Notes before codes
Most studies find that AI scribes raise the bill. They should have to show what they do for patients, starting with the accuracy of their notes.
Listen to this special topic
read by an AI voiceThe software that listens to a doctor's visit and drafts the note is moving toward the bill. On Oct. 8 Abridge announced an agreement to expand the use of its scribe across the University of Texas System. In September it announced a product that checks hospitals' inpatient claims against the record before they are sent, so that discrepancies can be resolved before they become denials. The product is built, the company says, "on the same platform and clinical record already used at the point of care," and the hospital's own staff decide what to change.
Scribes may ease doctors' paperwork and improve patients' satisfaction, and most studies of their effect on billing find increases. A tool that raises the bill should have to show what it gains for patients, beginning with whether the record it writes is right. One whose notes get worse should be fixed or withdrawn quickly, whatever it earns.
The gains for doctors are real but small. In an observational study of outpatient clinicians at five American academic medical centers, using a scribe was associated with about a quarter of an hour less documentation for every eight hours of scheduled visits, no significant change in after-hours time in the electronic record, and extra visits that brought in little money. In a randomized trial at UCLA, one product cut the time doctors spent on notes by about a tenth and the other made no significant difference, though both showed signs of easing burnout. In a retrospective study of emergency visits at Mass General Brigham, AI scribes saved less time per note than human scribes did and produced longer notes, though in a primary care preprint, note length did not change. None of these studies measured whether patients fared better.
Tested blind against clinicians on simulated visits, the scribes' notes fell short. Researchers at the Veterans Health Administration had 11 AI scribes and 18 clinicians write notes from recordings of primary care visits acted out by standardized patients, and raters who did not know who wrote each note scored them on a standard measure of quality. "In every case, human notes scored higher, in some cases dramatically so," the American College of Physicians said in announcing the study. The clinicians did not write under the pressures of a real clinic, which may have flattered them. In a randomized trial in everyday practice at the University of Wisconsin, by contrast, the authors concluded that documentation time fell "without compromising diagnosis, billing compliance, or note quality."
On billing, the broadest evidence is also the weakest. Trilliant Health, a health care market-intelligence firm, analyzed claims from six health systems that had announced AI scribes and found that from 2018 to 2024 the share of new patients' visits billed at the two highest levels rose by roughly 12 to 20 percentage points. But it had no comparison group or dates for when the scribes arrived, and the rise coincided with a 2021 rewrite of the coding rules. Studies with a comparison group are few and point both ways: a coding audit found higher visit levels too, though an emergency study at Mass General Brigham found no difference in the physician work billed per shift hour. In private, those who pay and those who are paid sound surer. Investors, health plans and providers at a roundtable held by the Peterson Health Technology Institute, a nonprofit that evaluates new technologies, agreed, as its executive director recalled, that "it's quite clear scribes are increasing coding intensity."
Somebody pays for higher codes. Higher-intensity visits carry higher copays for patients, and Hannah Neprash, a University of Minnesota researcher studying the question, warns of surprise bills: a scribe could turn a preventive visit into a problem visit, "which the patient may not learn about until they see a higher-than-expected bill." Experts expect insurers eventually to lower their rates, hurting doctors and hospitals that did not adopt the software. As KFF Health News has reported, Sister Irene Kraus, once chief executive of the Daughters of Charity National Health System, was famous for coining the phrase "no margin, no mission": a hospital needs money to care for anyone. When the mission becomes the margin, the record is written for the claim, and patients can pay more for the same visit.
The scribes' defense is that the higher codes are the right ones, because the note now records work that was always done. Doctors have long under-coded their visits, those who make it say, and a scribe simply follows the rules "that much more accurately and consistently" than people did, as Trilliant's research chief put it. Its best evidence is a blinded audit at an academic orthopaedic practice, where an independent coder also rated visits documented with a scribe at higher levels than dictated ones, though overall coding accuracy was lower with the scribe, 68% against 74%. Some of the rise may well be correction. But an audit checks the note against the coding rules, not against what happened in the room, and none of these studies measured whether patients did better, the gain that would justify a bigger bill. Nor does relieving burnout, the best case for the scribe, depend on a higher code.
Discharge summaries, written for the doctors who take over a patient's care, are where the technology has done best. In a retrospective study at Stony Brook published on Oct. 2, hospitalists and primary care doctors preferred the AI's summaries of long hospital stays in 19 of 20 cases; the AI left out less, and the estimated risk of harm did not differ significantly. Omission is still the error to watch: the primary care doctors, who act on the summaries, flagged more omissions overall than the hospitalists did and judged harm more likely. In a Dutch study, blinded clinicians rated AI summaries as good overall as physicians' own but less complete, and trusted them less often; in a pilot at Stanford, hospitalists flagged omissions in 25 of the 100 drafts on which they gave feedback. Hospitals that let AI draft the summary should have the discharging physician check what it leaves out, and judge the tool by whether patients get their follow-up, not by whether doctors prefer it.
