The scribe market is noisy
Every health IT vendor now claims AI scribe capabilities. The term has been stretched to cover everything from basic transcription to fully autonomous documentation. Clinics need a framework to evaluate these claims.
The four levels
Level 1: Transcription
The system converts speech to text. That is it.
Value: you get a searchable record of what was said.
Limitation: you still write the note. The transcription is raw material, not a clinical document.
Trust requirement: low. The output is clearly a transcript.
Level 2: Structured summarization
The system converts speech to a SOAP note or similar structure. It identifies chief complaint, history, assessment, and plan sections.
Value: you get a first draft that saves typing time.
Limitation: the summary may miss context, reorder information misleadingly, or include irrelevant details. It has no access to the chart.
Trust requirement: medium. Every section needs review.
Level 3: Chart-grounded documentation
The system generates notes using both the conversation and the patient's chart data. It can reference labs, medications, prior visits, and documented problems.
Value: the note is anchored to facts. Hallucination risk drops because the model has to cite.
Limitation: the system still operates visit-by-visit. It does not reason across time.
Trust requirement: medium-high. Citations are verifiable.
Level 4: Longitudinal clinical leverage
The system remembers the patient across visits. It knows what changed, what was tried, what worked, and what is pending. It can:
- pre-populate visit context before the patient arrives
- flag discrepancies between the conversation and the chart
- suggest plan items based on protocols and history
- generate follow-up tasks and orders
Value: documentation becomes a byproduct of clinical reasoning, not a separate task.
Limitation: requires deep EHR integration. Cannot be bolted on.
Trust requirement: high, but verifiable through citations and audit trails.
How to evaluate
Ask vendors:
- Does the scribe access the patient's chart? (Level 2 vs 3)
- Does it cite specific data points in the note? (Level 3 quality indicator)
- Does it use prior visit context? (Level 3 vs 4)
- Can you see exactly what data the system used? (Trust infrastructure)
- Does it trigger downstream actions or just produce text? (Level 4 indicator)
The investment decision
Level 1-2 scribes are commodities. They save typing time. The ROI is modest and the switching cost is low.
Level 3-4 scribes are infrastructure. They change how clinicians work. The ROI is significant but the integration depth is higher.
Choose based on what problem you are solving. If the problem is "typing," Level 2 is fine. If the problem is "cognitive load and after-hours work," you need Level 3 or 4.
The standard to hold
A scribe that cannot show its sources is not a clinical tool. It is a liability generator. Demand citations. Demand audit trails. Demand longitudinal context.
Frequently Asked Questions
What platforms combine AI documentation with workflow automation?
Platforms that combine AI documentation with workflow automation go beyond note generation to manage the work that the note creates. This means inbox triage, automated follow-up task creation, lab result routing, and care gap identification, all connected to the clinical record. Thyra is built around this principle for endocrinology and primary care, where longitudinal complexity (diabetes, thyroid, obesity, polypharmacy) makes the space between visits as clinically important as the visit itself.
