By Deborah Borfitz

March 4, 2025 | Generative artificial intelligence (AI) is being used by AstraZeneca to expedite its five-year ambitions to be an $80 billion company, deliver 20 new medicines, and be carbon negative. The initial value-driving activities include the creation of AI assistants to help with 3D location detection on CT scans and facilitate knowledge extraction for clinical documents as well as an intelligent protocol tool built in partnership with medical writers, according to Vaishali Goyal, the company’s generative AI lead in the R&D IT group focused on the development portfolio. 

Goyal was speaking at the recent Summit for Clinical Ops Executives (SCOPE) in Orlando about the journey with generative AI at AstraZeneca, which began by partnering with leadership to get them to see the “art of the possible.” With over 240 global trials in the R&D pipeline, the company is now running multiple pilots with a range of AI technologies to simplify its processes, she reports. 

Establishing leadership buy-in meant having tough conversations about how generative AI, including large language models (LLMs), could help boost efficiency to enable the company to meet its mission. To enable leaders to envision the potential of AI amidst all the hype, Goyal and her team spent about a year doing proofs of concept before conducting a retrospective analysis of existing platforms and capabilities and getting joint strategy alignment for use cases.  

The first step involved setting clear priorities and level setting about the different possibilities with AI and their probability of success, she says. To have 20 new medicines by 2030 requires thinking about ways to reduce the timelines for drug discovery and development, which “takes a lot of mindset shifts.”  

One key difficulty is that senior leaders have different definitions of success than folks on the ground, says Goyal, which many times has required finding a “happy medium.” Checkpoints were established throughout every type of development R&D IT took on.  

To better predict which patients are most likely to benefit from company-developed medicines, it was also important to be more predictive than reactive, she adds. It became the means for R&D to get a seat at the table alongside the moneymaking business side of the house. 

Trio of Use Cases

The generative AI initiative required analyzing capabilities of current data platforms and deciding which could and should be expanded with an LLM layer to digitize the company’s R&D processes, Goyal says. AstraZeneca has both partnered with outside vendors and built its own data products in-house.  

A cross-value chain was created for both development and research to identify data products and areas ripe for improvement, she continues. The decision was made to invest in the company’s data foundation to make it more robust, leveraging the power of AI to drive insights into R&D processes broadly across the portfolio while enabling the company’s core IT infrastructure to track those newly developed solutions. 

Use cases were developed in as little as one to four weeks, and in some cases over three to six months, to “fail fast” and move on to better solutions, says Goyal. Arriving at a joint strategy for these wasn’t easy, she adds, because “everyone wanted to invest in AI without actually knowing what it could do.” But that “year of innovation” led AstraZeneca to land on a highly targeted approach based on what would work. 

For the first presented use case, Goyal shares that the company invested in a radiomics platform to help speed up location detection on 3D CT scans in lieu of having radiologists manually annotate the images. It succeeded in reducing the amount of paid human expert time needed to do the analysis. 

The second use case was designed to help generate content for clinical study protocols. It provides access to a repository of information (e.g., informed consent form and older protocols) that can be tapped to build new studies. “We started small with...  a subset of sections,” says Goyal, but the AI tool has proven so valuable to medical writers in composing first drafts that it is now receiving further investment. 

The protocol tool is currently focused primarily on oncology and practitioners in later stages of clinical development, she later reports. The plan is to move toward earlier stages and “bring other therapeutic areas along this journey.”  

An AI-powered “development assistant,” the third use case presented, seeks to address common interdepartmental needs such as data extraction, search functions, and speedier analytics, she says. The project focuses on data products that were built in-house and embedding an LLM on top of them. 

This has allowed internal teams to have an AI assistant that can build charts, help with clinical analysis, and make comparisons between quality and clinical data across the portfolio. Although the tendency has been for everyone to work in their own silo, says Goyal, “we’re now able to partner better with other domains.” 

Change Management Approach

Managing change in the workforce is a bigger challenge than developing technology solutions, Goyal says. People need support through the process and to understand that “AI will not replace scientists, but scientists who use AI will replace those who don’t.”  

AstraZeneca’s change management approach has been to partner with different internal groups on the ground and seek top-down support, says Goyal. Specific tactics have included “mini-SCOPEs” within the company when different use cases and technologies for knowledge sharing are showcased. Learning modules and accreditation are also offered to AstraZeneca practitioners at the gold, silver, and bronze levels to encourage them to become more aware of the emergence of generative AI models such as ChatGPT and how to use them as well as where the company is placing its money. 

Success is being measured in part by surveys that have found 85% of stakeholders expecting generative AI will increase their productivity at work, 93% saying it is having a positive impact on their work, and 86% that AI tools will help them achieve their goals. “By these stats alone we know that we actively want to invest in [AI] at AstraZeneca,” she says. 

The clinical study protocol use case will be among the first to scale up because it’s one of the hardest documents for medical writers to create, she notes. It will not be easy due to the “different nature of writers and their exposure to [AI],” as well as the change management process involving a user acceptance testing phase where they do a cross-comparison validation of different protocol sections. 

For the summary section of the protocol, four out of five writers found the protocol tool useful, shares Goyal. “Now we’re trying to do this for each and every section because... each [one] will have a different level of complexity.”