Part 1: Starting with the Right Work
Artificial intelligence is moving quickly across life sciences, but responsible adoption is not defined by how many pilots an organization launches or how advanced its models appear. In a regulated environment, the more meaningful questions are whether the technology is being applied to the right work, whether it is grounded in trusted information, and whether human experts remain accountable for the decisions that matter most.
The FDA’s evolving use of AI offers a useful example. A recent BioPharma Dive article describes how the agency has expanded its internal AI capabilities, including Elsa, a large language model designed to support activities such as writing, summarizing reports, retrieving information, and reviewing lengthy regulatory histories. The article also notes that FDA-reported AI use cases increased by 148% between 2024 and 2025, reflecting a broader effort to use the technology across agency operations.
The most important lesson for pharma is not simply that the FDA is using AI. It is how the agency appears to be introducing it into work where efficiency can improve without transferring final accountability away from experienced professionals.
Begin with burdensome work, not irreversible decisions
According to BioPharma Dive, FDA staff can use Elsa to summarize industry comments on proposals and generate histories of regulatory submissions that may span several years. These are demanding, information-intensive tasks that can consume substantial time before a reviewer begins the higher-value work of interpretation and judgment.
The article also indicates that Elsa is being used to augment staff work rather than make final regulatory decisions. That distinction should guide AI adoption across pharma.
Too many organizations begin by asking, “Which decision can AI make for us?” A better starting question is, “Which burden can AI remove so our experts can make better decisions?”
Within a life sciences company, early use cases could include:
- Summarizing large volumes of medical, clinical, regulatory, or customer information
- Retrieving previous decisions and the evidence supporting them
- Comparing multiple versions of complex documents
- Drafting initial reports for expert review
- Identifying inconsistencies across submissions or internal materials
- Organizing insights from scientific, operational, or commercial sources
These applications may sound less dramatic than autonomous decision-making, but they often create more sustainable value. They can reduce manual effort, shorten the time required to find relevant information, and allow professionals to devote more attention to areas where context, experience, and accountability remain essential.
Responsible AI does not begin by replacing judgment. It begins by protecting the time and attention required for good judgment.
Trusted information matters more than fluent answers
The BioPharma Dive article explains that Elsa evolved from CDER GPT and uses retrieval-augmented generation to help reduce hallucinations. The system is designed to draw from a defined body of trusted information tailored to the needs of individual FDA centers.
This is a critical lesson for pharma organizations evaluating generative AI.
A model can produce a confident and polished response even when the underlying information is incomplete, outdated, or incorrect. In a regulated environment, fluency is not the same as reliability.
The quality of an AI system depends heavily on the knowledge environment surrounding it. Organizations must determine:
- Which sources the system is allowed to use
- Which documents are authoritative
- How approved and outdated versions are distinguished
- Who owns and maintains the source information
- Whether outputs can be traced back to supporting evidence
- How conflicting information is handled
- How frequently the knowledge base is reviewed and updated
This is where many AI initiatives reveal weaknesses that existed long before generative AI arrived. Fragmented repositories, inconsistent metadata, duplicate documents, unclear ownership, and outdated content can all undermine the performance of an otherwise capable model.
The AI challenge is therefore also a data-governance and knowledge-management challenge.
Organizations that invest only in the model may produce an impressive demonstration. Organizations that invest in trusted information, ownership, traceability, and lifecycle management are more likely to create systems that can survive real-world scrutiny.
Design around the user’s role and context
The FDA approach described in the article also suggests that trusted information can be tailored to the needs of different centers and staff roles. A user in one area of the agency may require a different knowledge base, level of detail, and workflow from someone in another.
Pharma companies should resist the temptation to create one generic AI experience for the entire enterprise.
A regulatory professional, medical reviewer, clinical operations leader, commercial strategist, field colleague, and data scientist may all benefit from AI, but they do not need the same information or the same level of control.
Their tools may require different:
- Access permissions
- Knowledge sources
- Output formats
- Validation requirements
- Review and approval steps
- Risk thresholds
- Audit and documentation standards
The stronger model is often a shared governance foundation with role-specific experiences built on top of it.
This approach improves relevance because the system is grounded in the user’s work. It also improves trust because employees are more likely to adopt a tool that understands their context, fits naturally into their workflow, and produces information they can verify.
Enterprise scale should not mean identical experiences for everyone. It should mean consistent standards, shared controls, and fit-for-purpose experiences for different users.
Prepare information for both people and AI-assisted review
The BioPharma Dive article also raises an important future question: if regulators use AI assistants to support review activities, how should sponsors and CROs prepare their submissions?
Tala Fakhouri, a former FDA AI policy official quoted in the article, suggested that greater transparency about regulatory AI use could help organizations prepare submissions with data labels and information structures that are useful to both human reviewers and AI assistants.
This does not mean companies should write submissions for machines instead of people. It means that clear, consistent, and well-structured information will become even more valuable.
Life sciences organizations should begin examining whether their materials have:
- Consistent terminology and definitions
- Clear data labels and metadata
- Traceable supporting evidence
- Documented assumptions and decisions
- Reliable version control
- Easy-to-follow links between claims, data, and sources
- A clear explanation of what changed and why
These improvements benefit human reviewers today, regardless of how extensively AI is used in the future. They reduce ambiguity, improve retrieval, and make complex information easier to interpret.
The broader lesson is that AI readiness begins upstream. It starts with how information is created, governed, labeled, connected, and maintained—not only with the system that eventually reads it.
Responsible scale begins with disciplined choices
The FDA’s journey offers pharma a constructive model for beginning responsibly: apply AI to burdensome, information-heavy work; ground the technology in trusted sources; tailor the experience to the user’s role; and keep people responsible for final decisions.
This is a more disciplined path than pursuing automation simply because it is technically possible.
The first stage of AI maturity is not defined by how much work has been handed over to a machine. It is defined by whether an organization understands:
- Where AI can genuinely improve the work
- Where human judgment must remain central
- Which information can be trusted
- Who is accountable for the output
- How the technology fits into existing decisions and workflows
Organizations that answer those questions first may move more deliberately, but they will also be better positioned to move from experimentation to responsible scale.
Source: Kelly Bilodeau, “The FDA was all in on AI. Will that change?”, BioPharma Dive, July 28, 2026.
If this area interests you further or are interested to be part of the 2026 Pharma CX Marketing Summit in October, please do get in touch with us.