Physicians are trained to manage complexity under time pressure. We carry twenty patients, track a dozen pending studies, remember which consultant owes us a callback, and keep it straight. Then we go looking for a job and manage the whole thing out of an inbox and a legal pad.
A hospital-employed position search generates a startling volume of material: recruiter emails, competing time zones, credentialing packets, site visit logistics, contract drafts that differ in ways that matter enormously and are almost invisible on first reading. Most of us handle it the way we handle a busy call night, by reacting. The good opportunities and the bad ones arrive in the same font.
Over the past several months, I ran my search differently, using a frontier AI model as a working assistant rather than a curiosity. Not to decide anything for me, but to do what a good chief resident does: hold the list, prepare the material, surface what I need to see, and never let something quietly fall off.
Treat the search as a service, not a mood
Every opportunity is a patient on the list, with a status, a pending item, and a disposition. I asked the model to build a tracking spreadsheet: one row per opportunity, with columns for institution, contact, position type, call structure, compensation model, stage, last contact, next action, and when it is due. It took under a minute. The second pass mattered more. I described how I actually think about these jobs and asked what columns I had failed to request. It suggested endoscopy volume expectations, APP support, ownership structure, and whether the position was a replacement or a growth hire. That last one proved among the most predictive fields in the sheet.
Real physician voices, twice a week
Free, and one click to unsubscribe.
The tracker is not sophisticated. It is the difference between knowing where things stand and thinking you do.
Define your criteria, then actually search against them
Most of us run a passive search. We wait for recruiters to surface things and react to what arrives, which produces a candidate pool assembled by other people’s incentives.
The alternative is to define criteria explicitly and search against them. I worked through what I actually wanted: geography and its acceptable radius, employment model, call structure, procedural volume, teaching involvement, group size, whether I wanted to build a service line or join a functioning one. Writing that down took an hour and clarified more than the previous month of browsing.
Then let the model search against it. With web search enabled, it works through health system career pages, academic postings, and society boards far faster than you can, screening out what fails your criteria before you read it. Ask it to identify systems in your target geography that plausibly have a need even where nothing is posted. A hospital that recently opened beds, lost a physician, or announced a service line expansion is a real lead that will never appear on a job board. The output is a short list you assembled deliberately rather than a long list that assembled itself around you.
The CV and the letter that goes with it
Most physician CVs are archaeological. Mine had been added to for three decades and edited almost never: accurate, complete, and organized so that a reader had to work to find anything.
Rebuilding it was the highest-yield afternoon of the search. I asked the model to restructure rather than rewrite: consistent formatting, reverse chronology, dates in a form a credentialing office can verify without calling me, clinical and administrative material separated so a division chief and a credentials verification organization (CVO) each find what they came for. Then I asked it to flag every gap, inconsistent date, and unexplained transition before someone else did. It found several. Better me than a credentialing committee.
Two rules. The facts stay yours: A model asked to “improve” a CV will smooth a description into something more impressive and less true, and in credentialing that is not a stylistic problem but a disqualifying one. And resist letting it flatten your voice into recruiter English. Physicians reading your CV can tell.
The introductory letter is where the technology earns its keep, because it should be different every time and almost nobody has the energy to make it so. Keep one letter that is genuinely yours, then for each opportunity feed in your research and ask for a version tailored to that group. What are their gaps? What in your background addresses them? A letter naming the service line they are trying to build reads nothing like one that could have gone to forty places. Then rewrite the first and last paragraphs yourself. Those are the ones that get read.
Correspondence and the logistics tax
Recruiter email has a low signal-to-noise ratio and a high anxiety cost. Somewhere in forty messages is the one that actually fits. I paste batches in and ask for structured extraction: what position, what location, what employment model, what is being offered, and what is conspicuously omitted. That last question is the valuable one. A model reading forty emails side by side notices, faster than I will at ten at night, that six describe call coverage in detail and four say nothing at all. It also drafts responses, and a polite decline that keeps a door open takes me fifteen minutes to write well and thirty seconds to edit.
Coordinating a Webex with a chief in another time zone and a two-day site visit around a clinical schedule produces nothing when done correctly and disaster when done poorly. I use the model as a scheduling scratchpad, then ask it to build a site visit briefing packet: the itinerary, who I am meeting, their background, and questions tailored to each. Asking the chief what you already asked the recruiter is a wasted meeting.
The dossier
Before any serious conversation, I build a research brief on the practice and the health system: ownership and recent transactions, whether the division is employed or independent, referral base, competing systems, leadership turnover, public quality reporting, financial condition of the parent system, and any notable litigation or regulatory action. The model is excellent at assembling public material quickly. It is not a substitute for calling someone who works there, and it will occasionally state something outdated or wrong with complete confidence. I treat the output the way I treat a well-prepared student presentation: useful, organized, requiring attending review.
Extend the same approach to the place itself. Housing, schools, commute, state tax structure, licensure timelines, malpractice climate. Physicians routinely evaluate a job in enormous detail and evaluate the life around it by driving through a neighborhood once.
Contracts and compensation
Two agreements can look similar and be profoundly different: compensation structure, wRVU thresholds and conversion factors, call obligations, tail coverage, restrictive covenant geography and duration, termination-without-cause notice periods, what happens to your production bonus if you leave mid-year. A side-by-side table across offers, clause by clause, with plain-language explanation of what each provision means in practice, surfaces the asymmetries fast. One agreement I reviewed had a without-cause termination clause that read as entirely standard until it sat in a table beside another offer’s.
