Pfizer: generative AI on AWS — 16,000 search hours saved yearly and impact estimated at up to $1B annually

• Driving Patient-Centric Innovation in Life Sciences Using Generative AI with Pfizer — AWS (кейс-стади)
• Three Takeaways from Pfizer at AWS re:Invent Keynote — AWS Industries Blog, 2023-11-28
• Pfizer at AWS re:Invent 2023 — Lidia Fonseca Keynote (1.3 млрд пациентов, 10%→80% облака, цель 19 препаратов за 18 месяцев) — AWS (страница keynote)

Background

Pfizer is one of the world's largest pharmaceutical manufacturers: in 2022 its medicines and vaccines reached more than 1.3 billion people. Behind that number is a years-long digital transformation, which Chief Digital and Technology Officer Lidia Fonseca described in her AWS re:Invent 2023 keynote: the company went from 10% to 80% cloud-based infrastructure, and per the AWS Industries blog migrated 12,000 applications and databases plus 8,000 servers in 42 weeks, saving over $47 million annually (the AWS case page phrases it more cautiously — 'tens of millions of dollars a year').

Back in 2019 Pfizer built the Scientific Data Cloud — a platform aggregating multi-modal data from hundreds of laboratory instruments. That data foundation was battle-tested by the pandemic: just 269 days passed from announcing the COVID-19 vaccine co-development with BioNTech to the FDA's emergency use authorization, and manufacturing capacity grew from 220 million doses across the entire portfolio to 4 billion doses of Comirnaty in 2022. The manufacturing Digital Operations Center gave plant teams a shared view of processes and added 20% throughput.

On that foundation, 2021 saw the launch of PACT (Pfizer Amazon Collaboration Team) — a joint program with AWS for rapid data and AI prototypes in the PSSM division (Pharmaceutical Sciences Small Molecule). The program's logic: short prototyping cycles by blended Pfizer-AWS teams instead of months-long internal builds. PACT has pursued 14 projects, including generative AI; 5 moved to production. The flagship of the generative wave is Vox, an internal platform accessing Anthropic's Claude 2.1 through Amazon Bedrock — and in total, per Fonseca, the company is developing AI across 17 priority use cases, from scientific and medical content generation to manufacturing.

Problem

Fifteen hundred scientists in the PSSM division were drowning in documents. A drug in development generates some 20,000 documents: experiment protocols, stability reports, regulatory dossiers, manufacturing documentation. Finding the right information in that corpus takes hours, sometimes days; at division scale, search consumed thousands of person-hours a year. The program set a measurable target — cut data discovery time by up to 80%.

The second problem was innovation speed. Building a prototype internally took three months or more — and that assumed available specialists with the right skills could be found, which was rarely the case. In pharma, where bringing a drug to market takes years and the competition for data engineers and ML specialists runs across every industry at once, the talent bottleneck was becoming strategic.

The third layer is the industry's regulatory context. Pharmaceutical data is among the most sensitive categories: intellectual property on molecules, clinical trial data, manufacturing secrets. Any generative AI solution had to run inside a protected corporate environment — public chatbots are ruled out for such a scenario. Hence the architectural requirement: models must come to the company's data, not the data to the models. At the same time the company could not wait years for a 'perfect' platform: the 2023 generative AI wave demanded fast but safe experiments.

Solution

The central element of the solution is Vox, Pfizer's internal platform where scientists search documents by voice and chatbot, asking questions in natural language. Under the hood are two AWS services: Anthropic's Claude 2.1 via Amazon Bedrock handles question understanding and answer synthesis, while Amazon Kendra provides intelligent enterprise search across documents. The AWS Industries blog calls Vox a 'Pfizer-certified generative AI platform' — a solution that passed internal security and compliance reviews, not a sandbox experiment.

Equally important is the operating model it was built in. PACT works as a rapid-prototyping pipeline staffed by blended Pfizer-AWS teams: a typical prototype ships in no more than 6 weeks — versus at least 3 months for internal development. Vijay Bulusu, Head of Data & Digital Innovation for PSSM, links that gap directly to the talent shortage: even finding people with the right skills for an internal project was a problem. The joint-team format with a cloud provider removed it without hiring.

Behind the generative flagship stands a broader portfolio: of PACT's 14 projects, five reached production. Predictive manufacturing tasks are covered by Amazon SageMaker (the ML platform), Amazon Lookout for Equipment (equipment anomaly detection), and Lookout for Metrics (metric anomalies). Company-wide, per Lidia Fonseca at re:Invent 2023, generative AI is being developed across 17 priority use cases — from scientific and medical content generation to manufacturing.

The scaling logic is also characteristic: the program did not try to cover the whole company at once. It started in one division — PSSM, small molecules — refined the process, accumulated patterns, and then transferred the experience to the large-molecule division without repeating the trial-and-error cycle. Bulusu sums up the cultural effect: 'With access to the talents and technologies of AWS, we've changed our innovation culture and done a lot in a very short time.'

Result

The program's measured results: up to 16,000 search hours saved annually for 1,500 PSSM scientists and a 55% reduction in infrastructure costs (per the AWS case study). Prototypes ship in 6 weeks instead of 3+ months, five of PACT's 14 projects run in production, and the small-molecule division's experience has been transferred to large molecules. Vox made the corporate document corpus — some 20,000 documents per drug in development — accessible through a natural-language question.

The company stated the scale of its ambition publicly at AWS re:Invent 2023: Pfizer estimates its priority AI use cases will deliver savings of $750 million to $1 billion annually. It is important to distinguish the genres of these numbers: 16,000 hours and 55% are retrospective measurements of a specific program, while $750M–$1B is the company's own forward-looking estimate across a portfolio of 17 use cases, with no published methodology. We present the two categories separately and suggest reading them differently: the first is fact, the second a stated goal.

In our view, this case is interesting above all as the anti-example of the 'big bet': instead of one megaproject — a portfolio of 14 rapidly tested prototypes, of which slightly more than a third reached production. That funnel (14 → 5) is not low efficiency but the normal economics of innovation: testing a hypothesis in 6 weeks costs incomparably less than a year-long project that 'can't be canceled because too much has been invested'. Note also the sequence of layers: the Scientific Data Cloud (2019) and the cloud migration preceded the generative wave — Vox was built on a ready data foundation, which likely explains the speed.

Our second editorial observation: a public impact estimate at CDTO level is a management instrument in itself. By naming the $750M–$1B range in a keynote, Fonseca moved generative AI from the category of IT experiments to that of corporate commitments to the market — with the corresponding resource priority. For companies stuck in pilots, that may be the most reproducible element of the case.

Lessons learned

1. In pharma the fastest generative-AI ROI is not molecule discovery but document search: 20,000 documents per drug turn enterprise search into a gold mine.
2. The 'joint team with a cloud provider' format solves the AI talent shortage: a 6-week prototype without diverting your own engineers.
3. A portfolio of prototypes beats a megaproject: the 14-projects-to-5-in-production funnel is the normal economics of innovation, not low efficiency.
4. Generative AI is the top layer, not the first: the Scientific Data Cloud (2019) and cloud migration preceded Vox — without the data foundation the 2023 speed would have been impossible.
5. Distinguish measurements from forecasts: 16,000 hours and 55% are measured results; $750M–$1B is the company's estimate with no published methodology.
6. A public impact estimate at CDTO level elevates AI from an IT initiative to corporate strategy — and a commitment to the market.
7. Start with one division and replicate: the large-molecule group adopted PSSM's (small molecules) playbook without repeating the trial-and-error cycle.

Frequently asked questions