Blog

We Read 445 FDA Rejection Letters. Most Are About the Factory.

2026-08-31

FDA began publishing Complete Response Letters in 2025. As of the 13 August 2026 update there are 445 of them in the openFDA transparency API, issued between 2002 and 2026.

We parsed all 445 letter texts, not the metadata. The headline:

In 236 of 445 letters (53%), FDA raised a manufacturing problem and no clinical deficiency at all. Only 42 letters (9%) were the reverse.

The drug usually worked. The making of it was not ready.

How we counted, and why it matters

Every CRL is organised under FDA's own ALL-CAPS section headings: FACILITY INSPECTIONS, PRODUCT QUALITY, CLINICAL, NONCLINICAL. We extracted those headings from each letter rather than searching the text for keywords.

That distinction is the whole analysis. Keyword-matching this corpus produces nonsense, because most of a CRL is boilerplate. Search for "labeling" and you will find it in 99% of letters, and conclude that labeling is the great unsolved problem in drug development. It is not. It is a standard paragraph.

Here is how often each heading appears:

The first five are near-universal. SAFETY UPDATE is a standing request to include a safety update when you resubmit. PROPRIETARY NAME and CARTON AND CONTAINER LABELING are procedural. OTHER is the closing section about the one-year clock.

The signal is in the headings that do not appear every time.

The substantive distribution

Collapsing the bolded categories:

96 letters (22%) carry no substantive deficiency heading at all. Those are applications held up on labeling, naming or procedural grounds.

This corroborates The FDA Group's analysis of 89 letters, which found 56% with facility issues and 41% with product quality problems. Our numbers come from a sample five times larger and from FDA's own section headings rather than a reading of the prose, and they land in the same place. Two independent methods agreeing is worth more than either alone.

If your CRL cites a facility, it will not tell you what is wrong

This is the practically useful finding, and it surprised us.

Facility sections are almost entirely formulaic. Here is one in full, from a 2018 letter to AB Pharmaceuticals:

During a recent inspection of the Macleods Pharmaceuticals Limited (FEI # 3007517881) manufacturing facility for this NDA, our field investigator observed objectionable conditions at the facility and conveyed that information to the representative of the facility at the close of the inspection. Satisfactory resolution of the observations is required before this NDA may be approved.

And a 2016 letter to Salamandra, LLC:

During a recent inspection of Sun Pharmaceutical Industries Limited, Halol-Baroda 389350, Gujarat, India, our field investigator conveyed deficiencies to the representative of the facility. Satisfactory resolution of these deficiencies is required before this application may be approved.

Neither says what the deficiencies were. They were conveyed at the close of the inspection, which means they are in the Form 483, not in the CRL.

Two consequences. If you are responding to a facility CRL, the letter is not your specification: the 483 is, and you need it from the site, which is often a contract manufacturer you do not control. And if you are researching a competitor's CRL hoping to learn what went wrong at their plant, the letter will not tell you.

A trap in the dataset

Every record carries an approval_status field, and aggregating it produces a finding that looks dramatic and is not real:

Read naively, first-cycle outcomes collapsed after 2023. They did not. The two release batches were selected on opposite criteria: the earlier release covered CRLs for products that went on to be approved, the later one products still unapproved. That split measures FDA's publication policy, not industry performance.

Any "percentage of CRLs that end in approval" drawn from this dataset is measuring the wrong thing. We have seen the figure quoted.

What else is in there

445 letters, 311 companies. 79 companies appear more than once. Accord Healthcare leads with six; Tanvex BioPharma, Eli Lilly, Fresenius Kabi and Regeneron have five each. A CRL is a normal event in drug development, not a mark of an unserious sponsor.

328 NDAs, 103 BLAs. Biologics are 23% of the corpus.

Median length about 8,700 characters, longest over 45,000.

98% name a person. 438 of 445 carry a company_rep, the sponsor's named regulatory contact.

Caveats

The texts are OCR'd from scanned PDFs and the quality varies. Redactions appear as artefacts, so facility names and FEI numbers are sometimes unreadable. Our heading extraction found at least one heading in 438 of 445 letters; the remaining seven are OCR casualties.

More importantly, this is not a random sample of CRLs. It is the set FDA has chosen to publish, and as the table above shows, that choice was not neutral. Everything here describes the published corpus, not the population.

What to do with it

The generalisable lesson is that first-cycle failure is usually a manufacturing-readiness problem, and manufacturing readiness is the part of a programme most often owned by someone else: a contract manufacturer, whose inspection history becomes your approvability issue.

The specific lesson is narrower and more useful. Read the letters issued in your own therapeutic area by your own review division. Reviewers are consistent, and a CRL is the clearest statement you will ever get of what a specific division found insufficient about a product like yours, in its own words. That was simply not available two years ago.

For the mechanics, see our glossary entry on Complete Response Letters: what a CRL is, the three options that follow one, the one-year clock, and Class 1 versus Class 2 resubmissions.

How we did it

The corpus is one API call:

https://api.fda.gov/transparency/crl.json?limit=100&skip=0

Paginate to 445, filter letter_type to COMPLETE RESPONSE (the endpoint also carries a handful of tentative approvals and refuse-to-file letters), then extract ALL-CAPS headings from between the "in its present form" sentence and the "Within one year" closing. Anything outside that window is letterhead or boilerplate.