Agentic AI Governance: Five Pressure Points for Data and AI Leaders

 
A model gives you an answer you can check. An agent takes an action at machine speed, and the consequence immediately becomes real. Putting a human in the loop of every agentic transaction undermines the tremendous benefits of automated agentic action. This heightens the need for Agentic AI governance and data security.

That shift ran through every point the panel made at the Agentic AI Governance Town Hall, led by GlobalLogic on July 16 for the Gartner CIO and C-Level Communities. Andrew Kalinovsky, our SVP of AI Business GTM, moderated and set the stakes: agents in production reason non-deterministically and act with little supervision, so while the enterprise gains unprecedented efficiencies, it also takes on a new class of risks in the same move. The governance built to check what a model says was never designed to govern what an agent does.

Andrew steered the discussion with two leaders already running agents in production: Reena Bajaj, Head of Data and AI, North America at Munich Re, and Sophie Sergeenko, Head of Data and Digital Innovation at Heineken. They brought insurance and consumer-goods vantage points to a discussion relevant for all industries, and independently landed on the same solutions.

Five themes stood out as essential considerations for data and governance leaders putting agents into production.

1. Agent Governance Is a Different Discipline

For the last several years, enterprises built AI model governance. Validate the training data, test for bias, monitor for drift, keep a person in the review path before an output reaches a decision. With empowered agents, the question moves from what a bot can say to what the agent can do: how much data it can access, which tools it invokes, what it modifies, and where it triggers action downstream. A wrong answer from a model is easy to spot. A plausible chain of actions that quietly produces the wrong business outcome is not.

That is why the panelists agreed that specialized governance controls should be established to enable the safe use of autonomous AI agents, such as agent identity management, decoupling agent activity from data access, query-time governance, reasoning traceability and audit trails, and a human-on-the-loop philosophy.

2. Overprivileged Access Increases Risks of Sensitive Data Leaks

In early designs, productivity bots that support human operators often operate under the credentials of their user, which implies that the human is still responsible for all actions and outcomes. Autonomous agents, being independent actors, need to be treated differently from human employees. On one hand, they possess superhuman abilities to scrub through huge volumes of information; on the other, they may lack the most basic common sense and restraint of even junior team members. Reena Bajaj’s framing was blunt: treat agents as trusted by default, and you have built a new point of failure rather than a new capability.

The principle both panelists returned to is separation. What an agent is allowed to do should be evaluated against policy at the moment of each request, not granted once at onboarding. 

Sophie Sergeenko made the organizational version of the same point: give every agent its own identity, its own authorization, and a human manager accountable for what it does — the way you would for a new junior employee: capable but not trusted unconditionally.

3. The Case for Proportional Trust

This was the thread the conversation kept returning to. Business leaders interpret the “Zero Trust” approach as friction that slows the business down. Sophie Sergeenko proposed framing it as “proportional authorization” — that is, enough control to contain the risk of a given action, enough freedom for the agent to stay useful.

The calibration is function- and industry-specific. A mistake in an employee picnic promotion does not carry the same risk as a mistake in assessing insurance policy terms, so the level of permissions and controls for agents can vary widely based on risk. Reena Bajaj described the same model as progressive autonomy: automate low-risk actions, supervise medium-risk ones selectively, and hold high-risk decisions for human approval.

How you actually draw those thresholds, and optimize them over time, is the hard part, and Andrew Kalinovsky pressed the point that it is where most programs are still refining their approaches to risk assessment and conditional governance.

Recommended reading: The New Physics of Software: Converting Strategic Intent into Market Outcomes at AI-Speed

4. Governance Has to Run in Production

Output assurance in an agentic system cannot sit at the end as a review stage: it must be injected into the Agent reasoning process, which is iterative by nature, and also control interactions between multiple agents, which can lead to compound damage from minor inaccuracies. Both panelists described governance as something the system does while it operates: policy encoded as a runtime constraint, used as a deterministic validation layer for agents’ reasoning loops and tool use.

Reena Bajaj compared it to network security, where you do not inspect every packet by hand but define policies, watch for anomalies, and intervene automatically. For regulated organizations, there is a harder requirement underneath this. A regulator will not just ask whether your policy was sound. They will ask you to show that you have control mechanisms in place that lead to policy enforcement on a specific date for a specific decision. That means capturing the initial context, the agent’s chain-of-thought in human language, and its interaction with tools and other agents, and producing an audit package for compliance and post-incident reviews.

5. Oversight at Scale Means Humans On-the-Loop

The reflex answer to the nondeterministic nature of agentic reasoning is human review, but that does not survive the arithmetic at scale. An agent making thousands of operational decisions a day, based on large and complex datasets, cannot be micromanaged by a person.

Sophie Sergeenko was direct that checking every task will not scale, and that the paradigm has to shift toward humans designing and continually validating the guardrails rather than approving each individual output. So instead of human-in-the-loop control common for AI-assisted workflows, the agentic AI calls for a human-on-the-loop approach: people hold authority over intent, boundaries, and the trade-offs that carry consequence, while the system executes inside them.

Andrew Kalinovsky closed the thread with the operating guidance: start agents on tight controls and widen their authority as they earn it, since agents’ capabilities evolve and the governance program has to evolve with them. Most enterprises have not yet made that shift organizationally, and the decision-making authority and corporate politics will need to catch up with the rapidly progressing technology capabilities.

Putting Theory Into Practice

The topics raised in the panel discussion are what GlobalLogic is working through with data and analytics leaders for real-world applications, right now. The AI engineering community has already developed best practices for addressing Agentic AI governance challenges, and our teams have been actively putting them in production and refining them based on actual outcomes.  

If your company is evaluating how Agentic AI can be applied to generate value, but is uncertain about the risks – you are not alone, but don’t let that stop you. There are proven solutions to govern AI agents without stifling their superhuman performance abilities, and there are experts who can guide you from design to implementation.

Talk to our team about your first governed agent workflow.

Learn more:

•       Reliable AI in Mission-Critical Environments: Why Trust Matters in Agentic Systems
•       Agentic AI Architecture: How to Engineer Intelligent, Autonomous Enterprise Systems
•       Revolutionizing Investment Strategies with AI-Driven Insights and Streamlined Decision-Making