The memory clinic appointment has always had an awkward arithmetic to it. Three people talk, one of them does most of the talking, and only a fraction of what is said survives into the medical record. The daughter describes the night her father tried to drive to a house he sold in 1998. The physician, typing or scrawling or trying to hold it in his head until the end of the day, keeps what he can. What reaches the next clinician is not the conversation. It is the residue of the conversation. Into that gap, quietly and fast, has arrived a phone on the desk that listens to the whole thing and drafts the note itself.
A prospective implementation study published in JMIR Medical Informatics on June 23, 2026, by Cameron Harvey and colleagues, reports what happened when Hawaii Pacific Health Medical Group, which spans four hospitals and 70 clinics across the islands, ran one of these tools in ordinary practice. The platform was Abridge, integrated with the Epic record. Eighty providers were recruited for a three-month pilot from December 2024 through February 2025; 79 finished it, generating more than 25,000 notes across 23 specialties. Among heavy users, defined as those using the scribe in at least 60 percent of encounters, the record system's own metrics showed average time in notes per appointment falling from 7.14 minutes to 5.63, a 21 percent drop (P=.005), and time in notes per day falling 21 percent (P<.001). Self-reported after-hours burden moved too: 6 of 79 providers (7.6 percent) said they spent eight or more hours a week on notes outside clinic, down from 25 (31.6 percent) before the pilot.
The interesting parts are the ones that did not move, and the paper reports them honestly. So-called pajama time, the minutes logged in the record after hours, fell by 3.6 minutes on average, a 13 percent change that did not reach statistical significance (P=.19). Symptoms of burnout showed no significant difference overall, even though 49.4 percent of physicians reported none during the pilot. Patient satisfaction scores for whether the provider listened were essentially unchanged (P=.39). And the notes got longer, not shorter: progress notes grew about 7 percent, from 5741 to 6121 characters (P=.001). Less time, more text. Whether more text means a better record of a complicated visit is a different question, and this study did not answer it. Note quality here was self-reported, with providers saying that 84 percent of notes needed edits to less than a quarter of the content.
Someone else has been counting the errors. In the same journal on April 17, 2026, Sandra Taylor and colleagues at University of California, Davis Health reported a pragmatic pilot in which 31 physicians used an ambient listening scribe for 7545 clinic notes over two months in 2024, then formally reviewed a sample of 356 of them. Accidental omissions were the most common fault, in 64 notes (18 percent). Hallucinations, meaning content the AI produced that was not in the conversation, appeared in 41 notes (11.5 percent). Accidental inclusions appeared in 33 (9.3 percent), and bias in the drafted language was rare, in 4 (1.1 percent). Most errors were graded mild to moderate. But 19 notes (5.3 percent) contained something the reviewing physicians judged capable of serious or imminent harm if it went uncorrected. The authors' conclusion is the one that matters at the desk: the drafts were generally good, and careful clinician review remains essential.
For a family, the practical shape of this is simple enough to hold. Omission is the dominant failure mode, and a dementia visit is built almost entirely out of the things that are easy to omit: a change in sleep, a new word-finding slip, the fall nobody reported, the dose that gets taken twice some evenings. The scribe does not decide anything. NHS England, in guidance published on April 27, 2025 for executives adopting these products, frames them as documentation and workflow support that still sits inside a clinician's accountability and a safety and regulatory process, not outside it. The note remains a draft until a human signs it, which means the person who can catch a dropped detail is the clinician, and the person best placed to notice the detail was dropped is whoever said it in the first place.
None of this is a reason for alarm, and the evidence does not support one. It is a reason to know what is in the room. If a device is listening, families can ask what it is, whether they can decline it, and what the note ends up saying. That is a conversation with a clinic, not a medical decision, and it is the one lever a caregiver actually holds over a piece of infrastructure that is already in the building.