• +1 (602) 922-5990
  • This email address is being protected from spambots. You need JavaScript enabled to view it.
  • Mon - Fri 8:00a - 5:00p AZ
Industry Insights banner: Third-Party Risk and AI Governance in Life Sciences, with a research facility and molecular network.

Third-Party Risk and AI Governance in Life Sciences: Keeping Critical Testing and Manufacturing Running

A life-sciences organization can have validated equipment, approved procedures, and a quality unit and still be unable to release a batch, complete a critical test, or keep a multi-site laboratory network running if a contract manufacturer, a specialized testing lab, a cloud laboratory system, or a model used in development or quality cannot perform. In that moment the problem is not only a supplier-management finding. It is a continuity problem, a quality problem, and, where models are in the path, an AI-governance problem.

Life sciences is not a single regulatory silo. Drug and biological product manufacturers, blood and biologics laboratories, and medical-device firms live under different Food and Drug Administration frameworks, and many also sell or operate in the European Union. This article does not treat those frameworks as interchangeable, and it does not invent client or facility examples. It asks how third-party risk and AI should be governed so that testing, manufacturing, and release can continue when a partner or a model fails.

Why this matters

The critical path in life sciences often runs through organizations the sponsor or laboratory does not employ. Contract manufacturing organizations, contract research organizations, specialty and reference laboratories, calibration and metrology providers, cold-chain and logistics firms, and cloud-hosted laboratory information or quality systems all sit between raw material or specimen and a result, a batch, or a regulatory submission. Blood and biologics laboratories, and other multi-site testing networks, add a further dependency: the ability to shut down, transfer, and restart testing without losing continuity of critical services.

Those relationships are usually managed as quality-supplier or vendor files. That work is necessary. It is not the same as knowing which testing or manufacturing services cannot stop, which partners and systems those services depend on, how long a disruption can be tolerated, and what a controlled restart looks like. Shutdown and restart are themselves quality events. They need pre-planned steps, not an improvisation written after the site is already dark.

Where a laboratory or manufacturer creates, receives, maintains, or transmits electronic protected health information, the HIPAA Security Rule may also apply, including the contingency-plan standard at 45 CFR 164.308(a)(7). That is fact-specific. It should not be assumed for every life-sciences operation, and it should not be ignored where ePHI is actually in scope. Availability of electronic protected health information, where the rule applies, is an official requirement, not only a customer expectation.

Third parties on the critical testing and manufacturing path

A useful starting point is to name the services whose interruption would stop release, delay a clinical or development decision, or interrupt a critical testing service, and only then to name the partners. A single-source contract manufacturer, a unique assay performed at one reference lab, or a cloud laboratory-information system with no current alternate path is a concentration risk, even if the quality file is complete. Alternate capacity, specimen or material routing, and data recovery need to be designed while the relationship is healthy. For a broader discussion of proving third-party resilience, shared concentration, and what to do when a critical provider cannot recover, see Beyond Due Diligence: Proving Third-Party Resilience, The Hidden Risk of Concentration, and When a Critical Provider Cannot Recover.

Contracts and quality agreements should be readable as continuity documents as well as quality documents. They should address notification of deviations and incidents, access to information needed to investigate and to restart, data return and portability, business-continuity and disaster-recovery expectations, and the right to exercise or test those expectations where the service is critical. A quality audit that never asks whether the partner can recover, and a continuity plan that never asks whether the partner is in a state of control, are two incomplete views of the same relationship.

Multi-site laboratory networks have a particular problem of shared platforms and shared partners. If several sites depend on one laboratory-information system, one courier network, or one specialty vendor, a local plan at each site will not describe the network's true impact tolerance. Mapping those dependencies before disruption is discussed more fully in Mapping Dependencies Before Disruption.

AI in development, quality, and devices

Life-sciences organizations are using AI to support development decisions, manufacturing and quality analysis, image or signal review, and, in some cases, medical-device software. Some of that work is internal. Much of it is supplied or hosted by vendors, or depends on vendor data and compute. The official US starting point for drugs and biologics is still draft. On January 6, 2025, the Food and Drug Administration issued draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. The FDA page describes it as draft guidance, not for implementation, containing non-binding recommendations. It offers a risk-based credibility-assessment framework for establishing and evaluating the credibility of an AI model for a particular context of use when the model's output is intended to support a regulatory decision about safety, effectiveness, or quality. Because it is draft, organizations should not treat it as a final requirement. Because it describes how the agency has been thinking about model credibility, sponsors and partners who use AI in that context should read it carefully and watch for a final version.

