Using AI on client lab reports: what to check

AI is most useful on a lab report for the copying: reading values, units and dates off the page. It is least trustworthy when it moves on to interpretation, where it can sound confident about a result it has misread or about a client it knows nothing about. If you use it, check every value it gives you against the original page: the number, the unit, the marker name, the draw date and the lab. Before you upload anything, find out where the data goes, whether it's used for training and how long it's kept.

I build Sera, a lab-review workspace for practitioners, and Sera uses AI in one place, so read this with that in mind. The checks below apply to any tool, including mine.

Two reasonable positions

Some practitioners don't use AI in their practice at all. That's a defensible choice. Client lab results are health information, the client didn't agree to have them sent to a model, and a mistake in a number can turn into a mistake in advice. If that's your position, the rest of this article is still useful, because your clients may be using AI on their own results and bringing you the answers.

Other practitioners use it every week, mostly to avoid retyping a long panel. That's defensible too, as long as nothing it produces reaches a client file or a client conversation unchecked.

The two camps are closer than they look. Neither wants a model deciding what a result means for a client. The difference is whether the typing can be handed over.

Where AI is useful, and where it isn't

Reading values off a page. A lab PDF is a table of markers, values, units and flags, often across several pages. Pulling that into rows is tedious, and it's the part of the work where a tool's output is easiest to check: every number it gives you is printed somewhere on the page. It can still go wrong, which is why the checks below exist.

Interpretation. This is where I'd be most careful. A model asked to interpret a panel only has what's on the page and in your prompt. It doesn't have the range you work with, the client's history, medications or cycle day, or your scope of practice, and it can still produce a fluent answer. An answer that names a cause, or suggests what to take, is the diagnostic step, and it stays with you or the client's doctor. There's more on that line in explaining lab results within your scope of practice.

What clients bring you. When a client arrives with an AI summary of their own labs, treat it the way you'd treat a direct-to-consumer company's summary: ask for the lab's own report, and work from that.

What to check on every value

Do this with the original report open beside the result, page by page. Don't check against a second AI pass.

The value. Look for a moved decimal point (4.2 read as 42), a value taken from the neighboring row, and signs that got dropped. A result printed as "<0.5" or ">1000" is not the same as 0.5 or 1000. Check that a flag such as H or L wasn't read as part of the number, and that the lab's reference interval wasn't read as the result.

The unit. Copy it exactly as printed. Some unit changes leave the number alone: ferritin in ng/mL and µg/L is the same figure. Many don't. Vitamin D in nmol/L is about 2.5 times the ng/mL figure, as the vitamin D range page shows with its 50 to 80 ng/mL (125 to 200 nmol/L) target. A tool that quietly converts, or drops the unit, makes the number meaningless. There's more on this in comparing results across labs and units.

The marker name. Close relatives get merged. CRP and hsCRP are different tests with different cut-offs. So are total and free testosterone, and serum and RBC magnesium. Check that a calculated value such as eGFR was taken from the report and not worked out by the tool.

The draw date. Use the collection date, not the date the report was printed or received. A panel drawn in March and reported in April belongs in March on a trend chart.

The lab. Note which lab ran each test. Intervals differ between labs, and for some markers the number itself isn't comparable: the TPO antibody note says to read a result against the cut-off on the report it came from.

What's missing. Count the markers on the page and the markers in the output. Rows split across a page break, results in a comment field and a second table on the last page are easy to skip.

If a tool can't show you which line and page each value came from, these checks are slower but still worth doing. If you can't do them at all, don't save the result.

Privacy questions to ask any tool

These are questions, not answers. Ask them of any AI tool before a client's report goes into it, and get the answers in writing, from the provider's privacy policy or terms rather than from a sales page.

1. Where does the data go? Which company processes it, and in which country? Is the whole PDF sent, or only the text?
2. Is it used for training? Is client data used to train the provider's models, or anyone else's? Is that the default, or something you have to switch off?
3. How long is it kept? By the tool, and by any AI provider behind it. Can you delete it, and what happens to it when you do?
4. Who can see it? Staff, contractors, other customers? Is it kept separate from other accounts?
5. What agreement covers it? If you're a covered entity under HIPAA in the US, the rules require a written contract with a business associate that handles protected health information on your behalf (45 CFR 164.502(e)). Under GDPR, a processor's work has to be governed by a contract (Article 28). Ask whether the provider will sign one.
6. What did the client agree to? Your own privacy notice should say which tools see their data.

A personal chatbot account you also use for everything else is the hardest case to answer these questions for. Removing the client's name doesn't make a report anonymous either: most reports also carry a date of birth, an ID number, the ordering clinician and the draw date.

Keep the judgment where it is

Whatever reads the page, the reading of the results is yours. A useful rule: let a tool fill in rows, never a conclusion. Every value it gives you gets checked against the page, and every statement to a client comes from a range you can name and a source you can show.

Where Sera fits

Sera uses AI (Anthropic's Claude API) in one place only: reading values out of the text of an uploaded lab PDF. Every value is shown next to its page in the report for you to check before it's saved, and each saved value keeps the line and page it was read from. The ranges, scoring, patterns, retest suggestions and reports are fixed rules, not AI. If you'd rather not use AI at all, you can enter results by hand instead. Every range is published with its sources and an evidence grade. It doesn't diagnose or recommend treatment.

On the privacy questions, the privacy policy gives Sera's answers: the extracted text is sent to the API rather than the PDF itself, under the API terms the data is not used to train AI models, and Sera doesn't use client data to train AI either. Client data can be exported and deleted at any time.

It's free for three clients with no card. There's a 3-minute tour if you'd rather watch first.

Alina Radu is the founder of Sera and a software engineer. The idea started with her own Hashimoto's tests, scattered across years without the big picture.