Answers · By Hugh Donatello, Atlas Bio · Updated 2026-10-30

What Is an AI Governance Control Plane for Agents in Life-Science R&D?

What is an AI governance control plane for agents in life-science R&D? An AI governance control plane is the layer that holds the rules for every AI agent in one place and sends them to the points where agents act. It covers six things: which agent is acting, which models it may use, which data it may read, which tools it may call, which decisions need a person, and what evidence is kept. Enforcement stays distributed in identity, gateways and the systems themselves.

A control plane is an architecture pattern, not a single product

The term comes from networking, and NIST uses it in its zero trust architecture. NIST SP 800-207, Zero Trust Architecture (August 2020), splits access control into a policy decision point, made up of a policy engine that decides whether a subject may reach a resource and a policy administrator that carries the decision out, and policy enforcement points that enable, monitor and terminate the connection. In NIST's words, the logical components use a separate control plane to communicate, while application data is communicated on a data plane.

Applied to AI agents, the split is useful with one adaptation. The control plane is where the rules for agents are written, versioned and approved. The data plane is where agents actually work: model calls, reads from an electronic lab notebook (ELN) or laboratory information management system (LIMS), tool calls, draft reports. In SP 800-207 the policy engine decides each access request. For agents, many rules can instead be pushed to enforcement points in advance, while high-impact actions are still referred for a decision at the time they happen. Either way, the control plane defines the policy each enforcement point should apply and is checked against what those points report. Products sold as an AI or agent control plane implement parts of this pattern.

This page is about that architecture. Atlas Bio's companion answer on AI agent governance in life sciences covers the operating model: who owns each agent, how it is reviewed and when it is switched off. Governance decides the rules; the control plane is how those rules reach every agent consistently. This page is not legal or regulatory advice.

The six dimensions a control plane governs

An agent in an R&D organization does six kinds of things that need a rule. Each row below names what central policy decides, where it is enforced, and the closest anchor in a published framework or regulation. NIST describes the AI Risk Management Framework (NIST AI 100-1) as voluntary; the GxP anchors apply only where the records involved are regulated.

DimensionWhat central policy decidesWhere it is enforcedAnchor
IdentityEach agent has its own identity, a named owner and a defined scope; no agent borrows a person's accountIdentity provider (non-human identities, service principals), credential issuanceNIST AI RMF GOVERN 1.6 (inventory) and GOVERN 2.1 (roles); 21 CFR 11.10(d) limits system access to authorized individuals, so each agent's access should trace to an accountable person
ModelsWhich models and versions are approved for which use, and under which vendor termsModel endpoint or AI gateway; cloud tenant configurationGOVERN 6.1 (third-party AI risks); MAP 4.1 (risks of third-party components)
DataWhich data classes an agent may read or send to a model, such as unpublished sequences, assay results or clinical dataData-source permissions, data-loss controls, gateway rulesEU GMP Annex 11 (2011) §12.1 (logical controls restricting access to authorised persons)
ToolsWhich tools and actions an agent may call, with what scopeTool allowlists, scoped credentials, authorization in the downstream systemOWASP LLM06:2025 Excessive Agency (minimise extensions and permissions; complete mediation)
DecisionsWhich actions need a person before they happen, and who that person isApproval step in the workflow or the downstream systemMAP 3.5 (human oversight); OWASP LLM06 (require user approval); 21 CFR 11.10(f) operational checks and 11.10(g) authority checks where regulated records are involved
EvidenceWhat is recorded for each agent action, how it is protected and how long it is keptLogging at each enforcement point, collected in one storeMANAGE 4.1 (post-deployment monitoring). Supports, but does not replace, the audit trail of the system holding a regulated record (21 CFR 11.10(e); EU GMP Annex 11 (2011) §9)

Central policy, distributed enforcement

The design principle is that rules are written once and enforced where the action happens. OWASP's entry on Excessive Agency (LLM06:2025) puts the enforcement half plainly: implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed, and enforce the complete mediation principle so that all requests made to downstream systems via extensions are validated against security policies. A model that has been told not to change a LIMS record is not a control; a LIMS that refuses the change from that agent's identity is.

The policy half is what many organizations lack. Without a single place where agent rules live, each team configures its own assistant, each tool keeps its own allowlist and nobody can say which agents may touch which data. NIST AI RMF GOVERN 1.4 asks that the risk management process and its outcomes be established through transparent policies, procedures and other controls. A control plane is one way to make that true for agents: one policy record per agent, versioned, with each enforcement point configured from it.

Distributed enforcement also means deciding what happens when one component fails. If the policy store is unavailable, enforcement points should fail closed for high-impact actions rather than fall back to open access. That is a design choice to write down, test and keep under change control, because a silent fallback to open access defeats the policy without leaving a record.

In GxP work the decisions and evidence dimensions carry the most weight. 21 CFR 11.10(g) calls for authority checks to ensure that only authorized individuals can use the system, electronically sign a record, access the operation or computer system input or output device, alter a record, or perform the operation at hand, and 21 CFR 11.100(a) says each electronic signature shall be unique to one individual. An agent can prepare, route and flag. The policy should grant no agent identity signing or approval rights, and the regulated system itself should refuse them.

Logs gathered by the control plane do not replace the record system's own audit trail. 21 CFR 11.10(e) calls for secure, computer-generated, time-stamped audit trails that independently record the date and time of operator entries and actions that create, modify or delete electronic records, and EU GMP Annex 11 §9 asks that a system-generated audit trail of GMP-relevant changes and deletions be considered on the basis of a risk assessment. Where an agent writes to a regulated record, the LIMS, ELN or quality system should attribute the change to the agent's identity, traceable to its accountable owner; the control plane's logs add what the record system cannot see, such as the model, prompt context and tool calls behind the change. How FDA applies Part 11, including its audit trail provisions, is set out in its 2003 guidance on Part 11 scope and application together with the applicable predicate rules.

