VITAL: A Framework for Validating Agentic AI in Pharmaceutical Drug Development and Clinical Trial Operations
Organizations: AstraZeneca Boston. 35 Gatehouse Drive, Waltham, MA 02451, United States · AstraZeneca Barcelona. Avinguda de Roma, 81, 08029 Barcelona, Spain · AstraZeneca Cambridge. Francis Crick Avenue Cambridge Biomedical Campus Cambridge, CB2 0AA, United Kingdom · NVIDIA. 2788 San Tomas Expy, Santa Clara, CA 95051, United States ·
Abstract
Current assurance approaches do not establish whether evidence is sufficient to validate, monitor and govern agentic artificial intelligence (AI) within a pharmaceutical context of use. Plausible outputs may conceal failures in retrieval, reasoning, orchestration or human-control points, where reference standards are unavailable. We developed and operationalised VITAL (Validation by Intended use, Traceability, Autonomy and Lifecycle evidence), a validation framework for agents in clinical development, through landscape analysis, regulatory and standards mapping, expert interviews and multi-stakeholder refinement. VITAL integrates context of use, controls, behavioural validation and lifecycle evidence within a risk-adjusted framework. We applied VITAL across epidemiology and treatment-landscape analysis, site selection, enrolment rescue and trial-design trade-off support to translate autonomy, reversibility, transparency and patient impact into proportionate validation activities, evidence requirements and auditable reconstruction of system behaviour. VITAL may inform evidence requirements and lifecycle accountability for higher-risk pharmaceutical uses; evaluation should assess reproducibility across assessors, organisations and system configurations.
Figures & tables
| Base agentic risk ( ) | Agent autonomy | Risk context | Risk definition | Examples |
|---|---|---|---|---|
| 1 | Informational | Retrieve | Static information access, no synthesis, judgement or prioritisation | Search protocol text, retrieve information |
| 2 | Informational | Summarise | Synthesis or formatting without evaluative judgement | Summarise protocol operational requirements, extract visit schedule, synthesise interpretation |
| 3 | Informational to analytical | Surface signals | Descriptive patterns or trends without prioritisation or recommendations | Surface descriptive protocol burden indicators (intensive visit schedule, many procedures/assessments, tight visit windows) |
| 4 | Analytical | Apply rules | Deterministic rule-based scoring or classification, transparent thresholds or rubrics | Burden scoring rubric, classify feasibility risk via fixed thresholds |
| 5 | Advisory | Recommend | Advisory decision support, human decides, outputs non-binding, explainable rationale required | Recommend feasibility mitigations, rank endpoint candidates |
| 6 | Advisory (contextual) | Recommend with context | Context-aware advisory, adapts recommendations using multi-factor signals, still non-binding | Recommend and rank sites using contextual operational signals (geography, start-up timelines, staffing/capacity, competing trials) and propose enrolment rescue options |
| Eight control requirements | Description |
|---|---|
| Quality and governance system | The organisational structures, roles, responsibilities, procedures and escalation pathways through which the system is designed, approved, deployed and overseen over its lifecycle. |
| Data governance and integrity | The controls and processes ensuring the provenance, quality, representativeness, privacy, access management, and lifecycle integrity of the data used for prompting, retrieval, system outputs and evaluation. |
| Agent design controls | The independently enforced technical and procedural controls that bound effective autonomy, including least-privilege permissions, tool and network access, credential isolation, blocked actions, stopping rules, escalation conditions and fail-safe interruption. |
| Verification and validation approach | The structured testing strategy used to demonstrate that the system is built as intended (verification) and performs acceptably within its intended context of use (validation), including methods, test design and acceptance criteria. |
| Traceability of system behaviour | The ability to reconstruct system behaviour after the fact, including logging of inputs, retrieved sources, tool calls, outputs, versions and decision traces. |
| Operational monitoring and change control | The processes used to monitor system performance during operational use, detect degradation or drift and manage changes, including criteria and triggers for revalidation. |
| Five behavioural validation domains | Description |
|---|---|
| Reasoning and planning | Ability of the system to decompose objectives into coherent and logically ordered steps. |
| Tool use and execution | The correctness, safety, and appropriateness of interaction with external systems, including whether tool calls remain authorised and whether the agent attempts to expand privileges, evade restrictions or exploit its execution environment. |
| Retrieval augmented generation (RAG) | Whether the system retrieves, synthesises and cites relevant information from authorised sources whilst resisting indirect prompt injection and unsupported claims. |
| Single and multi-agent coordination | Assessment of single-agent performance and the integrity of inter-agent hand-offs, including message provenance, authentication, trust boundaries, replay protection and containment of a compromised component. |
| Human-agent interaction | Evaluation of human oversight and interpretability, including whether approval, pause and termination controls remain effective across extended trajectories. |
| Ten evidence categories | Description |
|---|---|
