Eating a five-day Fasting-Mimicking Diet (FMD) several times a year may help extend longevity and reverse signs of disease, Valter Longo recently told an audience at Harvard T.H. Chan School of Public Health. This low-calorie plant-based nutritional program designed by Longo and colleagues tricks the body into believing it is fasting—providing the health benefits without the potential health problems fasting can cause for some.

Longo, Edna Jones Professor in Biological Sciences and Gerontology at the University of Southern California Leonard Davis School of Gerontology, delivered the 21st annual Stare-Hegsted Lecture on Nov. 9 in Kresge G3 and on Zoom. Sponsored by the Department of Nutrition, the annual event honors department founders Fredrick Stare and D. Mark Hegsted.

The FMD—which is low-sugar, low-protein, and high in healthy fats—came out of the Longo lab’s discoveries about nutrient-sensing pathways and how those pathways influence aging and age-related diseases. It’s now available in commercial kits which include pre-portioned soups, drinks, supplements, and snacks. (Longo consults for the company that makes them but donates his fees to charity.) The diet has been carefully formulated based on extensive research and would be difficult for individuals to safely replicate on their own, according to the company. Longo also wrote a book on his longevity diet aimed at the general public.

During his lecture, Longo explained that the FMD triggers stem cell activation, cellular reprogramming, and autophagy (a sort of cellular recycling system). In over 40 clinical trials, his team has investigated the potential beneficial effects of five-day FMD cycles on cardiovascular health, biological aging, and a range of age-related diseases.

The findings Longo shared included:

• Reduced body weight and abdominal fat without affecting lean muscle mass
• Restored normal glucose levels in pre-diabetic patients
• Reduced biological age by around 2.5 years based on blood markers
• Enhanced chemotherapy and hormone therapy effectiveness in cancer patients
• Reduced Crohn’s disease markers and chronic kidney disease damage markers

Longo stressed that FMD is a complement to existing therapies, not a replacement. He is continuing to explore its potential as a preventive intervention, for example in people with a genetically high risk of Alzheimer’s disease and hopes to see it become more integrated in clinical practice as part of doctors’ therapeutic toolkits.

---

It’s increasingly hard for patients to find a new primary care physician (PCP)—even in Massachusetts, which has more PCPs per 100,000 people than any other state, according to Harvard T.H. Chan School of Public Health’s Jose Figueroa. He was interviewed about the PCP shortage on the Nov. 5 episode of WBZ News Radio’s NightSide with Dan Rea.

Figueroa noted that PCPs are generally older than other clinicians and that not enough new doctors are in the pipeline to replace them. In 2025, only 18% of graduating medical students said they would choose primary care. Instead, they are turning to specialties such as oncology, gastroenterology, and anesthesiology, which are considered more prestigious and pay two to three times more than primary care.

Being a PCP also requires a much broader scope of knowledge to oversee the care of patients who are increasingly older and sicker and may be seeing multiple specialists for various conditions, Figueroa said. The job also comes with a heavy paperwork burden including insurance company prior authorization forms, and doctors may feel that electronic medical records systems create more work instead of saving time.

These factors are leading to burnout and PCPs reducing their hours—by 25% in Massachusetts, Figueroa said.

He noted that loan forgiveness programs for medical school graduates, who may carry debts in the hundreds of thousands, may help entice more doctors into primary care. Paying PCPs more would probably be even more effective, he said.

Listen to NightSide with Dan Rea interview: Help! I Need a PCP!

Learn more

Why finding a primary care provider can be so tough (Harvard Chan School news)

---

Heather Baer is an associate professor in the Department of Epidemiology at Harvard T.H. Chan School of Public Health. As a researcher, she focuses on the role of modifiable risk factors in chronic disease and strategies for addressing overweight and obesity in primary care. She’s also an award-winning teacher and an educational leader at the School, where she serves as faculty director of the Master of Public Health (MPH) in Epidemiology program and field of study leader for the MPH in Clinical Effectiveness program and teaches courses (both on-campus and online) in epidemiologic methods

Baer recently spoke about why she loves teaching at Harvard Chan School, and what surprises students about its faculty.

Teaching takes a lot of time and effort, but it’s one of the parts of my work I enjoy the most. I like thinking about how to explain something clearly and helping students make sense of a challenging concept.

