IQVIA

The agent platform is a product of its own inside a business that sells much more. IQVIA.ai, launched in March 2026, is a platform and catalog through which customers deploy the company's agents and AI capabilities: literature review, the Med-R1 Deep Research Agent, a field force agent, commercial planning, cohort building, market insight and master data agents, with connectors into IQVIA data and customer systems. Those capabilities do not exist without the models. The company's larger business, clinical research services, health data licensing, consulting and trial technology, carries value without them, and most buyers come to IQVIA for that.

Human oversight is asserted across the portfolio; no agent's limits are published. The platform page says agent orchestration runs complex workflows while maintaining human oversight, the governance page keeps experts in the loop, and the agentic AI page describes agents that will reason, advise and act based on customer guidelines, while defining agentic AI as systems that act without constant human supervision.

The catalog lists each agent's availability but not what it may do unattended, what a reviewer sees before approving, or how an agent's action is distinguished from a person's. The company reports more than 150 agents deployed across its own teams and client environments.

The stack and one in-house model are named; versions and update practice for the agents are not. IQVIA.ai combines the company's data and models with NVIDIA Nemotron, the NeMo Agent Toolkit, NeMo Customizer for fine-tuning, NeMo Guardrails, Dynamo and LangChain. The company's own Med-R1 8B medical reasoning model is named with its size and benchmark results, and the literature agents are described as proprietary and domain-tuned.

The governance page says model and code are documented and source-controlled and datasets versioned. Which model runs each agent, what data tuned it, and how customers learn of a change are not published.

The model layer is named in part; no party list or data terms are published. The platform names NVIDIA Nemotron models, NeMo tooling for fine-tuning and guardrails, Dynamo and LangChain, and the company's own Med-R1 8B model; its agentic AI page names Microsoft as a partner. Fine-tuning on top of third-party base models is stated rather than left implicit. No sub-processor list is published for IQVIA.ai, and nothing states which parties receive customer data, or the retention and training terms that apply there.

Published research and benchmark results, with adoption breadth and little outcome data for the agents. The company reports more than 200 AI publications and publishes Med-R1 8B's scores on MedQA, PubMedQA and MedMCQA, which it says outperform other 8 billion parameter models.

For the agents, it reports more than 150 deployed across internal teams and client environments and 19 of the top 20 pharmaceutical companies having begun to use them, a bar that spans pilots through production, and a literature review time cut of over 75 percent without a method. No agent has a published outcome study.

Privacy technology is claimed and one safeguard is named; what reaches the models, and whose data trains them, is not stated. The company says it uses privacy enhancing technologies and anonymization at scale, and its 2025 agent release names NVIDIA NeMo Guardrails to safeguard deployment. IQVIA.ai is described as learning continuously from operational feedback through a data flywheel built with NVIDIA, and agents as drawing on experience gained across thousands of use cases.

Nothing published says whether a customer's prompts, outputs or feedback train models serving other customers, what is retained and for how long, or which anonymization standard applies to the data the agents read.

HIPAA is named in the security framework; the company's role for the agent platform is not stated. The annual report says its information security framework draws on HIPAA, describes the obligations of covered entities and business associates, and says the company is an industry leader in anonymization, with the data it holds anonymized.

None of this says whether IQVIA.ai processes protected health information, whether a business associate agreement is available for it, or which de-identification method underpins the data the agents read; re-identification risk is a property of linked data, which agents reasoning across sources make live.

A recognized attestation is named in the company's own filing, with its scope a question to raise. The 2025 annual report (Item 1C) says the global data centers and IT controls sit inside an annual SOC 2 Type II attestation by an independent audit firm, with controls aligned to ISO/IEC 27001, under an information security framework drawing on NIST, GxP, HITRUST, the ISO 27000 family, COBIT, GDPR and HIPAA, and that independent consultants assess program maturity at least annually. There is no trust center, and no report or certificate is published. Whether IQVIA.ai and client-stack deployments fall inside the SOC 2 scope is not stated.

No device claim is made for the agent platform, and it is scoped to research, commercial and operational work. Where agent output feeds a regulatory filing, such as clinical data review or trial master file work, electronic records and computerized system validation rules apply.

The annual report says the company's security framework draws on GxP guidance among others, and its research services operate under those rules, but no position is published for the agents: whether they sit inside a validated boundary, how a non-deterministic component is qualified, or what happens when an underlying model changes.

A governance framework with named process is published; results for individual agents are not. The AI governance page sets out five principles (fairness, transparency, respect, accountability, auditability), expert review of AI systems before use, a stated aim to use only outputs with high confidence in precision and supporting evidence, and safeguards that include documentation for algorithms, models and code, source-controlled code, versioned datasets, deployment through approved systems and logged activity.

The company links a paper on applying healthcare AI governance, cites the EU AI Act and Colorado AI Act in its annual report, and reports more than 200 AI publications. Fairness is described as constant evaluation and mitigation of bias, but no per-agent evaluation or results across groups are published.

Outputs are meant to be traceable and reviewed; nothing stands behind them. The governance page says the company aims to use only outputs accompanied by high confidence in precision and justified by supporting evidence, keeps experts in the loop, and logs activity with documented, source-controlled models; the literature and deep research agents return citation-grounded answers.

No error rate is published for any agent: the figures are time (literature review cut by over 75 percent) and, for the Med-R1 8B model, scores on public medical question benchmarks (MedQA, PubMedQA, MedMCQA), which test the model rather than an agent's output in use. "Healthcare-grade AI" is a registered trademark rather than a described standard. No warranty, indemnity or remediation commitment is published.

No clinical record integration, and the customer systems the agents connect to are not named. The systems that count here are the ones its buyers run. IQVIA.ai lists MCP connectors to the company's own sources (MIDAS, ChannelDynamics and other market data, real world cohorts, a scientific content index, clinical code lists) and to customer systems described only by category: trial master file systems and practitioner and organization data. No named trial management, data capture or master file product, interface documentation or production reference is published.

Two deployment paths are named; where either runs is not. The platform page says IQVIA.ai capabilities can deploy in IQVIA's technology stack or the client's, and the company reports agents running in its own teams and in client environments. No hosting region, residency option, tenancy separation, retention period or sub-processor list is published for the platform, and the stack draws on NVIDIA software whose serving location is not stated. The annual report says the company's global data centers sit inside an annual SOC 2 Type II program.

No price is published, and how the agents are sold is discoverable. Enterprise agreements with life sciences organizations usually sit inside existing data, services and clinical research relationships. IQVIA.ai presents a catalog with each agent's availability (now, next quarter or launch-ready) and says some capabilities deploy in IQVIA's or the client's technology stack.

Coverage is named and validated for part of it. IQVIA.ai spans clinical, commercial and real world work, and its catalog marks what is available now (literature review, deep research, field force, market insight, several connectors) against what is launch-ready or due later (commercial planning, cohort and master data agents). Published support covers the literature agent's time savings and Med-R1's benchmark scores; the other agents have adoption claims only. Evaluate against what is available now for your function.