| Context-of-use specification | Describes the intended use, user population, workflow context, permitted data sources, decisions, and system boundaries, providing a clear statement against which validation is assessed. |
| Governance operating model | Defines accountability for design, approval, deployment, exception handling and oversight, including the decision gates, escalation pathways, and the allocation of human-in-the-loop or human-on-the-loop control responsibilities. |
| Data specification | Demonstrates the provenance, quality, representativeness, access controls and lifecycle handling of the data used for prompts, retrieval, outputs, and evaluation with sufficient detail for an independent reviewer to assess fitness for purpose. |
| Design specifications and controls | Describes the system architecture and behavioural constraints, including agent roles, tool access, permissions, blocked actions, escalation triggers, and the guardrails used to prevent unsafe or unauthorised behaviour. |
| Validation and revalidation protocol | Sets out the validation test strategy, including test methods (e.g., retrospective checks, benchmarking against reference datasets, simulation or adversarial testing, subject matter expert review), acceptance criteria, and the rules for pass/fail decisions, deviations, and revalidation triggers. |
| Validation evidence package | Contains the structured set of outputs generated from validation activities, including test results, performance summaries, expert assessments, deviation records, and conclusions that demonstrate that the system meets predefined acceptance criteria for its intended use. |
| Calibration | Reversibility | Transparency | Patient impact |
|---|---|---|---|
| Baseline | 1.0, 1.25, 1.5 | 1.0, 1.25, 1.5 | 1.0, 1.25, 1.5, 3.0 |
| Moderately expanded | 1.0, 1.35, 1.55 | 1.0, 1.35, 1.55 | 1.0, 1.35, 1.55, 3.0 |
| Compressed | 1.0, 1.1, 1.25 | 1.0, 1.1, 1.25 | 1.0, 1.1, 1.25, 2.0 |
| Use case | Baseline calibration | Moderately expanded calibration | Compressed calibration |
|---|---|---|---|
| Epidemiology and treatment-landscape analysis | 8 1.0 1.0 1.25 = 10.0 - High | 8 1.0 1.0 1.35 = 10.8 - High | 8 1.0 1.0 1.1 = 8.8 - Moderate |
| Site selection and prioritisation | 5.5 1.0 1.25 1.0 = 6.9 - Moderate | 5.5 1.0 1.35 1.0 = 7.4 - Moderate | 5.5 1.0 1.1 1.0 = 6.1 - Moderate |
| Enrolment forecasting and rescue recommendation | 6 1.25 1.25 1.25 = 11.7 - High | 6 1.35 1.35 1.35 = 14.8 - High | 6 1.1 1.1 1.1 = 8.0 - Moderate |
| Trial-design trade-off support | 9 1.5 1.5 1.5 = 30.4 - Very High | 9 1.55 1.55 1.55 = 33.5 - Very High | 9 1.25 1.25 1.25 = 17.6 - High |
| Source category | Sources, n | Proportion |
|---|---|---|
| Peer-reviewed journal and conference publications | 72 | 63.7% |
| Regulatory guidance and standards | 15 | 13.3% |
| Preprints | 14 | 12.4% |
| Grey literature | 12 | 10.6% |
| Total | 113 | 100% |
| Validation challenge class | Sources, n | Proportion |
|---|---|---|
| Model drift and continuous learning | 98 | 87% |
| Integration with regulated workflows | 94 | 83% |
| Documentation and evidence artefacts | 92 | 81% |
| Data access controls and confidentiality | 92 | 81% |
| Scalability, workflow fit and adoption | 89 | 79% |
| Hallucination and factuality | 57 | 50% |
| Functional dimension | Sources, n | Proportion | Evaluation focus |
|---|---|---|---|
| Reasoning and planning | 96 | 85% | Completeness, ordering and appropriateness of planned actions |
| Grounding and retrieval | 93 | 82% | Relevance, provenance, synthesis and attribution of retrieved evidence |
| Human interaction and governance | 90 | 80% | Oversight, interpretability, interruption, escalation and safe use |
| Tool use and execution fidelity | 52 | 46% | Validity, authorisation and intended alignment of external actions |
| Multi-agent coordination | 40 | 35% | Hand-off integrity, conflict resolution and collective behaviour |
| Topic domain | Scope of inquiry | Relevance to framework development |
|---|---|---|
| Orientation to agentic AI in clinical development and clinical operations | Expert perspectives on perceived opportunities, uncertainties and concerns associated with agentic AI in clinical development and clinical operations. | Established the contextual assumptions, perceived value propositions and risk concerns that shaped subsequent exploration of validation requirements. |
| Operational pain points and candidate use cases | Workflow bottlenecks and candidate applications in which agentic AI could plausibly support complex, repeatable or multi-step tasks, including feasibility assessment, site selection, recruitment, monitoring, trial master file processes, query management and safety triage. | Informed the framework’s use-case-specific orientation and reinforced the need to define validation relative to an intended workflow context. |
| Distinction between agentic AI and conventional analytics | Characteristics distinguishing agentic systems from dashboards, predictive models or conventional analytics, including planning, retrieval, tool use, task coordination, recommendations and bounded action. | Supported the framework’s focus on agentic behaviour as the object of validation, including reasoning, planning, tool invocation, workflow boundaries and human-control points. |
| Use-case suitability and generalisability | Features that make a use case suitable for framework application, including generalisability across organisations or therapeutic areas, feasibility of test design, availability of metrics and ability to construct an evidence package. | Contributed to the selection and framing of representative validation scenarios and clarified the need for framework components that can accommodate heterogeneous pharmaceutical workflows. |
