Leaders from Revefi, Rhino Federated Computing, Precisely, AtScale and Google Cloud defined what makes AI truly agentic and explained why human review won't scale. They discussed human-on-the-loop oversight, two-tier governance, agents that govern agents and the rise of agent ops.
Key takeaways
- Agents pursue goals. Unlike a chatbot's one-shot answer, an agent takes multiple steps toward an outcome, and a fixed ETL job is a workflow, not an agent.
- Human in the loop won't scale. Kulandaivel expects a shift to human on the loop, with people approving only a subset of agent actions.
- Humans set the guardrails. Oversight moves to defining policies and stepping in when an agent crosses a risk or confidence threshold.
- Governance is about provenance. Teams need to know what data an agent used, why it acted and who approved it.
- Data trust comes first. If you don't trust your data for BI, panelists said, don't unleash agents on it.
A chatbot answers a question. An agent "gets things done," said Girish Bhat, SVP at Revefi. That distinction framed "The Agentic Shift: Designing AI That Can Reason, Plan, and Act," a Data Science Connect expert panel webinar on what it takes to run agentic AI inside large enterprises.
Brian Mink, co-founder and president of Data Science Connect, moderated the panel. Joining Bhat were Adrish Sannyasi, VP, Customer Solutions and Delivery at Rhino Federated Computing; Antonio Cotroneo, Director, Product Marketing at Precisely; Dave Mariani, Co-Founder and CTO of AtScale; and Shan Kulandaivel, Head of Product, BigQuery Advanced Analytics at Google Cloud. The session was brought to you by Precisely, Revefi, Google Cloud, AtScale, Rhino Federated Computing and SAP, among others.
What makes an AI system truly agentic?
Bhat described agentic systems as goal-directed: they take a series of actions over time to reach an outcome. Asked why cloud costs spiked this week, an assistant gives a one-shot answer. An agent, he said, digs into the workloads and queries behind the spike and can suggest how to cut the spend.
Kulandaivel said the defining feature is a built-in, non-deterministic reasoning loop that lets software "reason, act, learn," and improve, unlike rule-based systems. He described agentic AI as a discipline, much like software engineering, covering development, deployment practices, governance and security for agents.
Sannyasi warned against labeling everything an agent. A scheduled ETL job with a fixed input and output is "a predefined workflow," he said, though it can serve as a tool an agent calls. Real agentic work, in his example, is finding the best patient cohort for a clinical trial: discovering data, checking quality and simulating cohorts before recommending criteria.
Mariani said open-ended questions let a model generate queries, reason on the results and ask more. He wants that reasoning to stay creative while the data and calculations stay deterministic, which is where a semantic layer comes in.
Does human in the loop scale for AI agents?
Kulandaivel called human in the loop the right architecture for the first wave of agents because it builds trust. But he said it won't scale, since the number of agents will outpace the humans available to review them. He expects a shift to "human on the loop," where agents act autonomously on certain classes of decisions under audit and governance, and people approve only a subset of actions.
Cotroneo said the human's role moves toward defining guardrails and policies, then stepping in when an agent crosses a risk or confidence threshold, or when an action can't be explained or audited. In that model the person handles exceptions and edge cases while the agent may take on most routine work.
Sannyasi described a risk-based approach in Rhino's data harmonization engine. When the model is confident it proceeds, and otherwise it produces a short list for human review rather than asking people to check every record. Every step is logged and traced for debugging. Mariani went further, saying humans have moved "from being an operator to more of a supervisor" and that agents will increasingly supervise other agents against goals and KPIs.
How should enterprises govern autonomous agents?
Autonomy increases the blast radius of bad decisions, Cotroneo said, which is why more organizations want full traceability: what data an agent used, whether it was current and complete, and whether it carried the right context. He called agent governance "data governance under pressure."
Sannyasi described a two-tier model. Program-level guardrails define what an agent is supposed to do, while individual data owners set their own rules, such as blocking access to medical images. Audit logs and lineage let each owner review violations in their environment. Mariani added that data policies must be applied at the source "if you really want to be confident let your agents roam."
Bhat expects data, security and AI governance to converge, because "fundamentally it's all about provenance." If an agent changes warehouse compute settings, teams need to know why and who approved it.
What does it take to integrate agents into enterprise workflows?
Kulandaivel made three points. Integrate through standard protocols such as MCP and agent-to-agent. Use agents to govern agents, since "the governance problem is itself ripe to be agentified." And build agent ops as its own discipline, as site reliability engineering and DevOps did before it.
Bhat said embedding AI piecemeal is easy, but "the real hard part is the access, trust, and organizational lethargy and change." He recommended narrow, read-only access and integrating with existing workflows rather than rebuilding them.
Mariani warned that without shared business meaning, agents "are going to get answers wrong confidently." Cotroneo pointed to data spread across clouds, SAP and mainframes, and to the change management employees need. Sannyasi said data modeling, documentation and API design will become first-class skills, since a field like "account status" can mean different things in different organizations.
What mindset shift do teams need for agentic AI?
Each panelist closed with one shift:
- Sannyasi: The "bottleneck for agentic AI is not intelligence." It is bringing agents to the data with guardrails, auditability and security.
- Mariani: "If you don't trust your data, it's game over." If trust in data for BI is shaky, don't unleash agents on it.
- Bhat: "Stop developing three-year AI plans and strategies. In 3 months, everything is going to change."
- Kulandaivel: Unlock unstructured data and treat agent ops as a separate discipline.
- Cotroneo: Agentic AI amplifies whatever data foundation you already have, good or bad.
Where the agentic AI conversation continues
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions. Leaders who want to compare notes in person can join The AI Enterprise Conference on Tuesday, August 31, 2027, at Pier Sixty in New York City. Registration is coming soon.
Frequently asked questions
What is the difference between an AI agent and an AI assistant?
An assistant gives a one-shot answer to a prompt, while an agent pursues a goal over multiple steps and can act, according to Revefi's Girish Bhat.
What does human on the loop mean?
A model described by Google Cloud's Shan Kulandaivel in which agents act autonomously on certain classes of decisions under audit and governance, and humans review and approve only a subset of actions.
How do you govern autonomous AI agents?
Panelists recommended full traceability and lineage, data policies applied at the source, program-level guardrails plus rules set by individual data owners, and agents that help govern other agents.
What is agent ops?
A discipline Kulandaivel expects to emerge, like site reliability engineering or DevOps, combining technology, process and policy skills to deploy and manage agents.
Watch: The Agentic Shift: Designing AI That Can Reason, Plan, and Act on YouTube
Event: Webinars · More coverage: Webinars insights
