Runtime Governance Becomes the Test for Enterprise AI Agents

By Tom Whitfield

“Every organization expects that within the next two years, there will be more agents than humans working inside their company,” Blake Brannon, chief innovation officer at OneTrust, a privacy and data governance company, said in a TechTarget interview. The claim had no survey, sample size, supporting data or independent validation, so it signals operational pressure rather than measured adoption.

Enterprise governance must now operate at runtime, not only before deployment, because non-deterministic agents are easy to build, connect to other systems and act on their own. For employers and people leaders, that changes tool approval, employee identity and access rights, human-oversight thresholds and the evidence retained after an agent takes action.

What does shadow AI mean for employers?

Brannon compared shadow AI with earlier tension between a work BlackBerry and a personal iPhone or Android: Why carry two devices if work data can move to a personal phone? Organizations now face the same choice between blocking unapproved tools and steering employees toward sanctioned AI tools and models while securing data.

He rejects a blanket prohibition because the technology is spreading quickly. The practical takeaway is that employee tool choice cannot replace a documented approval and data-security process.

How must governance keep pace with agents?

Enterprise governance assumed decisions would be made by people and software would behave consistently. AI agents remove those assumptions, while governance programs that take weeks struggle to supervise a high volume of autonomous actions in real time.

Brannon said organizations must compress those reviews to seconds or milliseconds, analyze autonomous actions and enforce policy while agents reason and act. Failure at runtime, he argued, prevents organizations from scaling AI and creates a competitive disadvantage.

Employers must make agents functional and intelligent without introducing security problems. Runtime enforcement is now a condition of adoption, not a later control improvement.

What makes an agent deployable?

Brannon divides a successful agent into three requirements: intelligence, security and governance. Intelligence means completing the intended task; security means keeping data from the wrong person and preventing jailbreaks or backdoors; governance means acting permissibly under organizational policies.

He calls this a quotient because the three values multiply. If any one is missing or zero, he said, the agent's value or return on investment will be zero. A useful output does not compensate for failed controls.

How does existing compliance apply to AI?

Brannon said asking which AI-specific regulation applies is almost the wrong question. Employers should instead identify existing obligations tied to organizational actions and data use; he treats AI as an engine, much as compute has long been.

He argued that many existing regulations still govern data used in AI, sometimes without new AI-specific rules. He also sees compliance, regulation, best practices and traditional governance on a collision course: conflicting rules act like three parents giving an agent different instructions, so employers need one authority for guardrails.

The interview names no jurisdiction, regulator, statute or effective date, so it establishes no particular legal deadline or rule. Employers must still map AI actions to applicable obligations, but this source cannot identify them.

When should a human approve an agent action?

Brannon expects each organization and use case to have some unique oversight design, but proposed three tests for an agent decision: the risk involved, how reversible the decision is, and whether human-specific intuition is required.

Most organizations, he said, treat every use case and scenario as a unique “snowflake.” To move approvals from weeks to seconds, organizations should split decisions into sub-decisions that match earlier choices; his example was reusing a decision already made 28 times in the previous three weeks.

No universal approval threshold or test result follows. Automate repeat decisions selectively rather than every agent request.

How should employers govern agent identity?

Brannon's central warning was that agents often inherit a person's identity to access systems without inheriting that person's trust.

His example: an organization trusts a human employee to delete a CRM record, opportunity or contact, but not that employee's agent. Employers must choose inherited or unique non-human identities, attach permissions and define permissible actions.

Agents also need a recorded purpose or intent, which Brannon described as a core concept not usually associated with humans. An employee's access rights cannot automatically become an agent's authority.

What guardrails must be in place?

Brannon described “three core parts” to the governance solution, then set out four implementation steps:

• Maintain visibility into all AI used by the organization.
• Keep a policy system of truth and system of record that reflects organizational decisions.
• Provide a control mechanism able to allow or deny any AI agent from taking any action.
• Keep logs and evidence that cannot be manipulated.

Brannon said: “No one should ever grade their own homework.” An agent's account is not proof because it can manipulate its log, memory and thinking.

Control must operate where people cannot, including every agent tool call at 2 a.m. on a Tuesday, and preserve proof of what happened. The interview provides no technical standard for tamper resistance or independent verification.

Policy without credible evidence leaves employers unable to reconstruct autonomous decisions.

What should employers test next?

The immediate adoption test is not whether a pilot can produce an answer. Employers need to know whether each agent has an inherited or unique identity, whether routine sub-decisions can reuse approved precedent, whether high-risk or irreversible actions reach a human, and whether independent logs can reconstruct what occurred.

Those tests follow Brannon's framework, not a new standard in the source. Over the next two years, the practical benchmark will be whether employers can authorize low-risk agent actions in seconds while retaining human judgment for consequential ones and evidence independent of the agent itself.