The complete write-up of the 2026 evaluation programme behind every number on this site: the corpus, the method, the measured results, the capabilities involved, and the limits of what we can claim.
Medical experts do not lack judgement. They lack time, and the time goes to a specific place: rebuilding the case before the analysis can even begin. Reading hundreds of pages, sorting scans, reconstructing a chronology, confronting assessments that disagree, finding the one document that supports each finding.
We wanted a number, not an impression: how much of that preparation can be absorbed, and does absorbing it change what the expert actually sees in the case?
The programme ran on 25 real, anonymized cases from practicing Swiss medical experts: insurance and judicial assignments from their own files, not demonstration material. Cases were selected for the properties that make expert review expensive: volume, fragmentation, and disagreement between authors.
| Cases reviewed | 25 |
|---|---|
| Largest case | 1,100 pages |
| Most documents in a single case | 24 |
| Most authors in a single case | 8 |
| Formats present | Scanned PDFs, Word files, spreadsheets, handwritten notes, old faxes, DICOM imaging, audio dictations |
| Languages | French, German, Italian and English, including mixed-language files |
Each case was reviewed twice by the same expert. First with their usual method: the case as it arrived, the tools they normally use. Then with Avicen: the case uploaded exactly as it arrived, with no sorting, renaming or preparation.
Three measures were recorded for both passes:
Cases were anonymized by the experts before upload. Processing ran entirely on dedicated infrastructure, and no data was used to train any model.
90%
less preparation time
5×
more material inconsistencies identified
10×
faster source verification
Measured on 25 real anonymized cases, before and with Avicen, by practicing medical experts (2026 evaluation programme). Figures rounded for publication.
The preparation figure is the least surprising to anyone who has opened a 1,100-page case: when the chronology, the findings and the links to evidence are already assembled, preparation shrinks from evenings to minutes. What remains is the part that cannot be absorbed: the expert reading of what matters.
The inconsistency figure is the one that matters professionally. The experts did not read badly the first time: a case of that size simply exceeds what any reader can hold in memory. Five times more material inconsistencies were identified once the case was fully cross-referenced, and several were the kind that change conclusions.
The verification figure changes behaviour rather than workload: when the source of any statement is one click away, checking stops being a cost. Experts verified more, not less.
1,100 pages · 14 documents · 5 authors
One report documented 60 mg per day. A later summary listed 30 mg for the same period, with no explanation of the change. The dosage appeared across 14 documents with no consolidated treatment history, and the discrepancy had gone unnoticed in the initial review. Avicen flagged the 60 mg and 30 mg statements side by side, with their dates and source passages.
1,100 pages · 24 documents · 8 authors
Between two opinions, the assessment moved from complete incapacity to 50% capacity. No new examination, diagnosis or documented clinical improvement appeared between them. With both opinions side by side and the interval empty of clinical events, the unsupported change was visible in minutes.
760 pages · 18 documents · 6 authors
The decision sat in a clinical note; the evidence sat in a lab table, an imaging report and a later summary. The dosage increase followed worsening laboratory values while imaging remained stable, yet a later report attributed the decision mainly to the imaging. The gap between the stated rationale and the record sat on one screen, ready to cite.
The results above are not produced by one feature. They come from a chain, and every link in it was exercised by this corpus:
Avicen (2026). Twenty-five real cases, before and with Avicen: the 2026 evaluation programme. https://www.avicen.io/evaluation
Journalists and researchers: for methodology details beyond this page, write to contact@avicen.io. We answer within 24 hours.
A 30-minute call to scope one of your cases, then a free, guided 14-day evaluation. No patient data is needed to start.
Reply within 24 hours. No commitment.