The same applies to the numbers. Before responding to any offer, ask for a benchmark: what base, conversion factor, productivity threshold, call stipend, signing bonus, relocation, and vacation are typical for this specialty, region, setting, and level of experience. That is enough to tell you whether an offer sits at the tenth percentile or the seventieth. Verify specific figures against the actual surveys, because a confidently stated wrong number is worse than none when you are sitting across from a CMO. But knowing the range makes for a different negotiation than guessing.
This is preparation for your attorney, not a replacement for one. What it does is make you a better-informed client, which lowers your legal costs and raises the quality of the conversation.
Dictate what happened, then build the calculus
The most perishable information in a search is your own impression. You finish a forty-minute call and you know something. Three weeks and four calls later, you have lost it.
Physicians already solved this problem for patient care. We dictate. Walking to the car after a site visit, I dictate into my phone the same disorganized way I would dictate a consult: who I spoke with, what they said about call, that the chief seemed guarded when I asked why the position was open, that the endoscopy unit looked understaffed, that my wife would like the town more than I expected. No structure.
Then hand it to the model and let it do the structuring. It pulls the discrete facts into the tracker and keeps the impressions and reservations as narrative notes on that opportunity. An unusable voice memo becomes a maintained record. After eight or ten encounters, you have a real dataset: verified terms, benchmarked compensation, dossier findings, and your own contemporaneous impressions, all on the same rows.
That is when you can build the calculus. Ask the model to weight your criteria, but weight them yourself first, because the weighting is the decision. If call burden matters four times as much to you as base salary, say so and watch the ranking move. Then run the sensitivity: Which weights would have to change for the second-place offer to win? If a small change flips the answer, the two are effectively tied and you are deciding on something the matrix does not contain.
The exercise was most valuable when it produced a ranking I disagreed with. That reaction is information. If the numbers say take the job and your gut says otherwise, the model has just told you that your real criteria are not the ones you wrote down.
The part the matrix cannot score
None of this reads a room. It cannot sense that a division is quietly fracturing, tell you what the nurses think of the chief, or explain why the last person left. That comes from calling someone who was there. It will also state plausible things that are wrong, so verify anything you would act on, and it is no place for protected health information or credentialing files carrying identifiers.
Every physician I know who regrets a move regrets it for reasons that were never in the spreadsheet.
You are not the only person taking this job. If you have a partner, they are taking it too, and so are your children, and their scoring differs from yours. A position that ranks first on call and compensation while putting your spouse three hours from their work, or your daughter into her third high school, is not the winning offer. It only looks like one because the criteria were written by one person. If the model builds you a decision matrix, build a second with your partner’s weights. Where they disagree is the conversation you actually need to have, and it belongs before the ranking, not after.
Career stage changes the calculus in ways worth naming honestly. A group evaluating an experienced physician is quietly asking how many years they are buying. That question is real and you should expect it. Less often said is what those same groups have learned from watching early-career hires: Younger physicians move, and often, frequently chasing a compensation delta into a situation that turns out to be worse. The physician twenty-five years in is usually optimizing for something else, for the people, the culture, whether the environment lets them practice the way they believe medicine should be practiced. Institutions burned by turnover have come to see that as an asset. Stability has value, and experienced physicians supply it precisely because they are no longer maximizing pay.
That is worth saying plainly to a prospective employer, and to yourself, because the pull of the highest number is strong even when you know better.
The argument for using these tools here is the argument for using them anywhere in medicine. The administrative burden around the work has grown until it crowds out the judgment the work requires. So let the technology hold the list, assemble the research, and do the arithmetic. It is very good at all of it. Then close the laptop and decide the way this has always been decided, by asking whether you and the people you love can build a life there. No model has any idea. It never will.
This essay is cited in the KevinMD record on physician careers.
Brian Hudes is a board-certified gastroenterologist and hepatologist with more than thirty years of clinical experience and a recipient of his specialty board’s thirty-year certification award. He built and ran a private gastroenterology practice for twenty years, then spent more than a decade in hospital-based gastroenterology, most recently as chief of gastroenterology and medical director of GI and endoscopy at a 550-bed Level I trauma center in Pensacola, Florida. He is an assistant professor of medicine at Florida State University College of Medicine.
Those two halves of a career shape the way he writes about medicine and clinical technology. Outpatient practice and hospital-based care fail their clinicians differently, and many systems are designed without a meaningful understanding of both environments.
Dr. Hudes has been building clinical software alongside his medical practice for decades. In 1995, during his gastroenterology fellowship, he co-developed one of the first Windows-based endoscopy reporting systems in the United States. He is now founder and chief executive officer of AiMOS Systems Corporation, where he is developing an AI-native clinical platform for outpatient and hospital-based care. AiMOS has been accepted into NVIDIA’s Inception program.
He writes about the architecture of clinical information, administrative cost growth, physician workforce shortages, board certification policy, and the widening gap between what clinicians need and what the health care technology industry builds. He also publishes The Renaissance Physician, an independent publication examining medicine, technology, physician workforce, and the structure of modern clinical practice.
