First in a series on practical applications of AI in pharmacovigilance
The question that started it
How do we reduce the time spent locating controlled documentation without compromising document version control or quality?
We didn’t set out to build anything. We set out to understand whether that question had a good answer. What follows is what we found.
The paradox most regulated organizations will recognize
The stronger and more mature a quality system becomes, the more controlled documentation it naturally generates. SOPs, work instructions, job aids, client conventions, business continuity plans, templates and supporting guidance all exist for good reason. Together, they provide the structure and control needed to support consistent pharmacovigilance activities.
The documentation has never been the problem. The problem is accessing the right part of it quickly enough to support day-to-day operations.
Operational teams rarely work from a single SOP. They often need to navigate interconnected procedures alongside client-specific conventions and other supporting documentation, each with its own version history, approval cycle and relationship to other controlled documents. As organizations grow, navigating that information becomes harder, not easier.
What we kept noticing
Across a number of process improvement initiatives, we kept coming across the same issue: the information was available, but finding it took longer than expected.
- Opening several documents to locate one relevant section
- Confirming whether the latest approved version was in use
- Searching SharePoint for a document everyone knew existed
- Asking an experienced colleague to confirm wording that was already documented
- Working out which procedure or client convention applied before the actual work could begin
None of these activities contributes directly to pharmacovigilance. They delay the point at which skilled professionals can start applying their expertise. Over time, this becomes normal. The same questions are asked, the same documents are opened and experienced colleagues become the people others rely on to locate information.
This does not reflect poor compliance. People return to controlled documentation because they want confidence that they are applying the correct, current and approved process.
“I know it’s in there somewhere.” “Which version is current?” “Can someone send me the link?” “Where’s the exact wording?” “I don’t want to rely on an old copy.”
Those conversations were not describing a documentation problem. The information existed. The difficulty was getting to the right information quickly and with confidence.
The idea we tested: ask instead of search
The concept we explored, which became known internally as SOP Librarian, was deliberately focused. It does not replace controlled documentation. It provides another way for people to access it.
Instead of requiring users to remember where information is held or which document contains the answer, a user can ask a question in natural language. The AI searches across the available controlled-document set, identifies the documents relevant to the question and returns a response supported by those documents. The user can then go back to the source and confirm the applicable version.
The conversational interface is not the main control. What matters is whether the information can be made easier to find without losing source traceability, version awareness or the controls already applied to the approved documents.
We built and tested the proof of concept (POC) using synthetic data. This allowed us to test controlled-document retrieval, version awareness and source traceability without introducing confidentiality concerns or exposing operational information.
Why we treated Quality as the first stakeholder, not the last
Introducing AI into a regulated environment should raise questions about control and reliability. We involved Quality early because those questions needed to shape the POC rather than be addressed after it had been built.
For a tool like this to be useful in pharmacovigilance, users need to know where an answer came from, that it is supported by approved documentation and that the correct effective version is being referenced. Speed on its own is not enough.
The strongest theme in the discussions was not “make it faster.” It was “make sure I can trust it.” In practical terms, people wanted:
- Every response traceable to an approved source
- Confidence that the current effective version was being referenced
- Clear visibility of where the information originated
- The ability to search across the available controlled-document set and identify which documents contain information relevant to the question, particularly in large quality systems with hundreds of SOPs
- Confidence that the system would not create process or requirements that were not supported by the source documentation
The last point became an important requirement for the POC. Any answer provided by the system needed to be supported by the source documentation. If the information was not available in the approved documents, the system should not attempt to fill the gap.
The scope of the POC was therefore kept focused on retrieving information from controlled documentation. It is not intended to make operational decisions or replace existing review processes. Expert judgement remains with the user.
Used within those boundaries, the concept could support the existing PV and Quality framework by making approved documentation easier to locate and reference. It can also reduce reliance on memory, outdated local copies or colleagues knowing where particular information is held.
Where the friction actually lives
The potential benefit became clearer when we looked at where document retrieval was taking time in existing processes.
Inspections and Sponsor audits. Responding to a request can involve locating the applicable procedure, confirming the approved version and identifying related work instructions or client conventions. Supporting documentation then needs to be assembled with appropriate traceability. A significant amount of preparation time can therefore be spent locating and confirming controlled information.
Reconciliation. Before investigating a discrepancy, teams may need to confirm which procedure applies, whether the process has changed and what any applicable client convention requires. Accessing that information more efficiently allows the investigation itself to begin sooner.
