Most pharma commercial plans still assume two channels carry the message: a rep in a chair and a clinician on a search engine. Both assumptions broke quietly, and the data is now hard to argue with. The audience you spent decades learning to reach has changed where it goes for an answer, and it went somewhere you are not measured.
This is not a forecast about 2028. It is a description of the current quarter. The demand side moved first, the supply side of approved content did not follow, and the gap between them is where your message is now being rewritten by machines you do not monitor.
The behavior change, in numbers you can check
Start with physicians, because they are the audience pharma optimizes for. In a 2024 Wolters Kluwer survey, 68 percent of U.S. physicians said they had changed their view of generative AI and now see it as beneficial in healthcare, with 40 percent ready to use it at the point of care. That was sentiment. The usage data caught up fast. The American Medical Association found physician AI adoption rose from 38 percent in 2023 to 66 percent in 2024, a 78 percent jump.
The sharpest signal is a tool built for exactly this job. OpenEvidence, a medical answer engine, reported in July 2025 that it was used by more than 40 percent of U.S. physicians and handled over 8.5 million clinical consultations per month, up from roughly 358,000 monthly consultations a year earlier. A clinician with a question about dosing, interactions, or whether a therapy fits a comorbid patient is now, in a large share of cases, typing it into an assistant. Doximity's reporting puts a finer point on direction of travel: in its physician cohorts, AI use rose from 47 percent in April 2025 to 63 percent by early 2026, with literature search the single most common use case.
The patient side moved on the same curve. Rock Health's consumer survey found AI chatbot use for health information jumped to 32 percent of U.S. adults in 2025, double the 16 percent a year earlier, with ChatGPT and Gemini the most-used tools. OpenAI, for its part, says more than 40 million people ask ChatGPT healthcare questions every day. Bain frames the commercial consequence plainly: more than half of patients say AI tools are valuable for treatment information, and nearly one-third of HCPs say they use AI frequently for it, often through AI-generated search results they did not deliberately seek.
Read those together and the conclusion is not subtle. The question your brand exists to answer is now being asked of a machine, by both audiences, at scale, today.
Why the approved message gets stranded
Here is the mechanism that should worry a brand lead. Every dollar of MLR review, every reviewed claim, every fair-balance line and ISI was produced for a delivery system that is losing share: a rep visit and a clicked link to your page. The approved sentence was engineered to survive scrutiny in those channels. It was not engineered to be retrieved, ranked, and paraphrased by a model that the clinician reaches instead of you.
When a physician asks an answer engine about a therapy, the model does not hand them your reviewed page. It reads many sources, your label among them if you are lucky, and writes a fresh paragraph. Three failure modes follow, and none requires bad intent from the model.
The first is omission. If your approved content is not structured, retrievable, and clearly attributable, the model leans on whatever is. A patient forum, an old press summary, a competitor comparison, a superseded abstract. Your careful language never enters the answer because the machine never reached for it.
The second is drift. The model rephrases for fluency, and fluency is in direct tension with the exact wording fidelity depends on. A reviewed indication becomes a looser one. A hedged efficacy claim loses its hedge. A "commonly discussed" off-label use, lifted from ambient text, gets attached to your name in a confident aside.
The third is silent loss of fair balance. Safety language is the easiest thing for a paraphrase to drop, because the source the model summarized may have led with efficacy. The contraindication that your ISI guards so carefully simply is not in the answer the clinician reads.
The rep could carry the whole reviewed message and read the room. The model carries a paraphrase and reads a probability distribution. You spent the budget producing the first and your audience is now consuming the second.
What this looks like on one brand
Take a fictional brand, Varigel, approved for a single narrow indication with a known contraindication in patients on a common comorbidity medication. The label is precise, the MLR file is clean, the rep deck is current. None of that touches what happens next.