| Failure modes and delegation boundaries | Potential failures associated with agent involvement, including silent errors, inappropriate autonomy, cumulative error propagation, unsafe escalation, regulatory consequences, patient-safety implications and activities unsuitable for delegation. | Informed the framework’s treatment of intrinsic risk, autonomy limits, prohibited actions, escalation criteria and risk-proportionate validation depth. |
| Confidence and evidence requirements | Evidence required by functional, technical or governance stakeholders before accepting an agentic system in a production or operationally sensitive workflow, including test scenarios, acceptance criteria, auditability, reproducibility and reviewability. | Informed the evidence-layer structure of the framework, including validation protocols, evidence packages, audit trails and traceability requirements. |
| Record element | Purpose | Minimum information to retain |
|---|---|---|
| Validation disposition | Defines precisely what has been authorised or not authorised | The resulting disposition and its scope, including any restriction on users, workflows, decisions, data, tools or downstream actions |
| Decision claims | Identifies the propositions on which the disposition depends | The small set of claims that must be true for the disposition to remain justified with references to the evidence supporting each claim |
| Decision-critical evidence | Distinguishes evidence that influenced the judgement from evidence that was merely available | The principal test results, expert judgements, control evidence or other findings that materially affected the disposition referenced to their sources |
| Adverse or conflicting evidence | Prevents unfavourable findings from being obscured by an aggregate conclusion | Material failed tests, contradictory findings, unresolved deviations or expert disagreement, together with their effect on the final judgement |
| Uncertainty and assumptions | Makes explicit where the decision relies on incomplete knowledge or judgement | Material uncertainties, assumptions or areas in which available evidence was insufficient to provide a definitive conclusion and the consequence of that uncertainty for use |
| Decision conditionality | Identifies controls or operating conditions on which acceptability depends | Controls, restrictions, human interventions or environmental conditions whose continued effectiveness is necessary for the disposition to remain valid |
| Approach | Characteristic assurance evidence | Potential evidentiary use within VITAL | Conditions affecting applicability | Boundary of interpretation |
|---|---|---|---|---|
| VITAL | Context-of-use and risk assessment, documented control state, behavioural validation findings and lifecycle evidence | Provides the integrated evidence basis for determining acceptability within the specified pharmaceutical use | Evidence should correspond to the defined system boundary, validated configuration and intended use | A VITAL assessment does not replace specialised technical testing or establish compliance with external frameworks not explicitly assessed |
| CHAI Agentic AI Testing and Evaluation Framework | Testing and evaluation results generated against healthcare-oriented agent assessment criteria | May provide supporting evidence for relevant behavioural or human-agent interaction claims | Applicability depends on concordance of the evaluated task, system configuration, workflow conditions and acceptance criteria with the VITAL use case | Successful performance under the original evaluation conditions does not independently establish the VITAL risk classification or adequacy of the complete pharmaceutical validation evidence package |
| VECTR | Reliability-oriented evidence concerning agentic AI in drug-development applications | May provide supporting evidence for reliability, evidentiary integrity or traceability claims relevant to behavioural validation and documentation | The assessed agent behaviour, evidence sources and system state should remain representative of the configuration under VITAL assessment | Reliability evidence for a particular behaviour or workflow does not independently establish that all context-specific controls or validation requirements have been satisfied |
| CSV/CSA/GAMP 5-based assurance | Requirements, verification and testing records, traceability, controlled-change documentation and associated software-assurance evidence | May provide evidence for foundational system controls, deterministic components, data integrity, configuration management and lifecycle change | Evidence should remain applicable to the components and configuration incorporated within the agentic system and should reflect their current validated state | Component or software assurance does not independently demonstrate acceptable non-deterministic agent behaviour across alternative execution trajectories |
| Abbreviation | Definition |
|---|---|
| API | Application programming interface |
| CHAI | Coalition for Health AI |
| CSA | Computer Software Assurance |
| CSV | Computer System Validation |
| EMA | European Medicines Agency |
| EU | European Union |
| Topic | Item No. | Guide Questions/Description | Reported on / Page No. |
|---|---|---|---|
| Domain 1: Research team and reflexivity | |||
| Personal characteristics | |||
| Interviewer/facilitator | 1 | Which author/s conducted the interview or focus group? | Methods, Subject matter expert interviews, pp. 20–21 |
| Credentials | 2 | What were the researcher’s credentials? E.g. PhD, MD | Methods, Subject matter expert interviews, pp. 20–21 |
| Occupation | 3 | What was their occupation at the time of the study? | Methods, Subject matter expert interviews, pp. 20–21 |
| Gender | 4 | Was the researcher male or female? | Not disclosed |
AI Agent Governance
AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.
See all 472 papers