Designing trust into AI-assisted pharmacovigilance.
Reframing complex safety case processing through explainable AI, human oversight and a focused enterprise experience—designed to help teams review information with greater clarity and control.
Complexity is inevitable.
Confusion isn't.
Pharmacovigilance teams work across source documents, structured case data, medical assessments and regulatory requirements. The experience needs to make every step understandable, traceable and reviewable.
Make AI useful without making it unquestionable.
AI can accelerate extraction and surface patterns, but safety case processing still depends on human judgement. The design challenge is to make machine-generated information legible, verifiable and easy to correct.
ExplainabilityHuman oversightTraceabilityOne connected review workspace
Bring source evidence, extracted fields, confidence cues and review actions into a coherent context.
Clarity
Show what the system found, where it came from and what still needs attention.
Control
Keep review, correction, routing and approval decisions visible to the people accountable for them.
Confidence
Use confidence as a prioritisation signal—not as a substitute for factual accuracy or expert judgement.
From fragmented tasks
to a connected workflow.
When information is spread across documents, forms and disconnected review steps, users spend effort navigating the process instead of evaluating the case.
Multiple inputs
Reports arrive through different channels and formats, with variable quality.
Manual interpretation
Users locate relevant facts and transfer them into structured case fields.
Context switching
Reviewers compare source evidence, extracted data and case history.
High accountability
Completeness, consistency and regulatory readiness must be checked before progression.
Efficiency versus certainty
Automating too little preserves repetitive work. Automating too much can obscure uncertainty and reduce meaningful human review.
“Automate the repetitive. Illuminate the uncertain. Keep consequential decisions accountable.”
AI should assist the decision.
Not hide the evidence.
The experience is organised around transparent assistance: show the extracted information, expose its source, signal review priority and preserve a clear path for human correction.
Make uncertainty visible and actionable.
Confidence cues help users decide where to focus attention. They should be paired with source evidence, missing-data indicators and clear review actions.
High, medium and low are interface groupings. The underlying thresholds and calibration must be established with the AI/ML team and validated against representative data before being treated as operational truth.
Keep evidence within reach.
Reviewers need to understand not only what was extracted, but where the information originated. A source-document view reduces unnecessary context switching and supports verification.
Corrections should retain provenance and distinguish machine-generated values from human-reviewed or human-corrected information.
Assist
Extract, organise and summarise information where the system has a suitable capability.
Explain
Expose source evidence, uncertainty, missing information and reasons for review.
Verify
Keep human review, correction and approval explicit and traceable.
One workspace.
Multiple layers of context.
The redesigned screens bring together operational prioritisation, case assignment, AI-assisted review, duplicate investigation and regulatory validation.
Start with what needs attention.
A focused queue gives users a clear entry point into incoming work. Status, ownership and review needs should be scannable so users can identify the next appropriate action without opening every case.
Make ownership and handoffs explicit.
Assignment is more than moving a row between users. The interface gives the team a place to understand case context, assign responsibility and maintain continuity across the workflow.
Review the information, not just the score.
Field-level cues help focus attention. A robust production pattern should pair each extracted value with provenance, verification status and a way to inspect or correct it.
Support comparison before consequential action.
Potential matches need a reviewable comparison, not a black-box decision. Similarity should prompt investigation; it should not silently merge cases or override domain judgement.
Turn validation into a clear next step.
Validation feedback should identify the affected field, explain what needs attention and help the user resolve the issue. Submission readiness must remain distinct from AI extraction confidence.
Compare the experience.
Move the slider to compare an earlier application screen with the redesigned case review experience. This is a visual comparison of the supplied screens, not a measured performance result.
Use the slider to reveal the redesigned screen. The two source images have different layouts, so the comparison is illustrative rather than a pixel-aligned overlay.
Design for outcomes.
Measure what matters.
No production performance figures are claimed here. The following is a measurement framework to connect the experience to operational quality once baseline and post-launch data are available.
Speed alone is not success.
In pharmacovigilance, efficiency must be considered alongside accuracy, completeness, traceability and appropriate human oversight.
What to measure
Bars are conceptual indicators only—not real values or a claim of improvement. Define numeric targets with product and operations stakeholders.
Operational efficiency
Median case-processing time, review turnaround and time spent on repetitive data entry.
Data quality
First-pass completeness, correction rate, missing mandatory fields and validation failure rate.
Trust & governance
Appropriate override rate, source traceability, reviewer confidence and audit completeness.
Trust is a product
behaviour.
Trust is not created by adding an AI badge or a confidence colour. It is earned through consistent, inspectable behaviour across the entire workflow.
“The best AI experience is not the one that removes people from the process. It is the one that helps people make better-informed decisions.”
What this work reinforces
Designing AI for regulated environments requires clear boundaries, visible evidence, recoverable actions and accountable human decisions.
Client names are presented as text marks; this strip represents broader portfolio experience, not necessarily clients on this specific engagement.