How it differs from an AI gateway, IAM and a SIEM

Each of these tools is an enforcement point or a data source for a control plane. None of them is one on its own.

An AI gateway sits in the path of model requests. It can allow, transform, keep local or block a request and log it, which covers the models dimension and part of the data dimension. A model gateway also sees the tool calls a model asks for, and some gateways carry the tool traffic itself. Neither replaces authorization in the system being called, and neither decides who owns the agent.

Identity and access management (IAM) issues and checks identities. NIST SP 800-207 describes the ID management system as responsible for creating, storing and managing enterprise user accounts and identity records. IAM gives each agent an identity and grants, and it can enforce some agent rules, such as access to a specific model deployment or approval before a privileged role is used. It is rarely where the reasons are recorded: why a model was approved, which outputs need review, or what evidence an agent must leave.

A security information and event management (SIEM) system collects events. SP 800-207 describes it as collecting security-centric information for later analysis, used to refine policies and warn of possible attacks. A SIEM is where evidence from enforcement points can be correlated. It does not block an individual action before it happens; with automated response it can suspend an agent's identity after an alert, which stops later actions.

The control plane is the layer above these: the place that says what each of them should enforce for each agent, and the place that can show, for any agent on any date, which policy applied. In a small organization it can be a versioned policy register plus a regular check that each enforcement point's configuration matches it, rather than software.

A minimal setup for a 20-200 person biotech

A company of this size rarely needs to buy a control plane. The individual controls, such as least privilege, human approval and logging, are covered in Atlas Bio's answer on securing AI agents in life-science R&D. What a control plane adds is the discipline that keeps those controls consistent across every agent. A practical first version has five parts.

One versioned policy record per agent. It names the agent's identity and owner, its approved models and the terms they are used under, the data classes it may read, its allowed tools and credentials, the actions that always need a person, and the evidence it must leave. A spreadsheet or a file in version control is enough to start, provided changes are approved and dated.

Enforcement points configured from that record, not set by hand in each tool: the identity provider for the agent's identity and grants, one approved model endpoint, permissions in the systems the agent calls, and an approval step in the workflow. A periodic check that each enforcement point's actual configuration still matches the record. The ability to show, for any agent on any past date, which version of its policy applied. And a tested fail-closed path for high-impact actions if a component is unavailable.

The NIST AI RMF suggests where the policy record starts: GOVERN 1.6 asks for mechanisms to inventory AI systems, and NIST AI 600-1, the generative AI profile, suggests enumerating generative AI systems for that inventory (GV-1.6-001). MANAGE 2.4 asks for assigned responsibility to supersede, disengage or deactivate AI systems whose outcomes are inconsistent with intended use, so each policy record should also say who can switch the agent off and how.

Discovery research that creates no regulated records may need only the security side of this, but data that may later support a submission, for example in an IND, should be kept so its integrity can be shown if it is brought into scope. 21 CFR 11.1(b) applies Part 11 to records in electronic form that are created, modified, maintained, archived, retrieved or transmitted under any agency records requirement, and to electronic records submitted to FDA under the Federal Food, Drug, and Cosmetic Act or the Public Health Service Act. Where that applies, the evidence and decisions rows above become validation questions for your organization, under your quality function's oversight.

A control plane is easier to design once you know which agents, models and tools already touch your research data. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. Built and tested offline against a mock, not yet deployed: a template for a customer-controlled AI environment in your own Azure tenant. Planned, design only with no code yet: an AI gateway, agent permissions, evidence linkage and notifications. Decisions on GxP scope and validation stay with your organization, under your quality function's oversight.

Is an AI governance control plane the same as AI agent governance?

No. Governance is the operating model: owners, reviews, change control and retirement. The control plane is the architecture that carries those decisions to every place an agent acts and records what happened.

Is an AI gateway enough?

Not on its own. A gateway governs model requests. It does not issue agent identities, enforce tool permissions inside a LIMS or ELN, or decide which actions need a person. It is one enforcement point among several.

Does a control plane make an agent GxP-compliant?

No. It can supply access controls and authority checks, and agent-activity logs that complement, but do not replace, the audit trail of the system that holds the regulated record. Validation for the intended use and the procedures around it remain your organization's responsibility, under your quality function's oversight.

Do we need special software to start?

Usually not at 20 to 200 people. A versioned policy record per agent, enforcement points configured from it, a periodic check that configurations still match, and a tested fail-closed path cover a first version.

Which frameworks should the control plane map to?

A common combination is the NIST AI RMF for vocabulary, NIST SP 800-207 for the decision and enforcement split, OWASP's LLM and agentic guidance for threats, and 21 CFR Part 11 and EU GMP Annex 11 (2011) where regulated records are involved. A revised Annex 11 and a new Annex 22 on AI were published in draft for consultation in 2025 and are worth watching.

  1. NIST SP 800-207, Zero Trust Architecture
  2. NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0)
  3. NIST AI 600-1, AI RMF: Generative Artificial Intelligence Profile
  4. OWASP GenAI Security Project: LLM06:2025 Excessive Agency
  5. OWASP Top 10 for Agentic Applications for 2026
  6. eCFR: 21 CFR 11.10 - Controls for closed systems
  7. eCFR: 21 CFR 11.1 - Scope
  8. eCFR: 21 CFR 11.100 - General requirements (electronic signatures)
  9. FDA guidance: Part 11, Electronic Records; Electronic Signatures - Scope and Application (2003)
  10. European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (2011)
  11. European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (2025)

Talk to Atlas Bio.

Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.

Contact Us →