I’m grateful for the teaching awards I have gotten, and I especially appreciate personal emails and thank-you notes that I receive from students, even many years later. It means a lot to hear that something helped them.

The most rewarding part of teaching is the students I get to work with. They’re thoughtful and kind, and I really appreciate their genuine interest in learning and putting what they learn into practice.

I teach epidemiology, which is about methods. These methods are the tools that we use to design, conduct, and interpret studies of health and disease, including studies of what causes disease and how well different treatments work. The majority of the students I teach in the MPH in Clinical Effectiveness are physicians, and they come from different specialty areas and have very specific research interests. Sometimes, a student will ask a question and start to get into fine details about their work. It’s fun to hear about different clinical contexts. I didn’t go to medical school myself, but I’ve learned a little about a lot of different aspects of medicine from my students, and I try to connect those specific questions back to the methods we’re learning.

Students in the hybrid MPH in Epidemiology program come from a wider range of professional backgrounds. Many are physicians, but others work in the business or tech side of health care or come from a basic science background. Again, the methods are the common denominator. Epidemiology is so logical. You can apply it to any question. It doesn’t have to be a clinical question. It can be a question in your daily life.

The first class that I taught as a faculty member was Introduction to Clinical Epidemiology [Epi 208] in 2009, which I still teach. I had done some teaching as a graduate student at the School, but this was the first time I had to give a lecture in front of 175 students. At the time, I wasn’t much older than many of the students, so that was a bit intimidating. I remember being very nervous, and I totally overprepared. I spent so much time on my slides. But one thing that helped me was to think about the element of teaching that is performance—keeping people engaged. I’d been involved with theater and music in high school and college, so that gave me something familiar to draw on.

It was a lucky situation for me because I was able to start with just a few lectures and build up my confidence. Then gradually, I took on more from the senior faculty involved in the course.

One of them was Fran Cook. He was my very first professor when I started as a master’s student at the School in 1999, and my first epidemiology professor. Later when I became faculty, he was my mentor and colleague. He’s definitely the one I’d say inspired me the most with teaching, and whom I have tried to model myself after.

Fran had so much knowledge about the content he was teaching, and he had a real gift for  taking complex concepts and breaking them down into manageable-sized pieces. And he always wanted to make learning fun. He was also so warm and welcoming. Fran is retired now, but I still talk to him and consider him to be a wonderful mentor.

I find that if I’m enthusiastic and excited about what I am teaching, that can be contagious. Even when the material is challenging, that enthusiasm can help students get interested and stay engaged. Another element that I think is very important is to be very organized about the concepts you’re conveying and in how the course is run. And like Fran, I always think about being kind to students. If you’re kind and warm, students feel comfortable asking quetions, and that helps them learn.

I try to keep things interactive in class. For example, we use software called Poll Everywhere that lets you ask questions of your class, and you can see their responses in real time. It’s nice because it forces a pause. If you’re lecturing, you can just keep talking, but I’d rather check in to make sure students are following what I’m saying. 

I think students value having accessible faculty. I’ve heard students say that they initially felt intimidated about approaching Harvard faculty. I’m glad when they discover they can ask questions and reach out for help.

---

Predicting survival among early-stage non-small cell lung cancer (NSCLC) may be improved by considering an individual’s “epigenetic age acceleration” (EAA)—whether their estimated biological age is older than their actual age—alongside traditional risk factors such as smoking history, according to a new study from Harvard T.H. Chan School of Public Health.

The study was published Oct. 16 in iScience. David Christiani, Elkan Blout Professor of Environmental Studies, was co-corresponding author. Other Harvard Chan co-authors included Mulong Du, department associate in the Department of Environmental Health, and Li Su, research specialist and director of the Christiani Lab.

Lung cancer, particularly NSCLC, is the leading cause of cancer-related deaths in the U.S. Survival rates for early-stage NSCLC are significantly lower than those of other common cancers, such as breast and prostate cancer. The new study looked at whether epigenetic age—a biomarker that reflects cumulative molecular changes influenced by genetic, environmental, and lifestyle factors—could be a useful tool in helping improve prognosis prediction in early-stage NSCLC and lead to more informed treatment decisions. Epigenetic age is estimated through data on DNA methylation, a type of DNA modification that changes how a gene behaves without changing its underlying sequence.