FDA also issued draft recommendations in January 2025 on lifecycle management and marketing submissions for AI-enabled device software functions. That draft is directed to device developers and to FDA staff. It is still useful reading for a company that places or depends on such software, because it emphasizes lifecycle control, not a one-time clearance narrative. It is not, by itself, an operational-resilience program.

The National Institute of Standards and Technology published the Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, on January 26, 2023. It is voluntary and organized around Govern, Map, Measure, and Manage. It describes trustworthy AI as valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair, with harmful bias managed. Those characteristics are a practical bridge between a development-quality discussion and a continuity discussion: a model that cannot be monitored, explained, or withdrawn safely is not a resilient dependency, whatever its performance was in a validation set.

ISO/IEC 42001:2023, published in December 2023, is the first international management-system standard for AI. It specifies requirements for establishing, implementing, maintaining, and continually improving an AI management system in organizations that provide or use AI systems. Certification is voluntary and is not performed by ISO. For a sponsor or laboratory network, the standard can put internally developed models and vendor-supplied models under the same roles, risk assessment, and lifecycle controls, including change management.

Organizations that place or deploy AI in the European Union also need to account for Regulation (EU) 2024/1689, the EU Artificial Intelligence Act. The Act entered into force on 1 August 2024 and applies in phases. Most remaining obligations have applied since 2 August 2026, with some high-risk AI rules applying later. Medical devices and certain other product-safety frameworks are part of how the Act treats higher-risk systems, and some of those dates are later than the general August 2026 application. Whether a particular model or device is in scope is a legal and factual question. The operational questions do not wait on that classification: what is the context of use, what is the evidence that the model is credible in that context, who can change it, and what happens to testing, manufacturing, or clinical use if it is unavailable or withdrawn?

Where the three issues meet

The meeting point is a service the organization cannot pause without affecting patients, donors, or a regulated product: a critical test, a manufacturing step, a release decision, or a device function. The contract manufacturer's incident, the reference lab's outage, or the withdrawal of a model used to support a quality or regulatory decision becomes the sponsor's or laboratory's continuity and quality problem at once.

That is why AI governance should be written into supplier and quality-agreement practice, and why continuity plans should name models as well as machines and sites. Context of use and credibility are not only submission topics. They are also ways to decide whether a model is critical, whether a second method exists, and whether the model can be taken out of service without losing control of the process. Exercises should include the loss of a contract site, a specialty lab, a laboratory-information system, or a model used in a quality or development workflow, and they should include a controlled restart. Findings should have owners and dates, and they should feed both the quality system and the continuity program.

See also Connecting AI Governance and Resilience and From AI Principles to Accountability.

What organizations should do

Identify the testing, manufacturing, and release services that cannot stop, including critical testing services in blood and biologics or other laboratory networks. Map the contract manufacturers, laboratories, logistics providers, cloud systems, and models that support those services. Treat single-source and single-path dependencies as board-visible issues.

Make quality agreements and vendor contracts carry continuity terms: incident and deviation notification, data return, recovery expectations, and the right to test those expectations where the service is critical. Plan shutdown and restart in advance.

Where ePHI is in scope, align the contingency plan with 45 CFR 164.308(a)(7). Where it is not, do not invent a HIPAA duty, and do not leave availability unaddressed.

Govern AI with a documented context of use, a credibility or performance assessment proportionate to the decision it supports, human oversight, monitoring, and a discontinuation plan. Read the January 2025 FDA draft guidances as drafts. Use the NIST AI RMF and, where a management-system frame is useful, ISO/IEC 42001. Do not claim conformity or certification you have not independently established. If you place or deploy AI in the EU, obtain a current legal reading of Regulation (EU) 2024/1689 for the specific use.

Give quality, operations, and executive leaders one picture of critical partners, critical models, test results, and open issues. See also From Plans to Proven Readiness and Operational Resilience: Where to Start.

How OpResONE can help

OpResONE helps life-sciences organizations strengthen operational resilience, third-party oversight, and AI governance as a single management problem, including mapping critical testing and manufacturing dependencies and designing exercises that include shutdown, restart, and partner or model failure. Our work is independent. We have no affiliation with FDA, HHS, NIST, or ISO, and this article is not an official interpretation of their documents.

We bring BRILLIANCE to RESILIENCE®.

Related resources

Request a Strategy Call

If you want a focused reading of how supplier and AI oversight lines up with the continuity of critical testing and manufacturing services, talk with OpResONE about where to start.

Request a Strategy Call

Sources