Key Takeaways
- When AI acts autonomously, bad data produces a bad action. The cost of poor data quality compounds fast in an agentic world.
- Agentic-Ready data is fundamentally different from reporting-ready data. Governance, lineage, and semantic context all need to evolve.
- The gap between successful pilots and responsible production is both technological and organizational.
There’s a shift underway in how data leaders talk about AI. For years, the conversation centered on generative AI and AI assistance. Now it’s centered on agentic AI — AI that acts. And that shift makes agentic AI data requirements fundamentally different from anything enterprises have faced before.
For the past few years, most of the work and energy has gone into co-pilots, chat interfaces, and model-driven recommendations. Those are genuinely valuable. But they still had a human in the loop, catching errors, applying judgment, and deciding whether to act.
Agentic AI changes that equation entirely.
When AI starts approving transactions, triggering customer outreach, or autonomously escalating service issues, the bar for data management and data integrity shifts.
There’s now a need for Agentic-Ready Data: the highest-quality data that is integrated, governed, and enriched for AI, automation, and analytics initiatives across the enterprise.
In a recent webinar, Agentic AI: What It Demands from Enterprise Data, I moderated a conversation with three practitioners to explore why trusted data foundations are more crucial than ever:
- Tamara Astakhova, Senior Partner Solutions Architect at AWS
- Dhruv Baronia, SVP and Head of Data and Analytics at Northern Trust
- Aniket Mane, VP of Data Platform and Enterprise Apps Engineering at thredUP
It was a candid conversation filled with practical insights that organizations across industries can benefit from. I want to share the biggest takeaways here.
How Does Agentic AI Change Enterprise Data Needs?
I kicked off the conversation by asking what fundamentally shifts when AI stops advising and starts acting.
Tamara framed it in terms of a few critical changes:
“When AI just advises, bad recommendations waste someone’s time. But when agents complete the full insight-to-action loop without human validation, bad data drives bad actions.” — Tamara Astakhova, Senior Partner Solutions Architect at AWS
She also flagged something that doesn’t get enough attention: the shift from static to stateful. Traditional AI asks and answers. Autonomous agents maintain persistent memory across sessions, which means that if bad information gets embedded, it can corrupt decisions across multiple future interactions.
Dhruv added that agentic AI fundamentally changes what accountability means: “When an agent takes a decision, you need to know what data is used, what decisions were made, why those decisions were made, and whether the agent has had the permission to access those resources and make those decisions.”
Aniket made a point that will resonate with anyone who’s built data pipelines: a lot of what we now call “agentic” has existed for years in automation systems.
The difference is that humans were quietly absorbing the inconsistencies. “We were always in the loop, like ‘if this breaks, we can trigger a rollback or go back to step one and run it again from step 3.’” Removing that human buffer is the real challenge. It requires knowing exactly what data is needed at every step of an automated workflow.
Key takeaway: Agentic AI raises the bar for data and removes the safety net humans once provided.
Why Does the AI Confidence-Reality Gap Keep Showing Up?
This question is directly connected to findings from the 2026 State of Data Integrity and AI Readiness Report, published by Precisely in partnership with Drexel University’s LeBow College of Business. 87% of data and analytics leaders say they have the data readiness needed for AI. But 43% cite data readiness as one of their biggest barriers.
So why does that contradiction keep showing up?
Tamara pointed to AWS and Harvard Business Review research that paints a stark picture: while 74% of leaders say AI is very important, only 26% are effectively leveraging it for positive outcomes. The gaps are consistent: only 13% say their data architecture is ready for generative AI, only 11% have the right governance in place, and only 5% feel their workforce is prepared.
Aniket traced the root cause to years of organizational fragmentation — the back-and-forth between centralized and decentralized data ownership, driven by the pressure to speed up. Every team ends up owning its own metrics, its own definitions, its own attributions. “That’s where integrity comes in — when people stop thinking about their silos. They have to come together as one organization and actually own that entire decision logic across systems.”
Dhruv connected the confidence gap to how pilots work: “Pilots have limited scope. They’re small, people are reviewing them while the pilot is running, and if data issues arise, they’re typically addressed manually on the spot to keep the pilot running. But really, these issues start arising when you move from pilot to production.”
Key takeaway: AI confidence is often earned in pilots. AI readiness has to be proven in production. Those are very different things.
REPORT2026 State of Data Integrity and AI Readiness
Findings from a survey of global data and analytics leaders.
How Does AI Data Governance Need to Evolve for Agentic Systems?
A common thread throughout the conversation was around how data governance needs to evolve as organizations turn to AI to execute decisions in real time.
Dhruv clearly reframed governance, calling for a shift from model governance to action-level governance.
Agentic AI requires asking a different set of questions than traditional model monitoring: Did the agent use the right data? Was it authorized to make this decision? Was the action reversible? If a high-risk decision was made, was there a human escalation path?
“Garbage in, garbage out — in the AI agent world, it’s accelerated garbage in, accelerated garbage out. The risk compounds very quickly.” — Dhruv Baronia, SVP and Head of Data and Analytics at Northern Trust
Tamara added that governance has to shift from reactive to proactive, with dynamic runtime guardrails, continuous monitoring for goal drift and unauthorized actions, and clear accountability frameworks that log reasoning chains and decision context, not just outcomes.
And as Aniket noted, it can’t be one-size-fits-all: agents handling marketing optimization, journal entries, and pricing decisions each need their own self-governance model. The guardrails have to move at the same speed as the decisions.
Key takeaway: Governance has to evolve from model oversight to action-level accountability — dynamic, embedded within each workflow, and operating in real time.
How Do You Scale From Pilot to Production Responsibly?
Dhruv described a staged approach along two dimensions simultaneously: scope (starting with a limited set of users and use cases, then expanding) and risk (starting in co-pilot mode, where agents recommend and humans approve, then gradually progressing to low-risk reversible actions). Both dimensions expand together, slowly.
Aniket’s experience was the most candid and useful for anyone who’s felt pressure to move fast. He watched a knowledge graph pilot show enormous promise but failed to scale. And he saw a chatbot initiative where the team that moved fastest actually failed fastest. The team that took the methodical path (and didn’t win the internal hackathon) was the one that eventually got to production.
“Patience is what is important when you go from pilots to responsible scale,” Aniket said, also citing the need to work with upper management and be transparent about failures.
After two quarters of iterations and honest conversations with leadership, that chatbot went live—and outperformed a commercial buy solution.
Key takeaway: The methods that have worked in software and data for years still apply: methodical staging, transparent failure, and patience.
Build a Trusted Foundation for Agentic AI
When I asked each panelist for one piece of practical advice, three distinct framings pointed to the same place: know where your data foundation is strong before you deploy agents, govern at the speed of decisions, and get the fundamentals right before you scale.
Agentic AI accelerates your work, but also accelerates all the consequences of errors. Successful organizations will start by building an Agentic-Ready data foundation.
That foundation is also where the Precisely and AWS partnership comes into play. By combining AWS cloud infrastructure and services with Precisely data integrity capabilities, organizations can build the frameworks that autonomous decision-making requires. As Tamara put it, without that foundation, “you’re just automating uncertainty.”
For more on what a trusted data foundation looks like, and where the gaps are showing up most often for data and analytics leaders, explore the 2026 State of Data Integrity and AI Readiness Report. And to hear more from Tamara, Dhruv, and Aniket, watch the full webinar on demand
