Correction: The previous version made unsupported claims about molecule generation, genomic and clinical data integration, and a contribution to Paxlovid. This version is limited to the use cases documented in the official AWS case study.
Problem and delivery model
A single drug can produce about 20,000 documents, leaving scientists to search historical data manually across tools. The Pfizer-Amazon Collaboration Team (PACT), launched in 2021, prototypes selected ideas and moves successful ones toward MVP and production.
Confirmed technology
Early search work used Amazon Kendra. The generative-AI search now accesses Anthropic Claude 2.1 through Amazon Bedrock. Scientists use Pfizer's internal Vox platform to search documents by voice or chat. A separate PCMM manufacturing prototype uses Amazon SageMaker and related services for anomaly detection and maintenance prediction.
Estimated architecture
AWS discloses the principal services and flows, but not the authorization model, indexing implementation, prompts, or evaluation controls. The overall status is therefore inferred.
- ▸Confirmed: Research documents and historical data—about 20,000 documents per drug—feed the search use case; PCMM equipment data feeds the manufacturing use case
- ▸Confirmed: Early search used Amazon Kendra; current generative search uses Claude 2.1 through Amazon Bedrock; anomaly detection uses Amazon SageMaker and related services
- ▸Confirmed: Vox accepts voice or chat questions for document search, while the manufacturing system detects anomalies and predicts maintenance
- ▸Inferred: Document-level authorization, evidence display, audit logs, and answer evaluation form a regulated-workflow control layer
Reported outcomes
- ▸Up to 16,000 hours of search and extraction time saved annually
- ▸Estimated 55% infrastructure-cost reduction
- ▸PACT prototypes generally completed within six weeks, versus an internal estimate of at least three months
- ▸Five of the initial 14 projects moved into production
These figures are estimates and statements in the AWS case study; they do not demonstrate clinical outcomes or drug approvals.