What problem does it solve?
Marketing authorisation holders must collect, assess and report adverse events, and the volumes keep rising: reports arrive from patients, doctors, clinical trial sites, call centres, social media, literature and business partners, in every format and language. Each report has to be checked for the minimum criteria of a valid case, triaged for seriousness because serious cases have short legal reporting deadlines, entered into the safety database and coded before medical review.
Most of that intake work is repetitive data entry done by trained staff, while the scarce skill in pharmacovigilance is medical judgment: causality, signal detection and benefit risk. Regulators face the same flood from the other side, receiving individual case safety reports and sponsor safety reports that must be extracted and loaded before anyone can analyse them.
- Pfizer reports that its Worldwide Safety organization processed approximately 1.4 million adverse events globally in 2019.AI in Drug Safety: Building the Elusive 'Loch Ness Monster' of Reporting Tools (2020)
How does it work?
- Collect reports from every channel. Emails, scanned forms, call centre notes, partner files, literature hits and web form or chatbot submissions arrive in one intake queue.
- Check validity and triage. The AI checks the four minimum criteria (an identifiable patient and reporter, a suspect product and an adverse event), detects duplicates and flags seriousness, such as fatal or life threatening outcomes, so the reporting clock is visible from day zero.
- Extract and code. It extracts patient, product, event, dates and narrative into the safety database structure (E2B), suggests MedDRA terms and product dictionary matches, and shows a confidence level per field.
- Route for review. Safety professionals check the extraction, especially low confidence fields and serious cases, and complete medical assessment, follow up and regulatory reporting.
- Help reporters report. On the public side, a conversational assistant can guide patients and professionals to the right form, ask for missing information and submit the report.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Internal tools, Email, Web chat, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Compliance quality, Speed and cycle time, Employee productivity, Lower cost to serve, Risk and loss reduction.
Indicative value
A drug company that takes in 100,000 adverse event reports a year
USD 266,667 to USD 2.7 million
Case intake effort released per year
How this is calculated
Formula: cases * intakeMinutes / 60 * effortSaved * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Adverse event reports taken in per year cases, reports per year | 100,000 | 100,000 | The reference company. |
| Intake minutes per report today (triage, data entry and coding) intakeMinutes, minutes per report | 20 | 40 | Editorial assumption. Replace with a time study of your own intake step. |
| Share of intake effort the AI removes effortSaved, fraction of intake minutes | 0.2 | 0.5 | Editorial assumption; the evidence on this page publishes no measured intake savings. Validate in a pilot before relying on it. |
| Fully loaded cost per case processing hour costPerHour, USD per hour | 40 | 80 | Editorial assumption covering internal staff and outsourced case processing. Replace with your own. |
What it leaves out: Intake effort only. It leaves out medical review, follow up and submission work, the value of fewer late reports, validation and inspection readiness costs, and the platform cost. No organization on this page has published measured savings, so treat the range as a hypothesis.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
U.S. Food and Drug Administration, Center for Drug Evaluation and Research
United States · Government and public sector · 2025
FDA's Center for Drug Evaluation and Research reports in the 2025 federal AI use case inventory that it uses OCR and AI, through the commercial tool ThinkTrends, to extract data from the Investigational New Drug safety reports that sponsors send in. The extracted data is converted to the E2B(R2) format and ingested automatically into the FDA Adverse Event Reporting System. The inventory describes manual extraction of these reports as labor intensive and time consuming for regulatory staff, and states the aim as faster processing and regulatory action on adverse events reported in clinical trials. It lists the use as deployed since March 2025.
No outcome disclosed.
U.S. Food and Drug Administration
United States · Government and public sector · 2024
FDA reports in the 2025 federal AI use case inventory that a conversational assistant built on the commercial platform Druid helps people who report an adverse event or product problem through the Safety Reporting Portal. It answers questions from a knowledge base, routes the reporter to the right form for the product type, helps complete it and submits the report to the portal through an API, with the stated aims of better data integrity and faster form completion. The inventory lists it as deployed since March 2024; no measured results are published.
No outcome disclosed.
Pfizer
United States · Pharma and life sciences · 2020
Pfizer's Worldwide Safety organization, which processed about 1.4 million adverse events in 2019, worked with industry experts to build an AI platform for the repetitive intake steps of adverse event case processing. In its first phase the model makes basic intake decisions, such as whether a report is a valid case and whether it is fatal or life threatening. The Drug Safety Unit in Rome was the first location to use it in live operations, and Pfizer described the aim as freeing safety professionals for signal detection and investigation rather than replacing them. No outcome figures are published.
No outcome disclosed.
