Industries / Health / Pharmaceuticals
Quality and records
Checking that each batch was made as written, and keeping records that prove it.
6
Milestones achieved, from 1910 to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1910
Tested before it shipped
Ehrlich offered Salvarsan to the many doctors who asked for it only after quality controls and further tests in his laboratory, and his team developed tests for whether the drug had oxidised.
Credit
Paul Ehrlich's laboratory
1997
Electronic records accepted
The FDA's Part 11 rule, issued in March 1997 and in effect from 20 August 1997, set the conditions under which electronic records and electronic signatures count as equal to paper records and handwritten signatures.
Credit
FDA
2008
A contaminant in heparin
US investigators led by the Centers for Disease Control and Prevention linked 152 adverse reactions in 113 patients in 13 states, between November 2007 and January 2008, to the blood thinner heparin made by Baxter. Vials held a contaminant, oversulfated chondroitin sulfate. A second team showed how it could cause the reactions and proposed an extra test to add to the existing checks.
LessonA test finds only what it was designed to look for.
Credit
D. B. Blossom and colleagues (CDC investigation); T. K. Kishimoto and colleagues
2013
Falsified data, criminal plea
The generic drug maker Ranbaxy pleaded guilty to three felony charges, including lying to the FDA, and agreed to pay $500 million in penalties, the US Department of Justice announced on 13 May 2013.
LessonManufacturing and test records are legal evidence, and false ones are a crime.
Credit
US Department of Justice; FDA
2018
An impurity from a process
On 13 July 2018 the FDA announced a recall of medicines containing valsartan, a blood-pressure drug, after NDMA, a probable human carcinogen, was found in active ingredient from a third-party supplier. The FDA said the impurity "was unexpected and is thought to be related to changes in the way the active substance was manufactured".
LessonA change in how an ingredient is made can create impurities that existing tests do not look for.
Credit
FDA
2026
AI-drafted records draw a warning
The FDA sent a warning letter to Purolea Cosmetics Lab of Livonia, Michigan, which said it had used AI to produce drug product specifications, procedures and master control records. The FDA cited 21 CFR 211.22(c), the rule on the duties of a company's quality control unit.
LessonAI can draft a GMP record, but the company's quality unit still answers for it.
Credit
FDA; reported by RAPS on 17 April 2026
Watch for
AI review of batch records
Enveda built CDP, a reviewer of executed batch records (the filled-in copy of a batch's written instructions) based on Claude. On four records from one manufacturing campaign, 1,061 scanned and mostly handwritten pages, it raised all 11 issues that human reviewers had found, counted across several runs. Anthropic reports that it also flagged 7 issues the humans had missed and cut review time by 89%, from over 25 hours to under 3.
Achieved whena party outside both companies runs the method on other records, from another product or site, and publishes the scores. The note "Seven more flagged" (Claude at Enveda) examines how the two counts were gathered.
Claimed by
Enveda and Anthropic, October 2026
Open
01
A public batch-record answer key
A set of realistic but fictional executed batch records for a made-up product, scanned, handwritten, with crossed-out corrections and with errors placed on purpose. Publish the answer key so that any review tool, human or machine, can be scored.
The briefCredit · unclaimed02
Test AI-drafted master records
A public test in which several AI systems draft master production records, the written instructions for making a batch, from fictional product briefs. Two reviewers with GMP experience score each draft against the items US rules require, such as the product's name and strength and a complete list of components. The drafts are never used to make anything.
The briefCredit · unclaimed
Open 01 · the brief
A public batch-record answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A set of realistic but fictional executed batch records for a made-up product, scanned, handwritten, with crossed-out corrections and with errors placed on purpose. Publish the answer key so that any review tool, human or machine, can be scored.
Why it is still open
The best-documented test so far used Enveda's own records, which Enveda says "cannot be made public" because they hold proprietary and personal data. No one outside can rerun it.
Built with what exists
Enveda's open-source companion code, open-doc-review (MIT licence); a word processor for the master batch record; volunteers to fill pages by hand; a scanner; Claude, Gemini and open-weight models to test. The advance is language models that read scanned handwriting a page at a time; the note "Seven more flagged" covers one.
First step
Write a two-page master batch record for a fictional placebo tablet, then fill 20 pages by hand with 10 seeded errors.
Done looks like
At least 500 pages holding at least 100 seeded errors in stated classes, an answer key checked by two people, and at least three review tools scored on how many errors they catch and how many false alarms they raise, all published.
Open 02 · the brief
Test AI-drafted master records
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which several AI systems draft master production records, the written instructions for making a batch, from fictional product briefs. Two reviewers with GMP experience score each draft against the items US rules require, such as the product's name and strength and a complete list of components. The drafts are never used to make anything.
Why it is still open
The FDA's 2026 warning letter shows a company using AI to write such records. I found no public measure of how often AI drafts leave out required items.
Built with what exists
the text of 21 CFR 211.186, turned into a checklist; Claude, Gemini, GPT and an open-weight model; a spreadsheet for scores.
First step
Turn the list in 21 CFR 211.186(b) into a checklist and score five drafts.
Done looks like
30 briefs, 4 systems and 120 drafts, each scored by two reviewers, with the miss rate for each required item and the agreement between reviewers published.