Two ways to use AI on a medical record
Generative summarization reads a document and writes new text about it: a paragraph describing the patient’s history, findings, and results.
Extraction finds specific values that already exist in the document, like an ejection fraction, an ECG date, or a pathology result, and returns them exactly as written, with a pointer to where each one appears.
Both use AI. They produce very different things, and in clinical intake the difference matters.
Why summaries are risky in clinical intake
They can invent values
A model that writes text can produce a number that isn’t in the record. In a summary, an invented value looks exactly like a real one.
They hide their sources
A paragraph rarely says which page each statement came from. To trust it, a clinician has to go back to the record, which is the work the summary was supposed to save.
They decide what matters
A summary chooses what to include. The specialist’s checklist should make that choice, not the model.
More passes aren’t more accuracy
Running more AI over every raw page adds cost and delay. It doesn’t make a blurry handwritten page any more readable. Knowing when to stop and ask a person does more for accuracy than another pass.
What extraction with citations gives you
- Only what’s needed. Values the intake checklist asks for, nothing else.
- The record’s own words. Nothing paraphrased or generated.
- A source for every value. Page and line, verifiable in one click.
- An honest “not found.” Missing items are flagged instead of filled in.
- A person for the hard pages. Low-confidence pages are routed to your team.
When a summary is fine
Summaries have real uses: drafting, orientation, reading something long for the first time. The problem is using them where a clinician must act on specific values. In intake and consult readiness, every value has to be checkable.
How to evaluate AI for medical records
- Ask whether the system generates text about patients, or only extracts.
- Pick any value it returns and ask to see the exact page and line.
- Give it a poor-quality or handwritten page and see what it does.
- Ask what it does when a required item is missing.
- Ask whether it runs the same processing on every page, and what that costs at your volume.
InfoNotData is extraction-only, cites every value to its page and line, and routes uncertain pages to a person. See how it works, or read what medical fax OCR is, and why it isn’t enough.