We read every Complete Response Letter the FDA has published and classified each one by the deficiency sections the agency itself wrote into it. 62.6% carry a manufacturing or quality section. Here is the full analysis, the method, the sensitivity checks, and the script that regenerates every number on this page.
62.6%, or 286 of 457 published letters, carry a manufacturing or quality deficiency section written by the FDA itself. Clinical-side sections appear too, and often in the same letter, which is exactly why the letters resist a single headline number. Here is the full split, four mutually exclusive groups that sum to the archive.
| Group | Letters | Share | What it means |
|---|---|---|---|
| Manufacturing or quality section only | 202 | 44.2% | The letter raised the plant, the product, or the microbiology, and no clinical section at all |
| Both sides present | 84 | 18.4% | Manufacturing and clinical deficiencies in the same letter |
| Clinical side only | 58 | 12.7% | Clinical, nonclinical, clinical pharmacology, human factors or device, with no manufacturing section |
| Neither classified | 113 | 24.7% | Letter used a different structure, or carried only template boilerplate |
Why 62.6% is a floor, not a ceiling. The rule counts a section only when the label sits alone on its own line in full capitals, the way the FDA sets its deficiency headings. Drop that condition and the figure rises to 64.8% (296 of 457). Keep it but narrow the denominator to the 444 records actually typed COMPLETE RESPONSE and it rises to 64.4%. Both alternatives run higher than the number we publish. Template boilerplate is deliberately excluded: SAFETY UPDATE alone appears in 394 letters, so counting it would measure the FDA's template rather than anyone's deficiencies. Full rule, label lists and spot-checks on the full rule, label lists and spot-checks.
67.6% of the published letters are marked approved. That is not a recovery rate, and treating it as one will mislead you. It is a fact about what the FDA chose to release. The first tranche, published July 2025, was drawn from applications that had already been approved, so the early letter years read at or near 100% approved by construction. Watch what happens as the release policy changes.
| Letter year | Letters published | Manufacturing or quality section | Marked approved |
|---|---|---|---|
| 2013 | 12 | 58.3% | 100.0% |
| 2014 | 8 | 50.0% | 100.0% |
| 2015 | 9 | 77.8% | 100.0% |
| 2016 | 20 | 60.0% | 100.0% |
| 2017 | 25 | 56.0% | 100.0% |
| 2018 | 33 | 72.7% | 100.0% |
| 2019 | 38 | 55.3% | 97.4% |
| 2020 | 35 | 71.4% | 100.0% |
| 2021 | 48 | 62.5% | 100.0% |
| 2022 | 25 | 52.0% | 100.0% |
| 2023 | 30 | 60.0% | 96.7% |
| 2024 | 69 | 72.5% | 20.3% |
| 2025 | 59 | 57.6% | 0.0% |
| 2026 | 31 | 67.7% | 0.0% |
68.9% of the 119 biologic applications in the archive carry a manufacturing or quality section, against 60.5% of the 337 small-molecule applications. A gap of 8.4 percentage points, in the direction anyone who has run a tech transfer would predict, and now with a number attached. The living process is harder to make twice than the molecule is.
That gap is the practical argument for putting modality-matched technical leadership in the seat before the filing rather than after the letter. A leader who has taken a biologic through a pre-approval inspection is reading a different risk surface than one who has only shipped tablets.
31 companies hold three or more published letters, and between them they account for 120 of the 457 records, 26.3% of the whole archive. The repeat names skew heavily toward generic and biosimilar manufacturers, where the molecule is settled and the rejection is therefore almost always about the plant, the process, or the paperwork. Rejection risk in this dataset is not evenly distributed. It concentrates in organisations, and it persists across filings.
Language prevalence across the full text of the archive. Facility-related language appears in 65.0% of letters and inspection language in 59.7%, which is the single most consistent signal in the corpus.
| Term | Letters | Share |
|---|---|---|
| Facility | 297 | 65.0% |
| Inspection | 273 | 59.7% |
| Stability | 112 | 24.5% |
| Impurity | 51 | 11.2% |
| Sterility or aseptic processing | 38 | 8.3% |
| Process validation | 21 | 4.6% |
Four credible sources report four different manufacturing shares. They are not contradicting each other. They are measuring different things on different slices. This is the reconciliation.
