Three things landed in the same seven days, and together they describe a handoff that has already happened. A survey of small practices found most of them running AI with nothing written down about it, and found that where anybody owns the question it is usually the owner. Pew found that most people want to be told when AI touches their care, and that nearly half cannot tell whether it has. ECRI, the nonprofit that has run a confidential problem reporting line since 1972, opened that line to AI errors and said plainly that nobody is counting them yet. None of it arrives as a rule. What arrives is the responsibility, and it lands on the smallest desk in the building.
Signal 1
Most Small Practices Have No Written AI Policy, and Half of Them Put It on the Owner
On August 24, 2026 Healthcare IT News reported results from the 2026 Healthcare Pulse Survey, a poll of 285 people working in dental, specialty medical, optometry and veterinary practices. It was run by Weave, a company that sells AI-enabled communication tools to those same practices, which is worth holding in mind the whole way through. Forty-one percent said their practice has adopted some AI workflow or that AI is already changing how they operate, and another 33 percent said they have looked at AI tools without committing to any. Then come the two numbers that matter most here. Fifty-eight percent have no formal AI governance structure of any kind, and 48 percent leave responsibility for AI strategy with the practice owner. On risk, 60 percent named HIPAA and data privacy as a leading concern, 51 percent said they worry AI will produce low-quality work, and 35 percent named regulatory uncertainty. The rest of the survey explains why the tools get bought at all. Nearly two thirds spend at least an hour a day on manual data entry and almost a third spend two hours or more, 80 percent said duplicate entry affects their workflow, and 7 in 10 said their software does not talk to itself. The work they most want automated is scheduling and confirmations, insurance verification, intake and forms, and billing reminders. Nearly two thirds had a staffing shortage in the previous twelve months, and close to a quarter said a shortage forced them to cut back on appointments.
Sources: Weave 2026 Healthcare Pulse Survey | Healthcare IT News, August 24, 2026
What this means for you
Read 58 and 48 together, because they are one fact said twice. Most of these practices have nothing written down about AI, and in about half of them the person responsible is the person who owns the place. That is not a committee and not a compliance officer. That is one person, usually the same one who answers the phone. Issue 022 covered Epic handing its customers a way to build AI agents and telling them to bring their own policies, and made the point that whoever puts the tool in place owns every decision about what it may do. This survey is that same handoff measured at the small end, where there is nobody left to hand it to. Two honest caveats before you take the numbers to heart. Weave sells AI tools to practices like these, so a survey showing unmet need is a survey that helps Weave, and you should read it that way. And the sample is dental, specialty medical, optometry and veterinary, with no health coaches in it at all, which tells you something on its own about who gets counted. Read it anyway, because it is the closest thing published this year to a picture of your end of the field. Then notice what nobody in that 58 percent is worried about. Not one of them is being audited. Nobody auditing you is not the same as nobody asking you. In a practice of one the asking arrives as your client, halfway through a session, wanting to know where the summary you just read back to her came from. You will answer that out loud, in the moment, because you never decided it beforehand. She will remember what you said, and you do not get to go back and decide it first.
Signal 2
Pew Found That Nearly Half of Americans Do Not Know Whether AI Has Been Used in Their Own Care
On August 25, 2026 the Pew Research Center published a survey of 3,488 US adults, conducted June 22 to 28 through its American Trends Panel. Seventy-two percent said it is extremely or very important that a doctor or other healthcare provider tells them when AI is being used in their healthcare, and that majority holds across age, sex, education, race and ethnicity. Asked about specific uses, eight in ten or more want to be told when AI analyzes a medical scan (81 percent), makes a diagnosis (81 percent) or explains lab results (80 percent). Seventy-two percent want to be told when it takes notes during an appointment. Even for back-office work the numbers stay above half, at 64 percent for ordering prescription refills and 56 percent for scheduling. Then the survey turns to control and the picture changes. Fifty-three percent say they have not too much say or no say at all over whether a provider uses AI in their care. Sixty-three percent would like more say, against 21 percent who are comfortable with the say they have. The sharpest finding is what people do not know. Sixteen percent say AI has been used in their care, 37 percent say it has not, and 46 percent say they are not sure. That uncertainty is flat across education, running from 46 percent among adults with a high school diploma or less to 48 percent among those with a postgraduate degree. Among the small group who know AI was used, 32 percent say they do not understand well how it was used, against 22 percent who say they do.