Bayer
Germany · Pharma and life sciences · 2018
In November 2018 Genpact announced a multi year agreement with Bayer under which its Pharmacovigilance Artificial Intelligence (PVAI) suite, which incorporates the Genpact Cora PharmacoVigilance software product, is applied to Bayer's existing pharmacovigilance database and IT systems. Bayer's head of pharmacovigilance said the partnership offered an opportunity to further increase the efficiency of its pharmacovigilance operating model and case processing, and Genpact said Bayer was among the first companies going live with the AI based Case Management module of the PVAI suite. Neither company has published outcome figures.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historical cases with source documents and the final database entries for testing
- MedDRA and product dictionaries under licence and version control
- Written case processing conventions and validity rules
- Intake channel inventory with volumes and formats
Systems to integrate
- Safety database that accepts E2B formatted cases
- Email, call centre, web form and partner exchange intake channels
- Literature monitoring sources
- Quality management system for deviations and corrective actions
Complexity: High
Extraction from mixed documents is proven, but pharmacovigilance is a regulated, inspected process with legal reporting deadlines. The system must be validated, integrated with the safety database, and designed so that no serious case is missed or delayed.
- 1
Map intake by channel and volume
List every source of reports, its format and volume, and start with high volume, well structured sources such as email forms and partner files.
- 2
Automate validity and seriousness triage first
Deciding whether a report is a valid case and whether it is fatal or life threatening is the first phase of Pfizer's tool; it protects deadlines and gives reviewers a prioritised queue.
- 3
Extract with confidence scores
Extract fields into the E2B structure with a confidence per field, and send low confidence fields and all serious cases to a person for verification.
- 4
Validate as a computerized system
Define the intended use, test against historical cases, document performance per field and case type, and put models, prompts and dictionaries under change control.
- 5
Monitor and extend
Track missed cases, extraction accuracy and timeliness weekly, and add channels and languages one at a time.
Guardrails
- Every case the AI marks invalid or non serious is sampled by a person; no report is discarded unreviewed
- Serious and fatal cases always routed to a safety professional with the reporting deadline shown
- Medical assessment, causality and regulatory submission decisions stay with qualified staff
- Personal and health data of patients and reporters masked in logs and prompts
- Models, prompts and coding dictionaries under validation and change control
KPIs to instrument
- Share of reports processed without manual data entry
- Field level extraction accuracy on a weekly sample
- Time from receipt to case creation, and share of expedited reports submitted on time
- Missed or wrongly invalidated cases found in quality samples
- Duplicate cases detected
Human in the loop
Drug safety professionals verify low confidence extractions and every serious case, perform medical review and causality assessment, and decide what is reported to regulators. The qualified person for pharmacovigilance owns the process, and quality staff sample cases the AI screened out.
Common failure modes
- A serious case screened out
- The AI marks a valid serious case as invalid or non serious and the deadline is missed. Sample every negative decision and track misses as deviations.
- Silent extraction errors
- A wrong dose, date or product enters the database and distorts signal detection. Show confidence per field and verify key fields on every case.
- Validation that ends at go live
- Performance drifts as new products, languages and sources arrive. Monitor accuracy continuously and revalidate after changes.
- A chatbot that discourages reporting
- A public reporting assistant that is hard to use lowers reporting. Keep a direct form and a human route and test with real reporters.
What are the risks and rules?
EU AI Act
Depends on design
Internal intake, extraction and coding for review by safety staff is not listed in Annex III and is usually minimal risk. A public facing reporting assistant must tell people they are talking to an AI under Article 50. The main obligations come from pharmacovigilance law and good pharmacovigilance practices, which require validated, inspectable processes, and from GDPR rules on health data.
Rules that apply
Guidance
- Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle (European Medicines Agency, Europe). Covers AI in pharmacovigilance, including adverse event report management and signal detection, in line with good pharmacovigilance practices, and expects marketing authorisation holders to validate, monitor and document these tools.
- Good pharmacovigilance practices (GVP) (European Medicines Agency, Europe). The EU rules for collecting, managing and submitting reports of suspected adverse reactions, which an automated intake process must still meet.
- Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (draft guidance) (US Food and Drug Administration, North America). Proposes a risk based credibility assessment for AI models that produce information or data used to support regulatory decisions about safety, effectiveness or quality.
Controls to put in place
- Validation package with intended use, test results per case type and field, and acceptance criteria
- Audit trail from source document to database entry, including AI suggestions and human changes
- Deviation and corrective action process for missed or late cases
- Periodic sampling of AI negative decisions (invalid, non serious, duplicate)
- Data protection controls for patient and reporter data, including transfers to vendors
Frequently asked questions
- Which parts of pharmacovigilance can AI automate safely?
- Intake steps with clear rules: checking whether a report is a valid case, flagging seriousness, detecting duplicates, extracting data and suggesting codes. Pfizer's first phase covers basic intake decisions such as validity and whether a case is fatal or life threatening. Medical assessment, causality and reporting decisions stay with qualified people.
- Do regulators use AI on adverse event reports too?
- Yes. The US FDA lists a deployed tool that extracts data from sponsors' IND safety reports and loads it into its adverse event database, and a chatbot that helps people submit adverse event and product problem reports through its Safety Reporting Portal.
- Why are there so few published results?
- Most deployments are described by companies and vendors without measured before and after data, and for older programmes, such as Bayer's 2018 agreement with Genpact, we found no published results. Validate on your own historical cases and publish internally what the system misses, not only what it saves.
How to cite this page
Blits.ai AI Use Case Library, "AI for pharmacovigilance adverse event case intake", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/adverse-event-case-intake. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published