For the scribe in the clinic, the first test is the note. Before a hospital buys a scribe, and at intervals afterward, clinicians with no stake in the decision should compare a sample of its notes with the visits they record and report the error rate to the medical staff; the authors of the blinded test wrote that "independent, vendor-neutral evaluations of note quality are essential before large-scale clinical deployment." The codes deserve the same scrutiny. A medical staff that sees each doctor's mix of visit levels before and after the scribe arrived, and checks a sample of the highest against the visit as well as the chart, will know whether the scribe is correcting old under-coding or overstating complexity; physicians should sign no level they would not defend to an auditor. What matters most is what happens next. A tool whose notes get worse, or whose codes rise while the care recorded stays the same, should be fixed or switched off at once, not when the contract ends.
Medicare, which holds the claims, is best placed to tell correction from inflation. Dr. Mehmet Oz, who runs the Centers for Medicare and Medicaid Services, said in September that in the short term AI "is going to be inflationary because it's going to turbocharge the ability of the current billing systems to work more effectively." The Minnesota team's study of Medicare claims before and after scribes arrived is scheduled to be completed in December. Medicare should then find out which practices use scribes and audit samples of their high-level visits, so that correction is paid for and inflation refused, practice by practice, not through rate cuts that fall on everyone. Scribe makers, some of which advertise the ability to "maximize the number and severity of diagnoses recorded," according to Dr. Neprash, should publish independent tests of what their notes get wrong.
The software that writes the note is now reading the claim. Abridge describes its products as following a case "from bedside to bill," and that is the right order. A scribe should first prove that it writes a record the next doctor can act on; what it does for the bill should come after.
- Abridge's claims review and the University of Texas agreementOn Sept. 14, 2026, Abridge announced a pre-bill review capability for clinical documentation integrity, coding and revenue cycle teams. "Built on the same platform and clinical record already used at the point of care," it reviews inpatient claims before submission and compares "the final coded diagnoses and diagnosis related group (DRG) against what the clinical documentation actually supports"; "Abridge doesn't change documentation, codes, or claim status." The company says pre-bill review "will be the first Abridge product built for the mid-revenue cycle" and that its products work with the same record, "from bedside to bill." On Oct. 8, 2026, Abridge announced an enterprise-wide agreement with University of Texas REAL Health AI after a pilot it says grew 73% in recording clinicians and 80% in monthly notes over 12 months, and said that this year it will support more than 100 million patient-clinician conversations across more than 300 health systems. The figures are the company's own.
- Time savedRotenstein and colleagues (JAMA; Mass General Brigham release, April 1, 2026; funded in part by the Agency for Healthcare Research and Quality and the Advancing a Healthier Wisconsin Endowment) ran a cohort study with a difference-in-differences analysis at five U.S. academic health care institutions: 8,581 ambulatory clinicians, of whom 1,809 received access to AI scribes, by opt-in at four of the five sites. Adoption was associated with 16.0 fewer minutes of documentation time and 13.4 fewer minutes of EHR time per 8 scheduled patient hours, no significant change in EHR time outside work hours, and 0.49 more visits a week; the release calls the revenue gain "statistically significant, but nominal ($167 per month, per clinician adopting an AI scribe)." In a randomized trial of 238 physicians at UCLA (NEJM AI, Nov. 26, 2025), Nabla users had a 9.5% decrease in time-in-note versus control and DAX users no significant change; potential improvements in burnout need confirmation, the authors say.
- Note quality, blinded and in practiceReddy and colleagues at the Veterans Health Administration (Annals of Internal Medicine, April 17, 2026; funded by the VHA) ran a cross-sectional evaluation: 11 AI scribe tools and 18 human note takers wrote notes from audio recordings of five primary care cases performed by standardized patients, and 30 blinded raters scored them on the modified PDQI-9 (maximum score 50). Human notes scored higher in all five cases; the largest gap was in the acute low back pain case, 43.8 against 20.3. The authors note: "Cases were simulated; human-generated notes were not generated under real-world constraints." In a stepped-wedge randomized trial of 66 practitioners at the University of Wisconsin (NEJM AI, Nov. 26, 2025; funded by UW Hospitals and Clinics and the National Institutes of Health), the authors concluded: "Documentation time decreased without compromising diagnosis, billing compliance, or note quality." In the UCLA trial, physicians reported that AI notes "occasionally" contained clinically significant inaccuracies, most commonly omissions or pronoun errors.
