AI Validation & Governance in Pharmacovigilance

HUMAN OVERSIGHT

Autonomy is a dial. Accountability is fixed.

Every workflow has a defined level of autonomy, agreed with you when your department is designed. It's not guidance or policy, it's system configuration, enforced by the quality management system and recorded in the audit trail on every single action.

In the loop

Approve before it happens.

The agent assembles before the case, reasons over it, and presents a decision-ready package - its evidence and its uncertainty made explicit. A qualified human authorises; the agent cannot proceed without that signature.

APPLIED TO

Causality assessment sign-off
Expedited report submission
Labelling and RSI decisions
Signal validation outcomes

On the loop

Supervise while it happens.

High-volume work runs autonomously and continuously. Humans don't touch every case: they watch the quality signal, review the sampled output, and intervene on the exceptions the system itself flags as outside tolerance.

APPLIED TO

Case intake and triage
Duplicate detection
MedDRA coding and narrative drafting
Literature and translation screening

Human in command

Own the system itself

Humans define what the agents are allowed to be. Scope, thresholds, escalation rules, competence boundaries. Quality standards are set by named, accountable people: QPPV and deputy, QMS owner, process owners. Any workflow can be halted, narrowed, or re-qualified at any moment.

HELD BY

QPPV and deputy
Quality management system owner
Agent authorization and scope board
Named process owners

Agents do the lifting. Physicians make the calls.

Illustrative dashboard

Agents do the lifting. Physicians make the calls.

Illustrative dashboard

Agents recommend, flag and prepare; a qualified safety physician takes every decision that matters.

QUALITY

Unified quality management

One Rulebook. No exceptions for AI.

Most AI in drug safety sits outside the quality system, treated as a tool that must be separately validated. iVigee's agents operate inside it - governed by the same rules, the same qualification logic, and the same deviation processes as the human professionals they work alongside.

The result: quality improvements that are built into the system, fully traceable, and auditable - not just claimed.

Agentic capability
across all

GPPC roles

About the GPPC framework →

Leadership roles

Accountable for pharmacovigilance, strategy and the overall PV system.

Managerial roles

Manage PV functions, teams, projects and vendor relationships

Scientist roles

Case assessment, signal, risk, data science and medical writing

Technical roles

Operate and administer PV systems, data entry, coding, and quality processes

Agentic capability
across all

GPPC roles

About the GPPC framework →

Leadership roles

Accountable for pharmacovigilance, strategy and the overall PV system.

Managerial roles

Manage PV functions, teams, projects and vendor relationships

Scientist roles

Case assessment, signal, risk, data science and medical writing

Technical roles

Operate and administer PV systems, data entry, coding, and quality processes

Each agent is held to the same standard of knowledge, skill and professional competence as the human professionals it works alongside for the simple reason that quality cannot depend on whether the work was done by a person or an agent. The standard applies equally to both.

VALIDATION

Don't take our word for it.

Take the regulators.

Every agentic workflow is assessed using the same risk-based credibility framework
regulators have established for AI supporting regulatory decision-making.

Credibility is not something a system earns once. It is established for each specific context of use, with the level of evidence determined by two factors: how much the model's output influences the decision, and the consequence of getting that decision wrong.

The Credibility Assessment Framework: seven steps

Define the question of interest

State the exact question the model must help answer - not what the technology does, but what decision it supports.

Define the context of use

Set out the precise role and scope of the model, including what other evidence and human judgement sit alongside it.

Assess the model risk

Evaluate the influence of the model's output and the consequence of an incorrect decision. Together, these determine the level of evidence required.

Plan how credibility will be established

Define in advance what evidence is needed, proportionate to the risk identified in step three.

Execute the plan

Generate the evidence: performance against reference datasets, behaviour at the boundaries, and the conditions where performance degrades.

Document results and deviations

Record what was found, where reality differed from expectations, and what actions were taken.

Determine adequacy for the context of use

Decide whether the evidence supports using the model for that specific purpose. Credible for one context of use does not automatically mean credible for another.

Step 3 determines everything else:

the evidence required and the oversight mode applied.
For high-risk decisions the agent can support the process, but a human authorises every outcome. No exceptions.

Built on regulatory thinking, not AI hype.

This is not a framework created for AI marketing. It applies the same risk-based approach regulators use for computational models supporting regulatory decisions.

Book a Consultation

FDA (draft guidance)

Credibility is established for a specific context of use, with evidence scaled to risk. The guidance is currently in draft.

EMA reflection paper

Marketing authorisation holders remain responsible for validating, monitoring and documenting AI model performance within the PV system.

CIOMS WG XIV

Describes the human-first oversight models used on this page: human in command, in the loop and on the loop.

FAQ

Questions answered

Through risk-based credibility assessment per context of use: define the question and context, assess model influence and decision consequence, scale the evidence to that risk, execute, document deviations, and judge adequacy - the same logic as FDA draft guidance and CSA principles.

An oversight mode where high-volume work runs autonomously while qualified humans monitor quality signals, review sampled output and intervene on flagged exceptions - one of the humans-first models (HIC, HITL, HOTL) described by CIOMS WG XIV.

No. Credible for one context of use does not automatically mean credible for another. Each workflow carries its own assessment, and any workflow can be halted, narrowed or re-qualified at any moment.

Named people are always accountable when an agent errs: the QPPV and deputy, the QMS owner and named process owners. Every action, human or agent, is captured in a single end-to-end audit trail.

Explore Further

← OVERVIEW

Back to AI in Pharmacovigilance

←

← OVERVIEW

AI in
Pharmacovigilance

←

INSIGHTS →

Articles and Videos on AI in PV

→

INSIGHTS →

Articles and Videos on AI in PV

→

LET'S TALK

Ready to transform your pharmacovigilance department?

Book a Consultation

LET'S TALK

Put Your Question to a Pharmacovigilance Expert

Tell us what you're working on, whatever the stage or market. We'll make sure your question reaches the specialist who leads that area, and that you get a real, considered answer back, not a brochure.

Actual iVigeeans ↑

Does It Really Matter if My PV Vendor Is Physician-Led?

Before you compare vendors or talk about outsourcing, you need to understand what great pharmacovigilance leadership really looks like. Dr. Henry Seto answers the question directly, and honestly. Press play to hear what most vendors won't tell you. ↓

More pharmacovigilance insights from people who live it.

From physician-led videos to practical guides and regulatory updates, everything in our Knowledge Hub is designed to help you make confident, informed decisions.

Start learning