Which Medical Specialties Are Most Exposed to AI?

Ask which specialties AI threatens and you get the same three answers: radiology, pathology, dermatology. Image in, label out. That answer is not wrong. It answers the wrong question.

Whether a machine can do the core task is one of five things that set what a specialty is paid. The other four are demand for the work, who controls the referral, how many people or systems can supply it, and what the payer decides it is worth. A specialty can be untouchable on the first axis and lose a third of its income on the other four. Radiology has spent ten years demonstrating this (the radiology decade is worth reading on its own).

The Specialty Exposure Map exists to score all five. It covers 97 specialties and is free behind an email gate: open the Specialty Exposure Map.

The question the map answers

Not "will AI do my job?" The question is "what happens to the dollars attached to my job?" Those are different questions with different answers, and the second one is the one that matters for a mortgage.

Clinician income has five inputs: demand, referral, supply, reimbursement, and time. AI reaches four of them without a machine ever seeing a patient, and each channel, with its evidence, is laid out here.

The map scores the first four and leaves out the fifth. Time is two-directional (administrative offload can extend a career; the productivity treadmill can shorten one) and it depends more on your employer than on your specialty. It gets its own treatment in the book rather than a number on the map.

What gets scored

Each specialty receives a score from 1 to 5 on five components. The scale runs one direction only: higher means more exposed, lower means better protected. The five components are averaged into a composite, and the composite is what you see first.

• Substitution (Sub). The question it answers: Can the core task be standardized and automated? What pushes the score up: Pattern recognition on digital inputs; predictable cases; fixed execution steps; tolerance for error; large training datasets already in existence.
• Demand (Dem). The question it answers: Can demand for the work itself be reduced or removed? What pushes the score up: Upstream prevention, better diagnostics, or medical management that removes the need for the downstream service.
• Referral (Ref). The question it answers: Is your volume gatekept by referrers who can redirect it? What pushes the score up: Dependence on upstream cognitive specialists or a health system that owns the funnel and can steer it.
• Supply (Sup). The question it answers: Can AI, workflow redesign, or task-shifting expand who can do the work? What pushes the score up: Work that advanced practice providers or AI-assisted generalists can absorb once the hardest step is standardized.
• Reimbursement (Rmb). The question it answers: Can policy, payment redesign, or AI-enabled efficiency reduce what is paid or move the payment elsewhere? What pushes the score up: Income concentrated in a few procedure codes; exposure to efficiency adjustments, site-of-service cuts, and algorithmic utilization management.

Substitution deserves a note because it is the component everyone starts with and the one most often scored on vibes. The book's procedure-level instrument scores the work itself on five properties: how standardizable it is, how predictable the case is, whether execution follows a fixed sequence, how much error the task tolerates, and how much data already exists to train on. That instrument produces a result people don't expect. Standardized, high-volume procedural work (routine scopes, cataract surgery, some excisions) scores as more exposed than its "proceduralist" label implies. High-variability, high-dexterity work scores lower. Being a surgeon is not the moat. Doing unpredictable work with your hands is.

The map takes that substitution score and folds in the other four components. A deep-moat specialty can still land in the exposed half of the map: the machine can't do the operation, but demand for the operation is shrinking, a health system owns the referral, and the payer has decided the professional fee is worth less this year than last.

How to read a score

A score is a national, specialty-level estimate. It is not a statement about your job. A hospital-employed physician in a consolidated market and an owner of an office-based practice in the same specialty can sit a full point apart on referral and reimbursement while sharing one row on the map.

For specialties where practice patterns diverge that much, the map shows a shaded band with a dot at the central estimate. The band is the range across practice types; the dot is the number used in the composite. Where a specialty was added after the original scoring pass, the row carries a marker showing it is provisional and pending validation.

A score is not a forecast; it describes exposure to pressure, not a prediction that the pressure arrives. It is not a career recommendation. And it is not evidence-graded in the systematic-review sense. The scores are structured judgment, applied consistently across 97 rows, documented so that you can disagree with any component of any row.

What the method rewards and penalizes

You can predict a fair amount of the map from the method alone. Substitution scored by itself puts image-and-pattern specialties on top, which is the answer everyone already had. Folding in referral and reimbursement moves a set of procedural specialties up the map that looked safe on substitution: work that is dependent on a referring specialty or a health system for volume, and paid through a small number of codes that CMS revalues on a schedule. Folding in supply moves specialties up whose routine work is already shifting toward advanced practice providers.

The specialties that score lowest tend to share three features rather than one: task variability that resists standardization, direct patient acquisition with little dependence on a gatekeeper, and payment spread across enough services that no single policy change moves the whole income. None of those is a specialty label. All of them are things a specialty can have more or less of, and things an individual practice can change.

Which specialties land where is on the map, sortable by composite or by any single component. I would rather you look at the row that applies to you and argue with it than take a ranking out of context.

How it was built

The scores live in a single source-of-truth file that regenerates the map. The print version of the book (an appendix to Chapter 5) shows the central estimates as of press time. The live version updates periodically, and the feedback form at the bottom of the map is how it updates: you tell me which component of your row is wrong and why, and I read every submission. The scores are directional estimates, not a vote, so a submission changes a score when it comes with a mechanism I hadn't weighed.

The substitution instrument is cited to the surgical-AI literature, including my own peer-reviewed work in it, and the reimbursement component is scored off public Medicare payment data.

Where the book picks up

The map tells you where your specialty sits. The book is about what to do with that. Chapter 5 derives the two instruments and makes the case that the procedural moat is a durability window of ten to fifteen years rather than a permanent condition. Later chapters use your position on the map to size a cash reserve against your own exposure rather than a generic three-to-six-month rule, and to decide how early and how hard to build income that doesn't depend on your hands.

Open the Specialty Exposure Map, and if the row for your specialty is wrong, tell me.

‍