- Trilliant's claims analysisTrilliant Health, a healthcare market intelligence company, analyzed national all-payer claims from 2018 to 2024 for six large hospitals and health systems selected because of public announcements about adopting AI scribing, "all of which have adopted ambient AI scribing in recent years." The share of high-intensity visits (99204-99205 for new patients, 99214-99215 for established patients) rose by approximately 12 to 20 percentage points for new and 7 to 12 points for established visits. The undated page gives no adoption dates and describes no comparison group; it says the trends "coincide with the 2021 revisions to E/M coding guidelines," that "the underlying causes of the changes are not clearly understood" and that the increase "is unlikely attributable to provider fraud." Asked whether scribes record more accurately or overstate complexity, its chief research officer, Allison Oakes, told Healthcare Innovation (March 30, 2026): "we don't necessarily have hard evidence at this point to say that it's one thing or the other."
- Coding and billing studiesArias and colleagues (Journal of the American Academy of Orthopaedic Surgeons, Aug. 31, 2026) compared 7,236 outpatient encounters documented by the same providers with ambient listening or dictation at one academic medical center, auditing a random 200 blind. Scribe-documented visits had higher E/M levels "as assessed by both auditors (3.24 vs 2.96) and providers (3.11 vs 2.82)"; "Overall coding accuracy was 68% for ambient listening and 74% for dictation," with 81.3% accuracy among higher-level ambient encounters. The authors conclude it was "associated with reduced undercoding." Dutta and colleagues (Annals of Emergency Medicine, June 11, 2026; 198,178 encounters at 4 hospitals) found AI scribes cut median documentation time 1.6 minutes per note and human scribes 3.3, and "Total wRVUs per shift hour did not differ among groups." A Washington University preprint (medRxiv, 2026) compared 220 primary care clinicians using Abridge with their own previous 150 days and found "a 2% increase total wRVU at day 150," gains the authors call "marginal."
- Costs, and who bears themCaroline Pearson, executive director of the Peterson Health Technology Institute, describing a roundtable it held earlier in 2026 (STAT, April 8, 2026): "The investors, the health plans, and the providers, in private, were like, 'OK, well, it's quite clear scribes are increasing coding intensity. One hundred percent.'" STAT's subhead: "Experts say insurers will eventually lower their rates, hurting providers who didn't adopt AI." A University of Minnesota release describes higher-intensity visits as ones "which come with higher copays for patients." Hannah Neprash, in an interview with the NIHCM Foundation: "Many ambient scribe vendors purport to enable higher-intensity billing through more complete documentation. Relatedly, some ambient scribes advertise the ability to maximize the number and severity of diagnoses recorded, yielding higher risk-adjusted payments for organizations participating in value-based payment." She warns of "surprise bills," which could arise "if ambient scribes convert preventive visits to problem-based visits."
- Medicare's view, and a study under wayCMS Administrator Dr. Mehmet Oz, at Oracle's annual health and life sciences summit in Orlando (Healthcare Dive, Sept. 24, 2026): "Short term, AI is going to be inflationary because it's going to turbocharge the ability of the current billing systems to work more effectively." A study led by Hannah Neprash and Paige Nong of the University of Minnesota School of Public Health, supported by the National Institute for Health Care Management Foundation, will analyze Medicare claims data "to compare clinicians at health systems before and after adoption of ambient scribe technology"; the study "is scheduled to be completed in December 2026." The foundation's grant runs from Jan. 1 to Dec. 31, 2026.
- Discharge summariesOsborne and colleagues at Stony Brook (npj Digital Medicine, Oct. 2, 2026), in a retrospective evaluation of 60 internal medicine hospitalizations lasting 7 to 21 days, found that hospitalist and primary care reviewers "preferred LLM-generated summaries for 95% of encounters," with fewer omissions and "no significant differences in estimated harm potential or likelihood"; "PCPs annotated more omissions and assigned higher likelihood of harm than hospitalists." Mehri and colleagues (eBioMedicine, May 2026; 292 paired summaries at a Dutch academic hospital, eight blinded clinicians) found lower completeness, similar correctness, greater conciseness and equal total scores; AI summaries were trusted in 249 (85.3%) cases, and physicians' by both reviewers in 279 (95.5%). In a single-unit Stanford pilot (JAMA Network Open, May 2026), the system generated 1,274 summaries for 384 discharges; physicians used AI content in 57.0% of discharges; feedback on 100 summaries noted omissions in 25; and time savings, up to 2.9 minutes, were not statistically significant.