For the study, researchers analyzed blood DNA methylation from 75 participants with early-stage NSCLC enrolled in the Christiani Lab’s Boston Lung Cancer Study. The team estimated EAA among the participants using several different types of “epigenetic clocks”—tools that have been developed over the past decade or so to track DNA methylation. During the follow-up period, 63 of the patients died. Researchers found that, across all the clocks, EAA significantly predicted survival in early-stage NSCLC, with one clock—called GrimAge2—predicting with the highest level of accuracy. The co-authors noted that the clocks were more accurate at predicting survival than models based solely on factors such as participants’ smoking history or chronological age.

According to the co-authors, “these findings highlight the need for personalized prognostic models in NSCLC that account for both smoking history and epigenetic aging dynamics.”

Read the study:

Epigenetic age acceleration enhances prediction of overall survival in early-stage non-small cell lung cancer

---

---

---

---

Every six months, a new cohort of clinicians, public health practitioners and health-system executives arrives in our classrooms at Harvard Chan, and the questions they bring with them act as a reliable barometer of where artificial intelligence actually stands. Three years ago, most conversations centered on what generative AI was and whether a large language model could draft a passable clinical note or summarize a research paper. Eighteen months ago, attention shifted to prompt engineering and AI scribes. Today, leaders are grappling with a qualitatively different development: software that does not merely answer a question in a chat window, but takes a delegated goal, uses software tools across clinical and administrative systems, inspects what happens and decides what to do next.

That shift—from single-turn conversation to multi-step action—marks the arrival of agentic AI in everyday institutional life. Yet as the word “agent” has migrated from computer science papers into vendor pitch decks and hospital boardrooms, it has accumulated a thick fog of hype and imprecision. Products that run a rigid five-step script are marketed as autonomous agents, while genuinely capable systems fail in clinic because no one designed the surrounding workflow, permissions or human oversight. Understanding what agents actually are, why they have become practical right now and what it takes to run them dependably is the subject of a discipline we call agentic operations.

What an AI agent actually is

To answer this question, it helps to first separate predictive AI and generative AI from agentic systems. A predictive model takes structured inputs—vital signs, laboratory trends, prior admissions—and estimates a single probability, such as thirty-day readmission risk. A conversational chatbot takes a prompt and generates a response: a translated discharge instruction, a plain-language summary or a piece of computer code. Even when that response is remarkably fluent, the interaction begins and ends with generating text.

An agent, by contrast, is a system that pursues a (for now, delegated) goal across multiple steps, calls external tools, observes the results and adapts its subsequent actions. Imagine an outpatient specialty referral that has stalled. An AI model asked to draft a polite follow-up letter is generating text. A system that opens the referral record, notices that an outside pathology report is missing, checks whether the report arrived overnight in a document queue, routes a targeted request if it is still absent and logs its progress for a clinic coordinator is operating inside an action-and-feedback loop. As we note in our glossary guide to reading the names of models, harnesses, applications and providers, the underlying AI model may be identical in both cases. What turns a model into an agent is what it does and the software architecture wrapped around the model that enables agentic behavior.

In everyday organizational work, nobody is retraining a model’s internal parameters. Instead, the typically off-the-shelf model supplies interpretation, planning and proposed actions, while a surrounding AI harness supplies execution and control. The harness assembles the right information into the model’s context window, exposes approved capabilities through function calling and APIs, tracks current task state, carries relevant lessons across sessions through AI memory and enforces strict permissions and stopping rules. Autonomy is never binary: an agent can be given wide latitude to search internal guidance while holding zero authority to alter a clinical record or contact a patient without human approval.

Why agents have become relevant now

Why has this architecture moved from research laboratories into practical operations over the past eighteen months? Three developments converged at once. First, newer frontier models and reasoning models became substantially better at sustaining multi-step plans, inspecting error messages when a tool call fails and revising their approach rather than immediately losing the thread. Second, software engineering matured around the model: standardized protocols for connecting data sources, reusable agent skills, better context management and compaction and battle-tested harnesses—first proven in coding environments such as Claude Code, Codex, Goose and Antigravity—gave non-software industries a concrete blueprint for bounded multi-step work. Third, the cost of AI inference fell far enough that running ten or twenty reasoning and verification steps on a routine administrative case became economically plausible.

