AI Turns Drug-Making Paperwork Into Instant Answers
Could shave months off the time it takes new medicines to reach you.
Making a single medicine generates an enormous paper trail — lab notebooks, manufacturing recipes, safety reports, regulatory filings. That information lives in different teams, different formats, and often different languages. So when a drug moves from the lab to a factory, or when a regulator asks a question, someone has to dig through all of it and re-check the details by hand. It's slow, expensive, and mistakes are costly.
The new system, described in a research paper, attacks this in two steps. First, it swallows everything — digital files, scanned pages, even handwriting — without losing anything, and organizes it into a tidy hierarchy of documents, sections, and passages. Second, AI agents (software that can take actions and make decisions on its own) read across those documents to pull out key concepts and link related facts together, building what's called a knowledge graph — essentially a map showing how pieces of information connect.
When you ask a question, a routing agent decides which approach to use: the plain-text search layer for simple lookups, or the connected-facts layer for questions that span multiple documents. The team tested it on 505 questions drawn from 38 development reports for a Sanofi small-molecule drug. It scored 95% on multiple-choice questions, but only 85% when graded more strictly by another AI — and it struggled most with comparison questions and questions requiring a view of the whole document set. That gap matters: a system that looks great on easy tests can still stumble where it counts.
The bigger picture is that drug companies sit on decades of trapped knowledge. If tools like this work at scale, handoffs between teams get faster, regulatory filings get smoother, and medicines may reach patients sooner. The same approach could apply anywhere with mountains of messy documents — law firms, insurers, engineering firms. The catch: this is a research demo on one company's program, not yet a proven product across the industry.
- The system reads scanned, handwritten, and multilingual documents without losing information, then links related facts into a searchable map.
- On 505 questions from 38 Sanofi drug-development reports, it scored 95% on multiple-choice but only 85% under stricter AI grading.
- It's weakest at comparison and big-picture questions — the exact kind regulators and executives tend to ask.
- The same 'turn document piles into answers' idea could spread to law, insurance, and engineering.
Why It Matters
Faster drug handoffs and filings could mean medicines reach patients sooner and cost less to develop.