Beyond Generative AI to Agents: CDO Vision San Francisco

The enterprise AI conversation over the past two years has been shaped by generative AI and its ability to generate content, summarize information, and accelerate drafting. That conversation is shifting. Systems that do not merely respond but reason, plan, and act autonomously across multi-step workflows are emerging as the next frontier, and for enterprises that get the transition right, agentic AI represents a fundamentally different kind of competitive advantage than the productivity gains of the generative era.

That shift was the subject of a panel discussion at CDO Vision San Francisco, part of the CDO Vision AI World Series hosted by AIM Media House. The series is an invite-only global leadership gathering that convenes senior Chief Data Officers, Chief AI Officers, CIOs, and enterprise AI leaders for peer-led conversations on practical AI transformation. The session was moderated by Suman Bhattacharya, Senior Director of Data Science and AI at DocuSign, and brought together a panel spanning some of the most consequential sectors in the agentic AI conversation: Suresh Teckchandani, Vice President of Product and Engineering at Ancestry; Elan Panneerselvam, SVP and Head of Enterprise Data & Analytics at CardWorks; Siwei Shen, VP of Engineering at Coinbase; Ajinkya More, Head of AI at Remitly; and Yogesh Kandlur, Senior Director of Data Science at Walmart. Together, the panelists represented genealogy and e-commerce, consumer finance, crypto infrastructure, cross-border payments, and large-scale retail, a range of perspectives that gave the discussion an unusually broad operational lens.

The panel's starting point was definitional, and the distinctions that emerged were consistent across speakers. Suresh Teckchandani framed generative AI as fundamentally reactive, chat experiences, recommendation systems, and content generation that respond to a prompt, versus agents that "perform a highly complex process for you, coordinating or orchestrating across multiple systems." Yogesh Kandlur put it simply: an AI agent runs a predefined task, while agentic AI adds "an orchestration layer" with planning, governance, and a feedback loop back into execution. Ajinkya More offered the sharpest framing of the distinction, describing it as the difference between an isolated intelligent entity and an orchestrated ecosystem, the analogy he used was an individual versus a team, with the deeper strategic choice being whether an organization is trying to build a smarter single agent or invest in the infrastructure that allows multiple agents to collaborate reliably.

That framing, planning, reasoning, and acting, surfaced repeatedly throughout the discussion, but so did a caveat. Ajinkya More argued that moving from a prototype agent to a production-grade enterprise system depends on a defined set of prerequisites, including observability into how an agent reasons and what actions it takes, clear evaluation rubrics measured against golden datasets, and feedback loops capable of maintaining or improving performance over time, finished off by guardrails to "ensure that the agent system is acting in accordance with the enterprise policy."

If the first theme established what agentic AI is, the second established the condition under which enterprises are willing to deploy it. Governance and observability, the panel agreed, are no longer secondary considerations bolted onto a working system, they are front and center as organizations move up the maturity curve.

The clearest articulation of this came from Suresh Teckchandani at Ancestry, who described deliberately approaching AI as "a bounded orchestration as opposed to unconstrained autonomous process, just go do stuff." The distinction allowed his team to leverage the power of AI with adequate governance and human oversight while still moving customer metrics like conversion and engagement in the right direction. A similar caution came from Elan Panneerselvam who emphasised that regulatory exposure means every agentic deployment, even something as operationally repetitive as financial reporting, still requires human-in-the-loop approval before reaching production, despite agents being able to compress multi-day reconciliation processes into a matter of hours. "Even though you have the rag and all the things," Panneerselvam noted,

The panel was equally direct about the risks of getting this wrong. Ajinkya More warned that unconstrained self-learning in live production systems, without clear governance boundaries, evaluation frameworks, and human oversight, is "going to hurt the business at some point, and not just your brand." As enterprises mature their agentic deployments, the consensus was that trustworthy AI, not fully autonomous AI, is the actual objective.

The panel's most consistent throughline was a refusal to treat AI adoption as an end in itself. Every example offered was grounded in a specific, often unglamorous operational problem: Suman Bhattacharya pointed to automated SDR workflows handling millions of inbound leads at DocuSign before a sales representative ever gets involved; Siwei Shen described blockchain upgrade monitoring at Coinbase, where supporting more than sixty blockchains and several hundred quarterly upgrades had previously required scaling headcount in step with chain growth; Elan Panneerselvam cited KYC, anti-money laundering, and sanction screening processes at CardWorks that involve hundreds of people manually reviewing the same categories of decisions; and Suresh Teckchandani pointed to personalization and content decisioning at Ancestry, where static workflows could no longer keep pace with a growing customer base.

The pattern that emerged was sector-agnostic: workflows that are highly repetitive and require significant human throughput are the ones being disrupted first. Workflows that are complex, highly regulated, or dependent on nuanced judgment, strategic decision-making, M&A evaluation, security incident response, remain firmly human-led, not because the technology cannot attempt them, but because the cost of an error and the ambiguity involved make full automation inappropriate for the foreseeable future.

That distinction became sharper still when an audience member raised a more skeptical question, whether enterprise AI adoption reflects genuine business value or institutional pressure, with boards pushing CEOs who in turn push their organizations to adopt AI primarily to avoid the appearance of falling behind. The panel's response leaned away from sentiment and toward measurement. Siwei Shen described moving away from tracking AI token usage internally, given how easily that number can be inflated, in favor of harder-to-game indicators like AI session time. Ajinkya More pointed to a more fundamental measurement gap: "the million-dollar question is how you measure the output," he said, noting that even leading AI labs like Anthropic and OpenAI, by his own observation, "do not measure the output yet" which means many organizations are still relying on input-side metrics, like the velocity of AB tests shipped and the increase in deployment frequency, as imperfect proxies while output measurement frameworks continue to mature.

The discussion reflected a broader evolution underway across enterprise AI. Agentic systems are increasingly viewed as the next competitive advantage, but the panel was unambiguous that autonomous capability alone is not the differentiator. Organizations that succeed will be the ones that pair agentic execution with governance, continuous evaluation, human oversight, and a discipline around measurable business outcomes rather than adoption for its own sake.

As Suman Bhattacharya put it in closing the discussion, the workflows most ready for disruption share a common shape: while anything "too complex and too regulated" and dependent on judgment becomes "more assisted" rather than automated. That distinction, more than any single technical capability, may be the clearest signal of where enterprise agentic AI is headed next.