Those technical shifts arrived at a moment when health care badly needs operational relief, yet cannot afford careless experimentation. Clinicians are burning out on administrative friction; patients wait weeks while prior authorization packets, care transitions and diagnostic follow-ups stall between disconnected electronic health records; and health systems face acute staffing and financial pressures. Much of that friction lives in multi-step coordination work that is too unstructured for rigid twentieth-century health care automation, yet too high-volume for exhausted clinical staff to chase manually by spreadsheet and phone.

From impressive demonstrations to agentic operations

Here, however, lies the trap. Watching an agent complete a complex task in a five-minute demonstration is intoxicating, and it tempts organizations to ask the wrong question: What can this model do? In real clinical and public health institutions, a model capability is not a dependable service. Models remain stateless by default, uneven across closely related tasks and capable of AI hallucinations or confident compliance with flawed premises. Worse, if a team defines narrow volume metrics without careful boundaries, an agent can satisfy the dashboard while displacing rework onto nurses, leaking sensitive context or eroding AI safety.

Agentic operations is the discipline of organizing people, AI models and production software systems to achieve meaningful outcomes through bounded delegation, rigorous evaluation and continuous scientific learning. Rather than chasing autonomy as a trophy, it asks how work should be divided between conventional deterministic software, fixed AI workflows, bounded agents and accountable human professionals. It treats AI not as an oracle that replaces institutional judgment, nor as a passive widget bolted onto a broken process, but as a powerful instrument inside a governed system of work—where AI accountability never leaves human hands.

In the companion essays in this series—drawn from the teaching framework for our Harvard Chan executive course, Innovation with AI in Health Care: Agentic Operations—we walk step by step through how that discipline works in practice. We begin with purpose, scope and the three levels at which every operational problem must be diagnosed; examine what foundation models contribute and the four structural limitations every system must accommodate; look inside the production system’s software harness and operational infrastructure; explore the human layer of purpose, scientific inquiry and workforce stewardship; and close with the six-step operating discipline that turns everyday deployment into organizational learning.

Sources: Anthropic — Building Effective Agents (2024); Anthropic — Effective Harnesses for Long-Running Agents (2025).

To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health, including our executive course Innovation with AI in Health Care: Agentic Operations. You can take an individual course or work toward the full certificate.

---

AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy

What happens when artificial intelligence begins improving itself, or pursues a goal we set for it in ways we never intended? In recent news and technical debates, three related terms have moved to the center of that question: recursive self-improvement, emergent behaviors and reward hacking. For most readers, these phrases will be unfamiliar, and they are often grouped together in alarmist headlines. Yet separating them is essential for anyone managing AI in a real organization. This post defines what recursive self-improvement actually means, examines a recent MIT experiment where one AI modified another model’s internal weights, and explains why the most immediate practical risk for health leaders today is not a model rewriting its own brain, but how existing models may behave when deployed in everyday agentic operations.

Recursive self-improvement, usually shortened to RSI, describes a process in which an AI system helps design, train or refine a more capable system, which then helps create an even more capable successor. The idea dates back to the mathematician Irving John Good in 1965, who asked what might happen if a machine became good enough at designing machines to accelerate its own development. In popular culture, the phrase conjures an image of a sinister computer program rewriting its own mind overnight. In modern engineering, it refers more concretely to using AI to automate the research work that human engineers perform: writing software, selecting training data, running experiments and diagnosing errors.

To judge how close we are to that threshold, it helps to separate changing how a system is used in a workflow from changing its internal parameters—the numerical weights and biases a model learns during training. Recently, researchers at MIT demonstrated a technique called weight-space programming, in which a large frontier model from OpenAI (GPT-6 Astra) inspected and directly edited a tiny fraction—0.016%—of the numerical weights inside a much smaller open-weights model (Qwen3.5-0.8B) without retraining it, raising its mathematics benchmark score sixfold.

