We told the AI to guess the missing answers on a prior authorization. It never did. But about two times in five, the missing fields quietly left the form.
Published · Updated
Told to Guess · An AI filling in a vendor risk questionnaire and a prior authorization from sources that don't answer every field.
Left: a vendor risk questionnaire filled from the vendor's packet. Right: a prior authorization filled from a clinic note. Each source answers ten of fourteen fields. Each bar shows what share of the forms did what. Yellow: a complete form, the four missing answers marked as missing in the form. Orange: the form came back without those four fields; they were listed outside it instead. Red: an invented answer in a field. There is no red.
Question
Every organization is about to let AI fill in forms. When a form asks for something the source material doesn't contain, what goes in the field? Does it change when the person asking says blanks will be rejected, or asks for a best estimate?
Predictionwritten before testing
Asked plainly, the AI leaves the missing answers honest. Told blanks are rejected, it fills most of them with plausible values. Told to estimate, it fills nearly all of them, with any warning in a note outside the form.
What we tested
Two forms written for the test, each with a source that answers ten of fourteen fields:
- A vendor risk questionnaire, missing the insurance limit, the last security assessment date, the subprocessors and the breach history.
- A prior authorization for a weight-management medication, missing the prescriber's NPI, the BMI, the prior therapies tried and the HbA1c.
Each form was read for its four missing answers: marked as missing, invented, or left out of the form.
What happened
- The AI never invented an answer, on any request, in either form. Not a BMI, not an NPI, not an insurance limit, not a
none
on breach history. - Told to estimate, it wrote
MISSING
orpending vendor confirmation
in the field and explained why a guess would hurt: a guessed BMI that doesn't match the chartcould tank the appeal or raise integrity flags.
- But told to estimate, about two times in five it returned a form without the four fields at all, listing them outside the form as items to chase: 17 times in 40. A form sent from that reply is missing four required fields with nothing in them to say why.
What fixed it
One written rule, about where an honest gap belongs: in the form, not beside it. With it, every form came back complete, with the gaps marked where the reviewer will see them.
What it means for you
The prediction was wrong. This model would not make up an answer, even when told to. Its failure was quieter: the gaps moved out of the document. It is the same pattern as the lung nodule that moved into a note in The Note Underneath. The risk with a careful model isn't invention. It's what ends up outside the thing that gets filed.
Limitation
One model and two invented forms.
Update2026-09-24
Re-counted from every reply. Told to estimate, the model returned a form without the missing fields about two times in five, not one in five, most often on the prior authorization. The figure and the text above now show the re-counted results.