Why senior data leaders aren’t ready to let AI act alone

5 key takeaways from Ataccama’s recent “Data Think Tank” event

Enterprises are facing immense pressure to deploy AI agents at machine speed, but senior data leaders are – at least for now – reluctant to enable fully autonomous execution. Not a single participant in Ataccama’s recent virtual Data Think Tank said they believed AI agents were ready to act without human supervision. 

Ataccama’s Data Think Tank is a monthly closed-door roundtable of more than a dozen data leaders working to implement AI at scale in their organizations; the forum gives participants a chance to share insights and best practices in confidence.

Moving from cautious “human-in-the-loop” pilots to fully automated, agentic AI requires shifting automated data quality checks directly into an organization’s core operations, Think Tank participants agreed. Data trust, they said, is now an urgent business imperative, and companies need to understand how each AI use case might introduce risk.

Five useful insights for data leaders, from the Data Think Tank event

1. “Data Availability Pre-Checks are invaluable”

Ataccama’s Field CDO, Larry Hunt, facilitated the session and told Think Tank members that complex machine learning algorithms aren’t always the main reason AI projects fail. He explained that it’s usually a lack of visibility into the trustworthiness of the organization’s raw data. 

Tactical tip: An AI intake process should be designed to determine whether the data required for a specific use case is fit for purpose – that is, if the data is digitized, accessible, and structured. If the data foundation isn’t there, the project will stall at step zero, Hunt said.

2. Unstructured SOPs are a secret AI bottleneck

At the Think Tank session, participants discussed how GenAI models designed for operational efficiency (like call center co-pilots or automated routing) depend heavily on internal context—namely, standard operating procedures (SOPs) and desktop documentation. 

Tactical tip: If an organization plans to deploy internal-facing agents, its first initiative must be to digitize, centralize, and structure tribal knowledge and desktop documentation so AI agents can actually consume it. When integrated and operationalized into AI workflows, internal data that captures an organization’s hard-earned knowledge about its industry and customers can be an invaluable competitive advantage.

3. The data leadership mindset is moving from “Absolute Quality” to “Risk vs. Reward”

Traditional regulatory reporting demands 100% data accuracy—there is no “almost good enough.” However, Think Tank participants agreed that applying this rigid “perfection mindset” across all AI initiatives can slow innovation, and they noted that business and data leaders need to ask: Is the financial or operational risk of an error lower than the business value of launching quickly? 

Tactical tip: Establish tiered Data Quality Gates. For example, consider defining an 80% confidence threshold for internal support agents (where a human verifies the answer), but enforcing strict 99%+ thresholds for direct client-facing execution or regulatory reporting.

4. “Hidden governance” already exists in your enterprise

Business teams often push back on data governance because they assume it requires building massive, slow, net-new frameworks from scratch. Ataccama’s Larry Hunt noted that most enterprises already have operational risk management, compliance, and process controls in place. They simply haven’t integrated data risk management into the broader framework. 

Tactical tip: Organizations should remember that “AI Governance” is not a brand-new bureaucratic burden but the natural extension of existing practices they can use to increase the chance of successful project outcomes.

5. Machine-readable metadata is essential for Agentic AI

At the session, some data leaders noted that, historically, human workers looked at a data asset, read the governance policy, checked the retention schedule, and manually decided if it was safe to use. As AI agents begin operating at machine speed, passive human-readable data catalogs are no longer sufficient. 

Tactical tip: To prepare for true autonomous agents, enterprise metadata must be translated into machine-readable logic. In AI, the semantic layer is where you build the rules that let AI agents check permissions on their own. If a company already has clear, unified “labels” connecting its raw data to meaningful business terms, that same system can be extended so machines can query it too: not just “what is this data,” but “am I allowed to touch it.”

Ready to build a trust foundation for autonomous AI?

Watch this space for Insights from future Data Think Tank events, and in the interim, learn from this new asset from Ataccama: Certified, Agent-Ready Data: A Practitioner’s Guide to the Data Trust Layer on Snowflake. 

The paper explains how to:

• embed automated data quality and trust checks directly into modern ELT pipelines
• make data certification a continuous part of everyday data operations
• build a dedicated Data Trust Layer on Snowflake to ensure downstream AI models and agents consume valid, compliant data