encounter. If the software introduced an unsupported conclusion, retaining it would not improve the record. If the available evidence supported an important condition that the physician failed to address, deletion would not make the omission disappear.

Second, a reviewer may question why the physician did not follow a suggested treatment or investigation. Here the clinical basis of the suggestion matters: the applicable evidence, the patient’s circumstances, what was already known, relevant alternatives, and the reasoning actually used. Agreement with a recommendation is not proof that its implementation was appropriate. Disagreement is not proof that indicated care was withheld.

Third, a reviewer may question why an AI-generated plan was absent from the signed note. Before calling that an omission, establish whether it was ever the clinician’s plan. A system’s proposed next step and a decision communicated to a patient are not interchangeable events.

These distinctions do not grant a clinician permission to dismiss useful warnings casually. They require the reviewer to articulate the actual criticism. Was there a failure to recognize a finding? A failure to act on an established risk? Inaccurate documentation? An uncompleted workflow step? Each deserves its own evidence and response.

A single label such as “AI recommendation rejected” cannot carry all of those meanings. It may describe an interface event while saying very little about the quality of care.

The order of analysis matters. Establish the clinical concern, identify the relevant standard, reconstruct the decision, and then determine whether the software history helps explain what happened. Beginning with acceptance rates reverses that order.

The tension in asking physicians to explain disagreement

The North Carolina Medical Board addresses this issue directly. Its medical-records position statement, amended in September 2026, says clinicians using AI remain responsible for responding appropriately to its recommendations. Following a proposed treatment should have a defensible rationale; treatment that deviates from an AI proposal should also have a documented rationale. The statement generally connects disagreement to the applicable standard of care.[4]

This is North Carolina guidance, not California law. It should be read accurately, including its expectations about both acceptance and departure. It does not say that every suggested code or generated phrase requires a separate rebuttal, and it does not declare an AI answer to be the clinical standard.

The practical concern is how an institution translates that guidance into workflow. If following the tool requires one click but disagreeing requires a lengthy defense, the interface can make agreement easier to document than judgment. That is a design risk worth examining before it becomes a performance metric.

An appropriate record should explain consequential clinical decisions. It should not become a transcript of every possibility a machine proposed. The distinction requires clinical leadership, clear examples, and attention to the volume and relevance of suggestions.

Institutions should ask whether their documentation rules make reasoning visible or merely make compliance convenient. A short, meaningful explanation of the actual assessment can be more informative than a long list of reasons for declining irrelevant outputs.

The goal should be a defensible account of patient care. Turning every discarded suggestion into a presumed deviation would create a different goal: defending the physician against the software’s imagination.

Automation bias runs in both directions

Researchers have a name for one of the risks at the center of these notes. A systematic review in the Journal of the American Medical Informatics Association described automation bias as the tendency to over-rely on automation, observing that decision support generally improves performance while sometimes introducing new errors that go unrecognized. The mitigators it identified included implementation factors, such as training and an emphasis on user accountability, and design factors, such as where advice appears on the screen and whether it carries updated confidence information.[11]

That literature cuts in two directions, and a fair review must respect both.

A clinician who accepts an incorrect suggestion because it was presented confidently and conveniently has made an error the review should identify. A clinician who overrides a correct warning has made a different error the review should also identify. Automation bias excuses neither.

But a review process can itself become a source of automation bias. If acceptance of generated content is treated as the default safe answer, and every departure as the event requiring explanation, the institution is training clinicians to defer. The vigilance that the research identifies as protective, which means questioning the output, checking it against the patient and declining what is unsupported, becomes the behavior most likely to draw scrutiny.

The design response is to evaluate decisions on their clinical merits rather than on their agreement with software. A review that asks “was this correct for this patient?” works against automation bias. A review that asks “why did you disagree with the tool?” may reward it.

Independent review requires a visible basis

FDA’s January 29, 2026 clinical decision support guidance reinforces the importance of reviewable reasoning in a specific regulatory context. For the statutory nondevice CDS exclusion it discusses, enabling a healthcare professional to independently review the recommendation’s basis is one of four criteria. FDA recommends usable information about intended use, inputs, development and validation, and relevant patient-specific information and unknowns.[5]

That is a framework for assessing software functions, not a declaration that a recommendation is clinically correct. FDA also cautions that failure to satisfy a nondevice CDS criterion does not automatically place a function under active device regulation; other exclusions or enforcement-discretion policies may apply.[6]

The practical lesson for healthcare organizations is an argument, rather than an additional FDA requirement: if independent review is central to the safety model, review must be possible under real working conditions.

Can the physician see the information used? Can a historical fact be distinguished from a new inference? Is the missing context obvious? Is there time to investigate a consequential discrepancy? Can a correction be made without creating confusion elsewhere in the record?

The answers affect what “reviewed” means. A signature field records an action. It does not demonstrate that the surrounding system made a particular error easy to detect.