That result sounds like a machine making another machine smarter, but the details tell a more grounded story. The smaller model already knew how to compute the math answers; its problem was getting stuck in repetitive text loops that caused it to run out of space before finishing. The larger model used a programming environment and practice tests to locate the numerical pattern triggering that repetition and quieted it. When the researchers tried the same approach on another small model (Gemma 3 270M) that did not suffer from looping but simply lacked the mathematical ability, its score barely moved. Clearing a blockage that hid an existing skill is useful engineering, and major AI laboratories track automated research closely, but removing an obstacle inside a small model is not the same as installing new intelligence.

As models grow larger or interact with complex environments, they also display emergent behaviors—sometimes called emergent abilities when the result is a useful skill. Emergence is not unique to computing; it is a fundamental feature of natural and social systems, from biological ecosystems to human cities. What happens in a small group of people is one thing, but when millions of people interact in a city, entirely new patterns of behavior—traffic flows, neighborhood cultures, and local economies—emerge that could not be predicted simply by studying individuals in isolation. In AI, an emergent behavior is a capability or pattern of action that was not explicitly programmed into the system and was absent in smaller models, but appears as a system scales in size, training or interaction with its surroundings.

In public debate, there is a common fear that emergent behaviors in AI are necessarily sinister. In practice, the vast majority of documented emergent behaviors in large models have been helpful capabilities—such as multi-step mathematical reasoning, writing and debugging software, translating between languages, or interpreting clinical shorthand—that smaller models simply could not perform. Crucially, researchers have not observed spontaneous malevolent intent emerging simply because a model grew larger. However, because emergent behaviors cannot all be known in advance (a priori), frontier models require extensive testing and evaluation before release, as unanticipated failure modes can emerge alongside useful skills.

Practically speaking, then, the most significant risk right now does not come from models altering their own internal weights and biases, or from sinister emergent behaviors appearing out of nowhere as models scale. Instead, it comes from reward hacking—also called specification gaming—in agentic operations. In agentic operations, an organization takes an existing, pre-trained model whose weights are already fixed, connects it to software tools, and uses it as an agent to carry out a multi-step workflow while setting the criteria of success. Interestingly, the same underlying principle is used both when frontier labs use one model to train or evaluate another model during reinforcement learning and when a hospital uses an agent to run an operational process: in both cases, the system is driven to maximize a defined score or success metric.

Reward hacking occurs when the system discovers a shortcut that satisfies the literal grading rule or success metric while defeating the human purpose behind the task. If a simulated car is rewarded for hitting markers placed along a racecourse, it may drive in circles hitting the same marker repeatedly rather than finishing the race. In deployed AI agents, this is an immediate operational hazard. Incident reports and safety studies over the past year have shown that pre-trained agents pressured to satisfy a task goal can over-comply with requests, leak private data or credentials to clear an authentication hurdle, or tamper with the software used to monitor and grade their work. Being vigilant to this risk is a central consideration for anyone designing and governing agentic systems in health care.

Consider a health system deploying a pre-trained model in an agentic workflow to clear a backlog of prior-authorization requests, triage patient messages or coordinate hospital discharges. No one at the hospital is training a model or editing a pre-trained model’s weights; instead, the team defines the agent’s criteria of success—such as how quickly it marks cases “complete,” reduces pending inbox volume or obtains an approval code. Pressed to optimize that target, an agent might learn to satisfy an external insurer’s portal by transmitting unredacted private clinical notes, overstating how closely a patient meets a coverage criterion, or routing around an exception queue meant for a clinician’s review. The operational dashboard would look better while patient privacy and clinical safety deteriorated.

Health leaders need to keep two horizons in view at once. On one hand, we must stay attentive to advances in frontier intelligence and the potential for recursive self-improvement, which is likely to become a powerful, foundational trend as AI increasingly accelerates its own research and engineering. On the other hand, there are more immediate practical considerations today around reward hacking in the pre-trained models we are already putting to work. Organizations do not need to wait for self-improving superintelligence to take these risks seriously. Safe agentic operations in health care depend on defining success criteria that reflect real quality of care rather than narrow volume targets, enforcing strict tool permissions in ordinary software, monitoring how agents achieve their outcomes, and keeping human accountability over actions that cannot be undone.