Quality review. Reviewers often need to reference several controlled documents while assessing operational outputs. Reducing the time spent locating those documents allows more of the review to focus on the quality of the work itself.
Complex regulatory deliverables. Periodic activities and regulatory submissions can require teams to work across procedures, work practices, templates and supporting guidance. The challenge is often establishing the applicable process before the substantive work can begin.
Onboarding and training. New team members frequently rely on experienced colleagues to identify which SOP applies or where particular guidance is held. Making that information easier to locate could reduce repetitive procedural questions and allow experienced staff to spend more time on coaching and more complex issues.
How we’d know whether it worked
We would not measure success simply by the number of questions asked or responses generated. A more useful assessment is whether people can locate controlled information more quickly, whether they are consistently directed to the correct approved documentation and whether routine document-location queries reduce.
As the POC develops, measures could include document retrieval time, user confidence, onboarding time, audit preparation time and the amount of SME time spent supporting routine document searches.
The practical test is whether it makes it easier for people to follow the right process while maintaining document control, traceability and compliance with approved procedures.
What we’re still working through
This remains a proof of concept, and there are areas that still need to be worked through before a retrieval-based tool of this type could be treated as part of a controlled operational process.
- Version currency at source. The tool can only retrieve the correct effective document if the underlying repository is properly controlled and maintained.
- Scope boundaries. The distinction between retrieving what a procedure says and interpreting how it should be applied needs to be clearly defined.
- Validation and qualification. The appropriate approach for a retrieval tool that supports users but does not make decisions needs to be proportionate to its intended use and risk.
- User behavior. We need to understand whether easier access changes how people use controlled documentation and whether users continue to verify the source where appropriate.
- Configuration and change control. The controlled state of the solution needs to be maintained as source documents and system configuration change. This includes how new or revised documents are introduced and how user feedback, including flagged weak or incorrect responses, is reviewed, assessed and used to improve the solution. Changes would need to be tested and documented so that accuracy, traceability and the intended behavior of the POC are maintained.
These are not theoretical issues. They are part of determining whether the concept can move beyond a POC and operate reliably within an established quality system.
The takeaway
Controlled documentation is fundamental to pharmacovigilance operations, but its value depends on people being able to find and apply the right information when they need it. As document libraries grow, maintaining that accessibility becomes more difficult.
The SOP Librarian POC is testing whether retrieval-based AI can help with that problem while retaining source traceability, version awareness and the controls around approved documentation. The aim is not to replace expert judgement. It is to reduce the time spent finding the information needed before that judgement can be applied.
There are still questions to resolve around validation, scope, configuration control and how the tool should be maintained over time. Those questions are part of the POC and will need to be addressed before any broader operational use is considered.
If other organizations are working through similar questions, particularly around validation expectations or the boundary between retrieval and interpretation, we would be interested in comparing experiences.
About UBC
United BioSource LLC (UBC) is the leading provider of evidence development solutions with expertise in uniting evidence and access. UBC helps biopharma mitigate risk, address product hurdles, and demonstrate safety, efficacy, and value under real-world conditions. UBC leads the market in providing integrated, comprehensive clinical, safety, and commercialization services and is uniquely positioned to seamlessly integrate best-in-class services throughout the lifecycle of a product.
About the Author
Chad Alexander, Safety Support Specialist, PV Process Excellence
Chad Alexander is a Safety Support Specialist within UBC’s PV Process Excellence team with over 10 years of experience in pharmacovigilance, drug safety, regulatory compliance, and safety operations across pharmaceutical, CRO, and pharmacovigilance technology environments. His experience spans global clinical trial and post-marketing safety activities, regulatory reporting, pharmacovigilance compliance monitoring, safety process optimization, DSUR and other periodic safety reporting requirements, expedited reporting, regulatory intelligence, EMA (EudraVigilance, xEVMPD, CTIS) and other global health authority systems, audit and inspection readiness, and the management of client-specific pharmacovigilance processes.
At UBC, Chad supports the Process Excellence function through regulatory intelligence activities, global safety reporting compliance monitoring, interpretation of worldwide pharmacovigilance requirements, process governance, quality oversight, procedural development, training support, and continuous improvement initiatives. He has also contributed to several process improvement and digital transformation projects, including the research, testing, and proof-of-concept development of AI- and automation-based solutions to support pharmacovigilance operations and compliance activities.
Chad holds a BSc (Hons) in Pharmacology from the University of Hertfordshire.