Microsoft’s transparency documentation likewise describes clinician review of generated content and recognizes possible errors, including omissions and unsupported content.[7] The existence of that responsibility should drive better review design. It should not end the investigation into how a particular error occurred.

Reconstruct what was available at the time

A serious review should resist an especially seductive form of hindsight: running an old encounter through a newer tool and treating the new output as what the original clinician should have known.

As an evidentiary matter, those are different situations. The later system may have access to additional records, different instructions, a different model, or knowledge of the eventual outcome. Its answer can generate a question worth investigating. It cannot retroactively place that information into the earlier examination room.

An institution using such analysis should be able to identify the inputs, version, purpose, and timing of the later review. Was the model used to organize existing evidence, identify possible issues, or supply an expected answer? Were its findings independently checked? Was the clinician allowed to address the actual factual concern?

This does not require an assumption that the original decision was sound. It requires an honest account of the conditions under which it was made.

A physician can make an avoidable error with incomplete information. A reviewer can also make an avoidable error by silently replacing incomplete information with a later, fuller record. Both errors deserve scrutiny.

Accountability has more than one address

The strongest objection to this argument is also the most important one: physicians remain responsible for the records they sign and the decisions they make. An explanation that software produced the words is not, by itself, an explanation of appropriate care.

That objection is correct, and nothing in these notes displaces it. The argument for examining AI’s role should strengthen accountability by making it accurate.

An organization may have legitimate documentation requirements, a sound review rubric, and qualified reviewers who identify a real problem. A physician may have received useful feedback and still need to improve. Recognition that a tool can err does not invalidate every finding involving an AI-assisted note.

But a complete improvement analysis should also examine decisions made beyond the individual encounter. Who selected the functions? Who configured how information appears? Who decided which outputs become easy to accept? Who established the review workflow, training, escalation path, and time available to resolve discrepancies?

These are proposed governance questions. They do not assign legal liability in advance. They identify where an organization may be able to prevent recurrence.

If clinicians repeatedly correct the same type of unsupported statement, treating each correction as an isolated physician problem misses a possible system defect. If clinicians repeatedly overlook the same misleading presentation, additional reminders may be insufficient. The organization needs to test the presentation itself.

Responsibility should follow the decisions each participant can make. The clinician must exercise appropriate judgment. The organization should make that judgment practicable. The developer should provide usable information about the tool’s intended behavior and limitations. A review process should be able to examine all three without assuming that one person’s signature explains everything.

Different consequences require different processes

Routine improvement work remains essential. HRSA’s Health Center Program Compliance Manual requires an ongoing QI/QA system, systematic assessment of records, and periodic clinical review. It also recognizes management and governing-board oversight and gives health centers discretion over methodology. It does not prescribe a universal chart sample or passing score.[8]

That framework leaves room for useful education, focused review, and correction. It also makes the quality of the chosen method important. If a rubric measures workflow completion, it should say so. If it measures clinical reasoning, its criteria should connect to clinical obligations. If the two are combined, the resulting score should not conceal the distinction.

Formal legal consequences involve additional questions. California’s section 805 reporting framework can cover specified medical-disciplinary employment actions as well as privileges actions. Section 809.1 connects formal notice and hearing rights to qualifying final proposed actions reportable under section 805.[9, 10] An unfavorable routine review does not automatically establish that those triggers have occurred.

Neither does the word “routine” settle the issue if the process or consequences later change. The institution should state what process is being used and what its findings are intended to support.

The governance proposal here is to make ordinary feedback sufficiently clear that escalation does not require reconstructing its meaning from memory. Identify the criterion, the evidence, the concern, the requested correction, and the opportunity to respond. More consequential action requires correspondingly careful attention to the applicable procedural rules.

What an AI-aware review should establish

Before a generated suggestion is used to assess a physician, five questions should be answered.

What is the actual clinical concern? Name the action or omission being evaluated. Avoid using disagreement with software as a substitute for identifying it.

What supports the expected care? Identify the clinical evidence and applicable professional standard. Explain why they fit this patient and this encounter.

What did the software contribute? Distinguish source information, generated inference, proposed language, and the clinician’s final decision. State when that distinction cannot be reconstructed.

What information was available to the clinician? Establish the relevant timeline. Separate facts known during care from material learned afterward, and separate a missed finding from a misleading display.

What does the finding justify? Education, a documentation correction, a system change, further clinical investigation, and formal disciplinary action are different responses. Explain why the evidence supports the response selected.

This is a proposed review method, not a claim that every jurisdiction already mandates this exact checklist. Its value is that it can substantiate a concern as well as disprove one. It gives an organization a stronger basis to act when care was deficient, and a stronger basis to correct its own assessment when the evidence points elsewhere.

It also gives physicians a constructive obligation. A request for fair review should come with a willingness to examine the underlying care, supply relevant context, acknowledge errors, and improve. Transparency should work in both directions.