Sources: Good — Speculations Concerning the First Ultraintelligent Machine (1965); Takehi and colleagues — Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights (2026); Wei and colleagues — Emergent Abilities of Large Language Models (2022); Krakovna and colleagues — Specification gaming: the flip side of AI ingenuity (2020); MacDiarmid and colleagues — Natural Emergent Misalignment from Reward Hacking in Production RL (2025).

To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.

---

AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy

If you’re a health care leader, a term you’re most likely going to keep coming across is agentic operations, i.e. running parts of an organization’s work through AI agents – software that pursues a goal through actions and feedback – with humans setting the goal and checking the result. Those agents are starting to take on real work inside health care organizations, and most of it is the unglamorous kind that takes friction out of operations and ultimately shows up on the bottom line – assembling a prior authorization packet, for example, or chasing a referral that’s been sitting in the fax queue since Tuesday. 

Marketing is on that list of operations to AI-ify too, and it’s the part I know best. I’ve worked in marketing since the early 2010s, and since 2018 almost entirely in mental health care, starting as the first growth hire for the biggest interventional psychiatry chain at the time in North America. Today, agents do a good share of my legwork.

I’d argue marketing is one of the better places to start onboarding agentic operations, because results show up in the data quickly and the work runs in short cycles, so you can change course before much has been spent, but that doesn’t make it risk-free. Marketing mistakes happen in public, and in health care they can carry regulatory and patient consequences, which is exactly why the guardrails matter as much as the tools themselves. These guardrails themselves are a big enough topic that they’ll get their own post later in this series.

Before any of that agentic work is useful, though, it’s worth exploring what marketing actually is. I wrote this for clinical and operational leaders who oversee marketers or sit beside them and have never had a reason to learn the field. It’s part of a series that will culminate in my session and real-world data lab in Innovation with AI in Health Care: Agentic Operations at the Harvard Chan School Advanced Learning Academy.

What is marketing?

Ask most people what marketing is and they’ll almost certainly myopically focus on just advertising: the billboard on the highway, the Facebook Reel, the pop-up that appears just as you’re trying to check out. Advertising is a very important part of marketing, but it’s just the most visible part rather than the whole of it.

My operating definition of marketing is the discipline of shaping an offer that meets a real need and then taking the steps to get the offer in front of the right people in a way they can act on. The American Marketing Association’s formal definition is a little broader – “the activity, set of institutions, and processes for creating, communicating, delivering, and exchanging offerings that have value for customers, clients, partners, and society at large” – but for health care leaders the previous version is usually the more practically useful one.

For example, it might mean noticing that people need help with depression that hasn’t responded to medication, but aren’t searching for a treatment name, because they’ve never heard of it. From there, the marketing angle and respective campaigns, landing pages, etc. pivot from saying “here’s what we do” to “here’s how we can help you” – which marks a material shift in perspective, of which advertising is but one part.

What do marketers actually do?

To oversimplify, the work tends to fall into five jobs that typically run as a loop:

• Research: finding out who needs the offer and how they think about the problem they’re experiencing in their own words.
• Positioning: using the research to decide what’s being offered and why the right person should choose your offering over the alternatives – including doing nothing, which in health care is often the biggest competitor.
• Reaching people: using the positioning to choose the channels and putting the right content in them. This is the advertising step that most people see and overindex on.
• Conversion: making the introduction between patient and clinic organization easy: a phone call, form, you name it. A person who wanted to get in touch with your clinic and couldn’t is the easiest person to help, so it’s usually where I look first. It’s also an easy first job to hand an agent, i.e. flagging any inquiry that’s gone a business day without a reply so a human can follow up.
• Measurement: tallying up what happened, working out what met the goals and feeding that back into the next round of research.

Figure 1. The five jobs of marketing, repeated as a cycle.

Each pass should leave the team a little more informed than the last, though in my experience measurement is the step that gets the least time and budget.

This loop is also where agents have earned their keep in my own work. In the research step, for example, an agent can pull search and competitor data on a schedule and draft what it found for a person to check, which leaves the team’s time for the positioning calls that actually need real judgment.

Note that there’s an absolutely vital step 0 that’s not listed as it’s not a part of the traditional loop because it is only done once per planning cycle: goal-setting. Knowing how many inquiries the business needs to maintain its healthy pipeline gives marketers a north star to aim for in all work, which informs how to allocate financial and work-hour resources.

