Why the typing moved before the judgment
This job is audio in, document out. A clinician dictates, and the transcriptionist turns that recording into a report that lands in the patient’s chart. Nothing has to be lifted, inspected in a room, or carried to a bedside. Speech recognition has been chipping away at the first pass of that work for years, which is why the question comes up more here than in most of the health care occupations we score.
Two tasks explain the split. Transcribing a clear dictation into a formatted report is pattern work, and machines are good at it. Spotting an inconsistency in that dictation and checking it with the provider is not pattern work. It needs someone who knows the chart, the specialty, and when a dose, a laterality, or a patient identifier does not add up. That second task is where the job’s remaining weight sits.
Volume matters too. The Bureau of Labor Statistics counted about 41,550 medical transcriptionists in the US and a median wage of $40,410 (BLS, 2025), with employment projected to fall 4.4% between 2025 and 2035 (BLS, 2025). That is erosion, not a cliff. Fewer first-pass typing jobs, more editing and quality-checking seats, and fewer openings for people with no clinical vocabulary yet.
What AI does, what it helps with, and what people still own
On the tasks where AI already carries the load, the pattern is clean dictation and routine formatting: converting a straightforward recording into a draft report, and expanding standard medical abbreviations into their long form. The share of task time in that group prints here: 27%. Our read on how much total task time software can handle today comes out at 48 out of 100, measured the way the coverage method describes.
The assist group is larger than people expect. Reviewing and correcting a speech-recognition draft is faster with the machine than without it, and so is checking terminology or formatting a report to a template. The person is still the one who signs off on what goes into the record. That shared slice reads 66%.
What stays with a person is narrow but load-bearing: querying a clinician about a discrepancy in a dictation, confirming the right patient and the right encounter, and deciding when a passage is too unclear to transcribe at all. That group holds 7% of task time. Small share, high consequence, since an error here follows the patient.
Good to know: this role needs no robot and no new hardware, so the physical brake that slows automation in nursing or maintenance work does not apply here.
What the evidence actually shows
Quality parity for this job carries an evidence grade of D. In our scale, that means no study has tested AI output against a qualified medical transcriptionist on this job’s real work, so we publish no parity number for it. Vendor accuracy claims are not a test, and neither is a transcription benchmark on clean studio audio.
What would settle it is specific: a blinded comparison of machine drafts and human-produced reports across accented speech, noisy clinic audio, and multiple specialties, scored by clinicians on error type rather than word error rate alone, with the rate of clinically significant errors reported separately. Until something like that exists, the honest answer is that the first-pass gap has narrowed and the review gap is unmeasured. Our quality parity method explains why we refuse to guess a figure in that case, and the full method pages set out the rest.
When the picture could shift
Most likely between 2043 and 2053 (8 in 10 of our scenarios). The replacement-year method explains what that window is and is not.
Two things could pull it earlier. Ambient documentation tools are being bought at the health-system level, not desk by desk, so adoption moves in large steps. And the cost comparison is lopsided: a software seat costs a fraction of a salary, as the cost panel on this page shows, which makes the business case easy to make even when quality is uneven.
Two things hold it back. Accountability is one: a record entry is a legal document, and someone has to answer for what it says. Audio quality is the other. Accents, cross-talk, background noise, and specialty vocabulary still produce drafts that need a trained reader, and the editing seat does not disappear just because the typing did. You can see how that mix compares with neighboring record-keeping roles on the compare tool.
How to stay needed in clinical documentation
Lean into the tasks machines leave alone. First, the provider query: get good at spotting a discrepancy in a dictation and raising it clearly and quickly. Second, identity and encounter checks, which are dull until they go wrong. Third, judging when audio is too unreliable to transcribe and saying so instead of guessing.
Two skills carry the most weight now. One is editing speech-recognition output at speed without losing clinical meaning, which is a different craft from typing. The other is coding and record literacy: knowing how a note feeds billing, quality reporting, and the legal record. Both move you toward the review-and-governance side of documentation rather than the keyboard side.
If you want adjacent ground, the closest work sits with Medical Records Specialists, Health Information Technologists and Medical Registrars, and Medical Secretaries and Administrative Assistants. All three keep the clinical vocabulary you already have. The wider other health care support occupations family page shows how the scores line up across that group, and our list of jobs most exposed to AI is a useful sanity check before you commit to a retraining plan.