Why Existing Blood Establishment Rules Do Not Cover AI Agents
Blood establishment computer systems, commonly called BECS, are validated closed systems governed by 21 CFR Parts 606 and 610. These regulations require full traceability of blood products from donor to recipient and mandate that software functions perform as intended within a controlled, documented scope. That validation framework was written for deterministic software behavior. It does not address a system that can interpret instructions, choose among available actions, or invoke tools dynamically at runtime. When an AI agent is introduced to support donor eligibility screening, inventory forecasting, or specimen tracking, the organization must decide whether that agent operates inside or outside the validated system boundary. That decision determines whether existing software validation obligations apply directly to the agent, and it has direct consequences for how access, logging, and change control must be designed before any production use.
Agent Identity Distinct From the Human Technologist of Record
AABB Standards and 21 CFR Part 11 both assume that actions taken in a blood establishment system are attributable to a specific, accountable person. Role-based access in most LIS and BECS platforms is built around licensed personnel, not automated processes. An AI agent does not fit that model. Governance requires establishing a separate non-human identity class for each agent, with its own credential, scoped permissions, and audit signature, so that every action it takes can be distinguished from actions taken by a technologist. Without this separation, chain-of-custody documentation becomes ambiguous, which is a direct compliance gap rather than a theoretical concern, since traceability from donor to recipient is an explicit regulatory requirement under 21 CFR 606 and 610.
Least Privilege as a Per-Use-Case Design Decision
Not every AI agent use case in a transfusion service carries the same risk. An agent supporting inventory forecasting may only need read access to unit counts and expiration data. An agent assisting with donor eligibility screening touches data and outcomes that directly affect whether a donation proceeds, which is a regulated determination. Governance leaders should map each proposed use case to explicit data access and action boundaries before deployment, rather than granting a single agent broad access across donor records, unit disposition, and cross-match workflows. Zero Trust principles, as described in NIST SP 800-207, support treating the agent as an unverified subject requiring continuous, per-action authorization rather than a one-time session grant. This distinction matters operationally because it limits the blast radius of an agent error or a compromised agent credential to a single, defined scope of action.
Audit Trails That Meet Existing Chain-of-Custody Standards
21 CFR Part 11 requires audit trails that capture who performed an action, what the action was, and when it occurred, in a way that supports record integrity and non-repudiation. AABB Standards layer additional documentation expectations onto quality systems used for donor testing, component processing, and distribution. An AI agent's audit trail needs to meet the same bar. That means logging agent identity, the specific action invoked, the data the agent accessed, the outcome of any policy evaluation, and whether a human overrode or approved the action. Quality assurance and regulatory affairs teams should confirm early that agent-generated logs can be reconciled with existing chain-of-custody documentation, rather than treated as a separate, parallel record that auditors have to interpret independently.
Human Accountability Remains With Licensed Personnel
None of the regulatory frameworks referenced here, including FDA's blood establishment regulations, 21 CFR Part 11, or AABB Standards, contemplate autonomous AI decision-making replacing licensed personnel in donor eligibility or transfusion decisions. FDA's 2024 draft guidance on AI-enabled device software functions emphasizes lifecycle management and transparency of AI system logic, reinforcing that oversight and accountability structures need to persist even as AI components are introduced. Governance models for blood bank AI agents should therefore preserve human sign-off for regulated determinations. The agent can support and accelerate work such as flagging inventory shortfalls or surfacing donor history for review, but the final action on donor eligibility, unit disposition, or cross-match interpretation should remain a documented human decision, with the agent's role clearly distinguished in the record.