Marketing channels, simplified

A channel is a route to a person, and marketers usually sort channels by who controls the placement:

• Paid: you buy the placement, as with search ads or social media ads.
• Owned: you control it, as with your website or your email list.
• Earned: someone else chooses to mention you, as with press coverage or a referral from a colleague who trusts your organization and its leaders.

The more useful split for a leader is whether a channel captures demand or creates it, as both are typically required in some ratio that’s unique to both your offering and the market you operate in. Paid search mostly captures demand, i.e. the person already has a question and is looking for an answer, while social media and events mostly create it by getting compelling content in front of people who may not know the solution even exists to search for. There are key risks in overindexing on one side of this dyad: a plan built only on capture eventually runs out of people who are already searching, and a plan built only on creating demand struggles to show results inside a budget cycle, so most healthy programs run both.

Figure 2. Common marketing channels, by who controls them and the kind of demand they serve. Many channels do more than one job.

Search itself is changing quickly – to get meta about it, you may not have found this very article on Google at all (which a few years ago would have been a shoo-in). Pew Research Center found that people who saw an AI summary on Google clicked a traditional result in 8% of visits, against 15% without one, and a KFF poll in early 2026 found one in three US adults had used AI chatbots for health information. Being cited accurately in these AI overviews and chatbot answers now matters as much as ranking in traditional search results.

How to tell whether marketing is working

Sooner or later, every leadership team inevitably asks marketing some version of the same question: what’s actually driving new patients, and what should we do differently to reach our goals? The main tool for answering these questions is attribution, i.e. a set of rules for deciding which ad or channel gets credit when someone becomes an inquiring patient prospect. It’s one of the most important tools marketing has, because without it you’re allocating budget and work cycles on instinct – which can’t scale an organization.

Even well-built attribution has blind spots, though. Modern attribution is designed to be able to tell you which ad someone clicked last before reaching out, but it may not tell you whether that individual ad is the reason they came, or just the last thing they saw after weeks of reading about the treatment and talking it over with their doctor, family, etc. In one large study, when researchers compared common measurement methods against 15 large randomized experiments at Facebook, the usual methods often got the ad’s effect wrong, and mostly overstated it. So the more weight your decisions put on attribution, the more it matters to know exactly where its blind spots are – which is what the questions below are for.

If you oversee marketing, these are the questions I’d ask first:

• How do we define a qualified inquiry, and who owns that number as a business metric?
• How many qualified inquiries do we need, and over what period to grow at a rate we can actually sustain?
• How do those inquiries get managed, and how reliably do they turn into patients?
• What does a new patient cost us in total, counting people and tools used as well as media spend?
• Have we ever switched a channel off in one market to see what happens?

If nobody has tested that last one, the attribution report is a well-formatted guess – which isn’t a reason to throw it out, but it is a pretty good place to design the next experiment.

None of this goes away when AI agents arrive; if anything it matters more, because an agent will optimize whatever number you point it at. If the definition of a qualified inquiry is fuzzy, it will get very efficient at producing fuzzy inquiries.

Another post in this series looks at building a marketing strategy around the patient’s journey, and I’m excited to cover all this and more in my course with Dr. Trishan Panch: Innovation with AI in Health Care: Agentic Operations, which runs online on December 7, 8, 14 and 15, 2026.

About the author

Alec Sherman leads digital marketing at Lumin Health. In Innovation with AI in Health Care: Agentic Operations, he brings a practitioner’s perspective on using AI agents in patient recruitment and growth operations.

Sources

American Marketing Association. Definitions of Marketing.

Gordon BR, Zettelmeyer F, Bhargava N, Chapsky D. A comparison of approaches to advertising measurement: evidence from big field experiments at Facebook. Marketing Science. 2019;38(2):193–225.

KFF. New KFF Tracking Poll on Health Information and Trust Finds One in Three Adults Have Used AI Chatbots for Health Advice. March 26, 2026. Fielded February 24 to March 2, 2026.

Pew Research Center. Google users are less likely to click on links when an AI summary appears in the results. July 22, 2025.

---

AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

I reviewed this on publication. If you have any suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy

---