| Source | Figure | Cohort | What it measures |
|---|---|---|---|
| Pharma Manufacturing (July 2025) | 74% cited quality or manufacturing (150 of 202) | The 202 letters in the first tranche only, every one of them from an application that was ultimately approved | Keyword classification of the letter text |
| McCarthy & O'Boyle, Drug Discovery Today (2026) | CMC 45%, efficacy 35%, safety 21% | 30 approved-after-CRL NDAs, drawn from 358 first-time small-molecule NDAs, CDER 2013-2023 | Peer-reviewed coding of the deficiencies behind each letter |
| Dilek et al., Ther Innov Regul Sci (2026) | Facility 65%, CMC 51%, labeling 44%, efficacy 26%, safety 26% | 43 novel therapeutics, all ultimately approved, 2020-2024 | Peer-reviewed coding; categories overlap, so shares exceed 100% |
| Jefferies (Andrew Tsai), reported by BioSpace (May 2026) | >50% manufacturing, 41% product quality, 27% more clinical data | Full public archive, then nearly 350 letters | Analyst assessment, reported by press, not a public standalone document |
| Phase 3 Search (this report) | 62.6% carry a manufacturing or quality section (286 of 457) | The entire published archive, every letter | The FDA's own section headers; a conservative presence floor, fully reproducible |
| RSM US LLP (February 2026) | >2.5 years average CRL to approval; over half of facility deficiencies were incomplete pre-approval inspections | First 200-plus released letters | Firm analysis of the released set |
Why 74% and 62.6% are both right, and why 62.6% is the one that now describes the archive. The 74% figure is real. Pharma Manufacturing counted it themselves in July 2025: 150 of 202 letters. Two things about that denominator have been lost as the number travelled. It covers the first tranche only, and the FDA drew that tranche entirely from applications that were ultimately approved. The archive has since more than doubled, and the letters added after it are mostly from applications that were not approved.
The rest of the gap is method. Pharma Manufacturing classified by keyword. We classify by the deficiency section headings the FDA writes into the letter itself, which is stricter. Run our own analysis with the loose rule instead and it returns 64.8%, close to the keyword-based figure on a comparable basis. The two numbers are not in conflict. They are the same phenomenon measured two ways on two different slices.
So the number to use now is 62.6%, 286 of the 457 letters analysed, rising to 62.5% across the 459-record live archive. It is the whole published archive rather than one tranche, it uses the agency's own deficiency sections rather than word matching, it is a conservative floor rather than a ceiling, and it regenerates from a published script every time the FDA posts more letters. The 74% describes 202 letters from the summer of 2025. This describes the archive as it stands.
More than half of the facility-related deficiencies in the FDA's first 200-plus released letters happened because the agency could not complete the pre-approval inspection in time (RSM US LLP, 5 February 2026). Part of that is agency capacity and foreign-site access, which a sponsor does not control. A lot of the PAI non-completions in the 2020-2023 window that produced most of these letters trace to travel restrictions, visa timing and inspection backlog, not to anything the applicant did.
A strong technical operations hire cannot make the FDA complete an inspection sooner. What the hire changes is the odds of passing once the agency arrives: a documented, audit-ready site, and a team that has been through a PAI before rather than one meeting it for the first time. I place CMC and quality leaders for a living, and that second part, the part a sponsor actually controls, is what I get asked to staff for.
Worth knowing if you sit on the CDMO side of the table: regulatory violations are the single most cited reason, at 26%, that a CDMO loses a bid (PharmaSource survey). The FDA's letter and the client's procurement decision are reading the same signal.
Worth separating from the other instrument people confuse it with. A warning letter is a compliance action against a facility, issued whether or not a drug is under review. A Complete Response Letter is the FDA's formal rejection of a specific marketing application, issued only inside that review. The full side-by-side is on the warning letter versus CRL comparison.
A plant can carry warning letters for years without a CRL ever being written against it, and a CRL can arrive with no warning letter in sight if the deficiency was clinical, or a first-time PAI finding rather than an ongoing compliance pattern. The two are related and not interchangeable. A warning letter is a smoke alarm for the site. A CRL is the fire department turning up at a specific address on a specific day. You want to have dealt with the smoke long before the second one shows up.
A board hears "Complete Response Letter" and reaches for the runway model. Fair enough: at an industry-typical Phase 2/3 burn of roughly $5-10 million a month (Adverum ran near $9.6M a month in R&D in Q1 2025, per its SEC filing, as one illustrative point in that range), a 1.28-year median delay is not a rounding error.
But the CFO's question, how long, and the CTO's question, what is broken, are different questions with different answers. CMC is plumbing. Nobody in the building thinks about the pipes until one bursts, and by then the fix is expensive and public. The 45% CMC share above says something about which systems get board attention before the letter arrives, and which only get it after.
A CRL usually says more about a company's systems than about any one person in the room, which is worth remembering before anyone gets blamed for it. The uncomfortable version cuts both ways. Some facility-cited letters trace to agency capacity nobody at the company could have moved. Some trace to a site that was not ready for an inspection whenever it landed. GMP run by heroics looks fine right up until the week the FDA shows up unannounced. GMP run as a system does not care which week it is.
Four limits, stated plainly, because an analysis that hides its boundaries is marketing.
Every number on this page comes out of one script run against the FDA's own file. Nobody has to take our word for any of it.
build_crl_dataset.py. It emits every table on this page and validates itself against the
published July 2026 baseline, 75 checks, before it will write output.More than half of the facility-related deficiencies in the released letters trace to a pre-approval inspection the FDA could not complete in time. Some of that is agency bandwidth, and no sponsor controls it. The rest is site readiness, and site readiness has an owner. If you are working out who that owner should be, that is the conversation we have every week.