Biotech diligence · updated 13 September 2026

Biotech due-diligence document gathering: coverage across 11 companies

Before an investor or a partner commits money to a biotech company, someone gathers everything the company has published: investor decks, regulatory filings, drug labels, research papers and clinical trial records. This is part of due diligence, the checking done before a deal. By hand, it falls to an analyst, one search and one download at a time.

We gave this job to an AI research run for 11 biotech companies. Before each run there was a checklist of what should be found, and we scored each run by how much of that checklist it covered.

5 of the 11 runs found everything on the checklist. The other 6 found at least 94%. Every file a run cited was really saved and not empty.

Companies: 11
Time, all runs: 191m 32s
Text read (tokens): 55.92M
Text written (tokens): 242k

What a research run is

hedwigai makes workbooks: documents such as reports, trackers and lists that AI research runs write and keep current.

A research run is an AI agent given one piece of work, here “gather this company's diligence documents”. It searches the web, opens and reads what it finds, saves the documents, and writes the workbook: a checklist of what it found and a short memo.

The seven kinds of document every run looks for

The checklist is built from the same seven kinds of document for every company.

01
Investor presentations
the decks on the company's own investor page
02
Website literature and posters
every paper and poster the company links from its site
03
Annual and quarterly reports (10-K and 10-Q)
the reports a US-listed company files with the regulator each year and each quarter — or the 20-F and 6-K, which a foreign company files instead
04
The company's own drug labels
the official prescribing information for each approved product, where there is one
05
Competitors' drug labels
the labels for the competing products, and for the treatment doctors use today (the standard of care)
06
Published research
a search of PubMed and Google Scholar for papers on the company's products, and on other products that treat the same disease or work the same way
07
Clinical trial registry entries
every study on the company's products listed on ClinicalTrials.gov, the public registry of clinical trials

How much each run found

Coverage is the share of the checklist a run actually covered, from 0 (nothing) to 1 (everything). The companies are not named. Each is described by its kind, because the kind decides how much there is to find. The dashed line marks 0.95, the score we aim for.

Average coverage 0.971.
5 of 11 at or above 0.95.
Quality 1.000 on every one of the eleven. Quality is a second score: of the files a run says it saved, the share that are really saved and not empty. 260 cited files were checked.

Time and cost

Time is how long each run took from start to finish. Cost is counted in tokens, the small pieces of text (about three quarters of a word each) that an AI model reads and writes. Read is everything the model took in; written is everything it produced.

Kind of company	Coverage	Time	Read	Written
Mid-cap clinical-stage immunology, US-listed	1.000	24m 15s	5.19M	19k
Small-cap clinical-stage oncology, US-listed	1.000	6m 35s	1.81M	9k
Clinical-stage neuromuscular, no approved products	1.000	10m 15s	4.48M	17k
Early-commercial, one recently approved product	1.000	22m 07s	5.52M	29k
Small-cap immuno-oncology, no approved products	1.000	25m 39s	7.69M	21k
Large-cap commercial, multiple approved products	0.947	23m 50s	6.13M	17k
Private, pre-clinical, no registry footprint	0.947	8m 58s	2.25M	25k
Clinical-stage oncology, no approved products	0.947	15m 48s	6.05M	33k
Clinical-stage neuro, ASX-listed foreign issuer	0.947	20m 32s	5.84M	24k
Commercial-stage, foreign private issuer (files 20-F)	0.947	18m 46s	5.56M	26k
Small-cap clinical-stage, one legacy approved product	0.944	14m 47s	5.40M	21k
All 11	0.971	191m 32s	55.92M	242k

A run reads roughly 231 times as much as it writes. That is the shape of the work. It reads filings, labels, registry pages and papers, and writes only a checklist and a memo. The reading is the job.

What can still go wrong

Some companies have less to find. A private company with nothing in trials has almost no public record. A foreign company files a 20-F where the checklist asks for a 10-K. Both of those scored below 0.95.
Quality checks that files exist, nothing more. A score of 1.000 means every cited file is saved and not empty. It is not a person's review of what the files say.
Eleven companies is a small test. Each figure is the last scored run for that company, and the page is updated as the test is run again. It is a record of where this stands, not a claim that the work is finished.