Insights
AI has reached 16.3% of the world. We work for the other 84%.
Sixteen point three percent. That is the share of the world's working-age adults who had used a generative AI tool by the close of last year, and we keep the number written down where we can see it. Six months earlier it was 15.1%. Say it the way you would say it to a friend across a table: roughly one in six people alive has ever touched the thing the rest of us keep calling world-changing. We don't find that discouraging. We find it clarifying. It means the room where AI is being decided is a small room, and most of the people we love in this profession are not in it — the practitioners who chose healthcare to sit close to other human beings, not to be early adopters of anything. A world should not be run by one in six. So the 84% is not a market to us. It is the assignment.
Six months of light years, and one point of movement
Here is the arithmetic that keeps us honest. Microsoft's AI Economy Institute, which measures adoption as the share of people aged fifteen to sixty-four who have used a generative AI product, put the world at 15.1% in the first half of 2025 and 16.3% in the second — a gain of 1.2 points across six whole months. The following quarter added 1.5 more, reaching 17.8%. Round it generously and you get somewhere under three points of human uptake in a year. Now set that beside what the technology did in the same stretch. Stanford's 2026 AI Index records models gaining thirty percentage points in a single year on Humanity's Last Exam — a test built on purpose to be brutal for machines and generous to human experts. On a benchmark of real software engineering work, models climbed from roughly 60% of the human baseline to nearly all of it inside twelve months. The cost of running a model at the level that stunned everyone in 2022 fell about 280-fold. Light years, on any honest reading. And the number of people coming along moved by about a point. That gap is the whole story, and notice what it is not: it is not a technology story. Capability is not the scarce thing anymore. A way in is.
The 84% are not standing out there unconvinced
It is tempting, from inside the small room, to assume the other five-sixths are holdouts — that they've heard the case and said no. The data does not describe holdouts. It describes people who were never actually reached. Adoption in the Global North now sits at 27.5% while the Global South sits at 15.4%, and the distance between them widened rather than closed, from 9.8 points to 10.6 in half a year. Diffusion is not tracking curiosity. It is tracking who already had the electricity, the bandwidth, the device, the spare evening, and the language. That last one deserves a pause, because it is the most hopeful thing in the whole report. When the tools got meaningfully better in Asian languages, adoption in South Korea, Thailand, and Japan moved faster than almost anywhere else on earth. Those people were not finally persuaded. They were finally addressed. Something in them had been waiting the entire time for the technology to speak the way they speak. We see the same shape at a much smaller scale, in ordinary practices, all the time. The practitioner who has not adopted AI is rarely a skeptic. Far more often she is someone who ran an eleven-hour day, drove home with charting still open on her laptop, and did not have a free Tuesday night to audition software written by people who have never stood in a treatment room. She is not behind. She is unaddressed. Those are completely different problems, and only one of them is her fault — which is to say, neither.
From the field
We sat with a practitioner this spring who apologized to us. That is the part we cannot shake — she apologized. She said she felt behind, that everyone at the last conference seemed to be running something she had never heard of, and that she had tried one tool and quietly given up because she couldn't tell whether it was safe to put a patient's name into it. Nobody had ever told her the answer to that question. So we spent the first hour not touching software at all. We just wrote down, in her handwriting, where her days were actually leaking — and documentation came out fourth on the list. The phone was first. She has two small, unglamorous things running now, both of which she chose, neither of which she needs us for anymore. Nothing about that hour was technical. We think about her every time someone quotes an adoption statistic as though it were a scoreboard, with winners on one side and laggards on the other. She was never a laggard. She was somebody nobody had sat down with.
44%
The share of medical group practices already using AI in patient visits whose leaders say it has not reduced staff workload — against 39% who say it has, and 17% who aren't sure (MGMA Stat poll of 244 practice leaders, August 2025). Hold it loosely: it is small, self-selected, taken at one moment, and "reduced workload" is felt as much as it is measured. But one detail inside it is hard to unsee — the practices using AI in a quarter or fewer of their visits were the least likely to report any relief at all.
In healthcare, the gap has already moved indoors
Medicine is well past 16.3%, and that is worth saying plainly before anything else. The AMA's 2026 survey of nearly 1,700 physicians found 81% using AI professionally — roughly double the share in 2023 — and 76% saying it gives them an advantage in caring for patients, up from 65%. Our profession is not the laggard here; it is running ahead of the world. But read the same survey one line further and you find people carrying real weight while they run: 88% are worried about losing skills they spent a career building, 88% want genuine validation of safety and efficacy before they lean on a tool, 86% want straight answers about privacy. And then hold that against what practices report on the ground — 71% using AI somewhere in the patient visit, and among them, more leaders saying the workload has not moved than saying it has. So in our field the divide is no longer only between the practices that have AI and the practices that don't. It has moved inside the building. There is a second 84% now: people who bought the thing, never got the two coached weeks, never named who owns it when the audio drops, and now pay a subscription for a tool they half-trust and half-use. Half-use is the most expensive place in this entire category to stand. You carry all of the cost, all of the vigilance, and almost none of the relief — and worst of all, you start to believe the disappointment is about you.
This is the work we chose, and it is slower than it sounds
Vision Optimized™ exists for both of those groups, and the method is the same for each. We start with the day, never with the product. Before a tool is named out loud we want to know what your Tuesday actually looks like — where the hour disappears, which task everyone quietly dreads, what the team has stopped complaining about because they assume it is permanent. Then we pick one bottleneck, the true one, and we are unsentimental about it. If the phone line is what is bleeding your practice, no documentation tool on earth will touch that, and we will say so before you spend a dollar. We insist on the coached weeks, because the practitioners who abandon these tools are almost always the ones who were handed a login instead of a hand. We insist that a human being's name goes next to the workflow, so that when something fails at 4:50 on a Friday it does not land on whoever is least able to say no. We teach it in your language, in your rooms, with your words — the same lesson those diffusion numbers taught at the scale of nations. And when the honest answer is not yet, we say not yet, and we mean it warmly. Prices in this category are falling, the tools are getting better every quarter, and waiting six months with your eyes open is a legitimate strategy. We would rather be the people who told you the truth than the people who sold you a subscription. That is the whole business model, and it is not a complicated one.
A world running on one in six is not a world that has been changed. It is a world that has been sorted — and quietly, without anyone deciding it, sorted along the same lines everything else has always been sorted along. We do not accept that, and we do not think you have to either. The correction is not more capability; we have more capability than we know what to do with. The correction is patience, in rooms, with individual people, one practice at a time. So we would rather help one practitioner who has never typed a prompt get a real hour back on a Thursday than optimize a system that is already humming. Because here is what that hour becomes, and it never appears on any invoice or in any adoption statistic: a practitioner who finishes the visit facing the person instead of the screen, whose hands are still, who hears the sentence that was not on the intake form — the one that begins, quietly, with "there's one more thing." That sentence is the entire practice of medicine. It is what everyone in that small room is ultimately building toward, whether they know it or not. Let's make sure it reaches the other 84%.