A clinician opens a medical answer engine and asks whether Varigel is appropriate for a patient who is also on the contraindicated comorbidity drug. That is a real, high-stakes prompt, the kind that 8.5 million monthly consultations are made of. The engine assembles an answer from its sources. In one run it cites the label and flags the contraindication cleanly. In another it describes the approved indication but omits the contraindication, because the efficacy summary it leaned on skipped the safety section. In a third it volunteers a second, off-label use it found in an old abstract the brand never promoted.
Now layer the patient. The same week, that clinician's patient asks ChatGPT what Varigel is for and whether it is safe with their other medication. The patient reads a paraphrase too, generated from a different source mix, and forms an expectation before the appointment.
Every one of these answers is being produced right now, on engines your audiences already use, and your brand team can see none of it. The rep call report tells you what the rep said. Nothing tells you what the machine said to the ninety clinicians the rep never reached this month.
What a brand should instrument
You cannot govern a channel you do not measure. Treat the machine's answer as a surface with a metric, and put three things in place.
1. Baseline Share of Answer on the prompts your audience actually types. Not your campaign headlines. The real questions: indication fit, dosing, interactions, comparison to standard of care, safety in a comorbid patient. Run them across the engines your audience uses, ChatGPT, Gemini, Perplexity, Google AI Overviews, Claude, and the medical answer engines in your therapeutic area. Record three numbers per prompt. Are you mentioned. Is the mention accurate against the label. Does anything read off-label or drop required safety language. That triple is your baseline, and almost no brand has taken it.
2. Give the machine a clean source to prefer. Models reward content that is structured, attributable, and consistent. Your approved indication, your safety language, your evidence, expressed in plain, well-marked, machine-legible form, becomes the easiest correct answer for a model to lift over the ambient noise. This is the part of the work that rhymes with SEO, but the payload is the reviewed sentence, not a keyword. Which means the inside of the house has to be in order before you feed the outside.
3. Monitor continuously and close the loop to MLR. A baseline taken once is a screenshot of a river. The answer moves when a model updates, when a competitor publishes, when a new abstract surfaces. The brands that hold their message will detect a new off-label drift the week it appears, trace it to the source the model used, and route it to the team that can correct the underlying content. That requires the outside signal and the inside review system to be one loop, not two tools owned by two teams.
The takeaway
The channel shift already happened. Physicians adopted AI tools faster than pharma adopted any prior technology, patients doubled their use of AI for health questions in a year, and the reviewed message that took months to clear is now competing with a paraphrase the brand never sees. Saying nothing is not neutral. It is a decision to let the internet write your label summary for the two audiences that matter most.
Juncture is built for the seam between the message you approve and the message the machine repeats. Inside, it pre-checks an asset against the label before MLR and gives the reviewer a 21 CFR Part 11 trail and an e-signature sign-off. Outside, its Answer Monitor measures Share of Answer and off-label drift across the engines your HCPs and patients actually use, continuously, so a deviation shows up as a deviation from a known-good source rather than as a forwarded screenshot weeks later. The same approved sentence you clear on the inside is the one Answer Monitor watches for on the outside.
Bring one brand and the twenty prompts your audience actually asks. We will show you its Share of Answer today and walk the line from any drift back to the approved sentence that should have been there.
Sources
- Wolters Kluwer, "Over two-thirds of U.S. physicians have changed their mind, now viewing GenAI as beneficial in healthcare," 2024. wolterskluwer.com
- American Medical Association, "2 in 3 physicians are using health AI, up 78% from 2023," 2025. ama-assn.org
- OpenEvidence, "The fastest-growing application for physicians in history announces $210 million round," via PR Newswire, July 2025. prnewswire.com
- Doximity, "2026 State of AI in Medicine Report," 2026. doximity.com
- Fierce Healthcare on Rock Health Consumer Adoption of Digital Health survey, "AI chatbot use for health information up from 16% in 2024," 2026. fiercehealthcare.com
- Healthcare Dive, "40M users turn to ChatGPT daily for health questions, OpenAI says," 2025. healthcaredive.com
- Bain & Company, "Pharma Commercialization in the Age of AI and Active Patients," 2025. bain.com