Why heart care keeps a person in the room
Cardiology is two jobs stitched together. One half is pattern reading: electrocardiograms, echocardiograms, stress tests, lab trends, long records. The other half is physical and personal: threading a catheter through a narrowed artery, placing a stent, deciding at the bedside whether a sick patient goes to the lab tonight or waits. Software is improving fast at the first half. The second half barely moves.
Two tasks show the split. Interpreting a resting ECG is a bounded problem with a clean input, and models can draft a read in seconds. Performing a cardiac catheterization is not bounded. The wire behaves differently in every vessel, the anatomy varies, complications arrive without warning, and a licensed physician signs for what happens next. The same is true of the consent conversation before that procedure, where a person weighs a frightened patient’s values against the numbers.
That mix is why the share of task time software can handle today sits at 23 out of 100 on our Can AI do it? scale. The scale and its inputs are explained on the coverage method page, and the wider approach is set out in our methodology.
What software does, what it assists, what people keep
Some cardiology work is already machine work. Flagging abnormal rhythms in long monitoring traces, measuring chambers and ejection fraction from stored images, summarizing a thick chart before clinic, and drafting clinical notes all fall in the group where AI handles the task rather than helping with it. That group accounts for 0% of task time on this page’s task split.
A larger band is assistance. Reading echocardiograms, scoring cardiovascular risk from imaging and labs, choosing between medical management and intervention, and tuning device settings all go faster with a model in the loop, but the call and the signature stay with the cardiologist. Tasks where AI helps rather than replaces cover 48% of the work.
Then there is the part that still needs a person: the catheter lab, the emergency decision, the family meeting about a failing heart, the follow-up visit where a patient admits they stopped their medication. Those tasks make up 52% of task time, and they are the reason the headline figure lands where it does. Our Still needs a human score for cardiologists is 75 out of 100 (higher is safer), explained on the headline score page.
What the evidence actually tests
Here is the honest position. Our evidence grade for cardiologists is D, which means there is no direct, graded head-to-head test of AI against cardiologists in our evidence set yet. Where evidence does exist, it sits at the task level: image measurement, rhythm detection, risk flags. Task-level accuracy is not the same as doing the job.
So we publish no quality-parity number for this occupation. A number would imply a comparison nobody has run. What would settle it is specific: prospective studies where an AI-led pathway and a cardiologist-led pathway handle the same patients, judged on patient outcomes rather than agreement with a reference read, plus results on the messy cases that make up real clinic lists. Until that exists, the sensible reading is assistance with strong supervision. You can see how we grade this kind of claim on the quality-parity page.
The market data points the same way. US employment for cardiologists is about 17,290, with projected growth of 4.8% between 2025 and 2035 and median pay of $496,010 (BLS, 2025). High pay creates strong pressure to automate the paperwork around the job, not the job.
When the picture could change
Most likely after 2041 (8 in 10 of our scenarios). How we build that window is set out on the replacement-year page.
Two things could pull the date earlier. First, regulators approving autonomous reads for narrow, well-defined studies, which would move routine interpretation out of the clinician’s queue. Second, real progress in dexterous robotic hands, since a meaningful share of this job’s task time is physical and the robotics tier for that work is dexterous humanoid manipulation.
Two things hold it back. Liability is one: somebody has to carry responsibility for a decision that can kill, and today that is a licensed physician. Procedural variability is the other. Catheter work is not a repeatable assembly step, and equipment, training and theater time cost far more than a software seat, which blunts the business case for replacing rather than assisting.
Good to know: the erosion to watch in cardiology is early-career task volume, as reading and reporting work thins out the training pipeline.
How to stay needed
Lean into the tasks that sit in this job’s needs-a-human group. Keep procedural volume up, since hands-on competence in the lab is the hardest part to copy. Own the complex decisions where guidelines conflict and the patient’s goals decide the plan. And take the hard conversations: consent, prognosis, deprescribing, end-of-life care in advanced heart failure.
Two skills pay off alongside that. One is knowing how to audit an algorithm’s output, including where a model fails on unusual anatomy or poor image quality. The other is supervising a mixed team, since more of the reading and triage will arrive pre-processed and someone has to check it.
If you are weighing nearby paths, look at general internal medicine physicians, radiologists and neurologists. Each has a different task mix and a different window. You can set two of them side by side with the compare tool, see the wider picture on the diagnosing and treating practitioners family page or the healthcare sector page, and check where this kind of work lands on our list of jobs that most need a person.