Sources: Pew Research Center, August 25, 2026 | Fierce Healthcare, August 26, 2026
What this means for you
Neither of these numbers means much on its own, so put them side by side. Almost three quarters want to be told when AI touches their care. Nearly half cannot tell whether there was anything to be told about. Between those two sits a duty nobody has been assigned. Issue 021 covered a poll finding that one adult in five keeps something back from a doctor out of shame or mistrust, and argued that the withheld question goes somewhere else instead. This is the same silence running the other direction. She is not being told, and she has stopped expecting to be. Here is what that looks like in a practice of one. Your client has a note-taker running in her physiotherapist's room, a summary in her portal that software drafted, and an assistant that answered her email at nine on a Sunday night, and she is not certain that any of it happened. She is certain about you, because you are the one she is in a room with. Which means the plainest account of AI she receives in a whole year of care is quite likely to be yours. That is a strange amount of weight to rest on a health coach, and it is where the numbers land anyway. Then there is the other half of it, the 32 percent who were told and still did not follow what they were told. Telling and explaining are two different jobs. The one she needs from you is the second one, and it takes about a sentence.
Signal 3
ECRI Opened Its Problem Reporting Line to AI Errors and Said Nobody Is Counting Them
On August 25, 2026 ECRI, the nonprofit that has run a free confidential problem reporting line for medical devices and technology since 1972, expanded that line to cover errors, malfunctions and near misses involving AI tools and AI-enabled devices used in care. It is asking providers, health systems and clinicians nationwide to send in any incident where an AI tool may have contributed to an error or introduced risk into how care is delivered. Every submission is triaged and investigated by ECRI's clinical and engineering staff, and ECRI writes back to the person who sent it with what was found. The reason given is a gap rather than an incident. Scott Lucas, ECRI's vice president of devices, therapeutics and technology, said there is no clear picture yet of what AI does downstream in health work, and that without a reporting dataset and analysis the industry cannot improve how these tools are designed and fitted into care. In ECRI's own survey of 124 respondents, most of them quality, safety, risk or compliance leaders, 31 percent had run into an AI output they believed was incorrect or misleading over the past year, 34 percent had not, and 35 percent were unsure. Nine percent reported that an AI error reached a patient or changed a care decision. The tool people encountered most was the ambient scribe, at 37 percent, ahead of decision support built into the health record and clinical chatbots, both at 31 percent.
Sources: ECRI news release, August 25, 2026 | Fierce Healthcare, August 28, 2026
What this means for you
Notice who ECRI asked before you weigh what they said, because the whole announcement rests on one answer from that group. Thirty-five percent were unsure. These are quality and safety leaders, people whose actual job is knowing what went wrong, and a third of them cannot say whether an AI in their own organization produced a bad output this year. Nine percent say one reached somebody receiving care. Nobody is claiming that is the real rate. It is the rate that is visible from inside organizations that already employ a person whose job is to look. Issue 020 covered a health system that passed the first AI certification in three weeks because it had four years of records to show, and made the point that what gets graded is the record you kept. This is the same idea one level up. What is being assembled here is the national record of what AI gets wrong in health work, and whoever files reports is who that record will describe. Notice which tool topped their list. The ambient scribe, at 37 percent, which is the AI most likely to be switched on in the room when your client finally says the difficult thing. So if an AI ever hands you something about a client that is confidently wrong, the visit summary that says improving in the week she told you she was worse, the supplement interaction that does not exist, that report is worth sending. ECRI asks healthcare providers, systems and professionals, which is a wider door than most reporting channels have ever opened to a coach, and it costs nothing.
The Pattern
Three findings, seven days, and they stack. Most small practices have nothing written down about AI and about half put the question on the owner. Most people want to be told when AI touches their care and nearly half cannot tell whether it has. And the one organization now trying to count what AI gets wrong has to do it by asking people to volunteer what they saw. Issue 022 argued that the AI your client uses is the one nobody is testing. This is the other half of that. The responsibility arrived before the instructions did, and it did not stop at the hospital door. None of the three was written with a coach in the room. Read them in order and the shape is plain enough: nobody wrote you a rule, your clients assume somebody did, and the record of what goes wrong is being built out of whatever gets reported. None of it arrives with a deadline. It arrives as a tool already running in your practice that you did not choose, and a written record that does not exist yet.
One Thing You Can Do This Week
Make the list of every AI that touches a client's information in your practice, then say one line of it out loud to a client this week. You are almost certainly in the 58 percent, and the list takes about fifteen minutes. Go through what you actually pay for or log into: your scheduler, your booking and questionnaire software, your email platform, your note-taking app, the telehealth room, whatever writes your session summaries, and anything with an assistant built into it that you did not choose and never turned off. For each one write the tool, what it sees, and whether anybody other than you ever reads what it produces. Some lines will be blank because you do not know, and the blank lines are the useful part, because 46 percent of the people you work with are sitting in the same blank. Then take one line and put it in a sentence a client would understand. Something close to this. I use a tool that turns our session into notes, I read and correct them before they are saved, and nothing goes anywhere else. Say it at the start of a session rather than burying it in paperwork. Most of the time nothing will happen, and that is the point, because the question stops being open. You can be the one practitioner in somebody's year who did not make her ask.