The shape of a defensible finding

The five questions above can be carried into the written finding itself. A finding that survives later scrutiny tends to contain the same elements, in roughly this order.

ElementWhat it statesThe AI-specific caution
The issueThe specific act or omission under reviewNot “disagreed with the AI,” unless that disagreement is itself the clinical issue
The standardWhat a reasonably careful clinician would have done in similar circumstances, and its basisA generated recommendation is not the basis
The facts availableWhat the clinician knew or could reasonably have known at the timeExclude information a later tool or record supplied afterward
The decisionWhat the clinician actually concluded and didDistinguish the clinician’s plan from the system’s proposed plan
The documentationWhether the record accurately reflects the decisionFidelity and appropriateness are separate findings
The system contributionWhat the software or workflow added, removed or obscuredState when this cannot be reconstructed
The conclusionWhether the standard was met, and whyExplained in clinical terms a peer could evaluate
The responseEducation, correction, system change, further review or formal actionProportionate to the established finding

A finding built this way can still be wrong, but it is wrong in a way that can be examined and corrected. A finding that consists of a score and a reference to a generated suggestion cannot be.

Keep the patient at the center of the standard

The hypothetical physician who removed a diagnosis may have made the right decision. The physician may have made the wrong one. An honest review must remain open to both possibilities.

What it cannot do is allow the software’s answer to decide which possibility deserves consideration.

We should welcome tools that help clinicians find important information and communicate care clearly. We should also insist that the institutions evaluating clinicians can explain the difference between a suggestion, a supported finding, and an obligation.

A useful system gives medicine another way to ask a question. It should not acquire the authority to answer every subsequent question about the physician merely because its words came first.

The standard belongs to the patient’s care: what a reasonably careful clinician should have recognized, considered, done, and documented in the circumstances. When an algorithm helps reveal that standard, examine the evidence. When it is offered in place of the standard, ask for the clinical reasoning.

Sources and scope

Sources checked through 19 November 2026 (Pacific). Each source establishes only what its own scope supports; vendor documentation describes product capabilities, not any organization’s configuration. Source numbering restarts in each essay of the series.

  1. Microsoft Learn, Intelligent diagnoses retrieval, updated August 3, 2026. Product capability and workflow documentation; not evidence of any particular organization’s configuration.learn.microsoft.com
  2. Medical Board of California, Expert Reviewer Guidelines, revised November 2023, especially the report format and standard-of-care sections.www.mbc.ca.gov
  3. Medical Board of California, Expert Reviewer Program.www.mbc.ca.gov
  4. North Carolina Medical Board, 3.2.1: Medical Records—Documentation, Electronic Health Records, Access, and Retention, amended September 2026. A North Carolina position statement, not California law.www.ncmedboard.org
  5. FDA, Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff, January 29, 2026. Nonbinding guidance interpreting the statutory framework.www.fda.gov
  6. FDA, Clinical Decision Support Software Frequently Asked Questions, especially Q4.www.fda.gov
  7. Microsoft Learn, Dragon Copilot transparency documentation, updated November 5, 2026. Vendor description of capabilities, limitations, and intended clinician review.learn.microsoft.com
  8. HRSA, Health Center Program Compliance Manual, Chapter 10: Quality Improvement/Assurance, last reviewed November 2025.bphc.hrsa.gov
  9. Medical Board of California, Section 805 reporting form, revised July 2026.www.mbc.ca.gov
  10. California Business and Professions Code §809.1. Primary statutory text reproduced in the 2025 code; official live publisher was inaccessible during research.law.justia.com
  11. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121–127. doi:10.1136/amiajnl-2011-000089. A review that predates generative AI; cited for the concept and the mitigators it identified, not for the performance of any current product.pmc.ncbi.nlm.nih.gov doi.org

Related reporting and resources on this site

About this page. This is Episode 7 of Record of Care, a fictional medical drama in seven episodes, followed by its clinical and governance notes. The story is fiction: its characters, clinic, documents and events are constructed, and nothing in it is evidence about a real person, organization, product or deployment. The notes cite real sources with their limits and label their proposals as proposals. First published 19 November 2026 as part of the essay series Clinical AI, Physician Autonomy, and the Record of Care; revised as a serialized work of medical fiction on 19 November 2026. The earlier edition is retained in the site’s editorial archive. Drafted with AI assistance and published on the author’s instruction. Sources were carried over from the first edition, checked through 19 November 2026 (Pacific). General information, not individualized medical or legal advice. Corrections: editorial standards · methodology.

Medical fiction with clinical and governance notes. First published 19 November 2026; revised as a serialized work of medical fiction on 19 November 2026, on the author’s instruction. Sources in the notes checked through 19 November 2026 (Pacific). Corrections: editorial standards.

Search KPSGill.com

Search the full publication — patient guides, policy, investigations, enforcement records and research.