How agentic AI amplifies data management challenges

1. Maintaining high data quality levels

Poor data quality is the leading cause of AI project failures, said Doug Gilbert, CIO and chief digital officer at digital transformation consultancy Sutherland. Agentic AI applications exacerbate data quality issues, especially if agents are configured to act autonomously.

"When the underlying data is incomplete, inconsistent, outdated, biased or ungoverned, the risks go far beyond mediocre outputs," Gilbert said. "They trigger cascading errors at machine speed in critical business applications."

Data management best practices are shifting accordingly, he added. Master data management is moving from an optional discipline to a foundational one for ensuring data quality, although MDM initiatives themselves remain challenging due to their technical complexity and high implementation costs. Organizations are also increasingly embedding data observability and quality gates directly into data pipelines rather than auditing them after the fact, according to Gilbert.

Data quality fundamentals must be at the front of the conversation when organizations deploy AI agents, noted Rosaria Silipo, a data scientist, author and co-host of the "My Data Guest" podcast. "An agent is only as reliable as the data it reasons on, and that is still a human responsibility," she said.

2. Stopping 'garbage agents' from scaling

Dan Federoff, vice president and head of data at IT consultancy Bridgenext, said agents won't question or push back on bad data. The danger, he warned, is they'll spread it across the organization through flawed analytics and actions before anyone notices.

This has inspired a variant of the traditional "garbage in, garbage out" adage in data management circles: "garbage data, garbage agents." In addition to strong data quality, lineage and governance processes, standardized prompting and clear engineering guardrails are necessary to prevent errors from proliferating, Federoff said. Managed effectively, agents safely scale execution within those boundaries, he added.

3. Providing sufficient context about data to agents

To function accurately, AI agents also require a context layer built on metadata and shared semantics that provides a clear understanding of data relationships and rules, according to Gartner. Without one, agents "are far more likely to hallucinate, introduce bias and produce unreliable results," Gartner analyst Rita Sallam said during a session at the Gartner Data & Analytics Summit in London in May 2026.

By 2027, Gartner predicted, organizations that prioritize developing unified semantics will increase agent accuracy by up to 80% and reduce agentic AI costs by up to 60%. Reducing errors and increasing trust in data for autonomous decision-making should push data leaders to invest in semantics "as a non-negotiable foundation" for agentic AI, Sallam said. She also expects regulators to demand greater transparency about the semantics that agents use.

4. Controlling data access by agents

Steve Touw, co-founder and CTO at data security platform vendor Immuta, said traditional identity and access management approaches fail to scale sufficiently for agentic AI. Trying to solve data security issues at the authentication level by simply proving an agent's identity is a mistake, Touw said. The real problem is authorization: governing in real time what data the agent is allowed to access.

When an organization has hundreds or thousands of agents acting on behalf of various users with different permission levels, managing agent permissions manually or through static roles is impossible, according to Touw. Ghost accounts and over-privileged agents become a ticking time bomb for data leaks.

"You have to move from gatekeeping access to orchestrating it," Touw said. He recommended decoupling security policies from the data storage layer. An abstract authorization tier ensures that each user's permissions travel with an agent, enabling data teams to automatically enforce fine-grained controls -- such as data masking or row-level security -- regardless of which system the agent pulls data from.  

5. Preventing agents from amplifying data bias

Data bias is a common issue in AI, and agentic systems compound the risks for organizations. If mismanaged, "every cache between your data and your agent is a lie waiting to be narrated," said Rock Lambros, director of AI security and governance at software vendor Zenity. Bias in training and production datasets similarly creates the potential for erroneous or discriminatory outcomes to go unnoticed, which can negatively affect business decisions and damage a company's reputation.

In addition to rooting out data bias, Lambros recommended treating an agent's session history as a data management problem. Old results can bias future analysis or decisions, he noted. For example, Lambros has seen an agent skip a valid data source because a process failed an hour earlier, then create a confident answer to an analytics query from incomplete data -- or worse, nothing. The fix, he said, is periodic session resets combined with tracking provenance metadata for every artifact the agent reads.

6. Managing large amounts of unstructured data

Agentic AI tools unlock greater access to unstructured data, which they can use for new types of analytics. Indeed, Stephen Catanzano, a senior analyst at Informa TechTarget's Omdia division, wrote in a March 2026 TechTarget article that AI agents often need to access unstructured data to get contextual information that conventional structured data doesn't provide. Otherwise, he said, their reasoning and decisions will be based on an incomplete understanding of the available data.

However, many organizations have vast repositories of unstructured data that are harder to wrangle and manage than structured datasets. Catanzano recommended deploying a data lakehouse that supports unified governance and management of both structured and unstructured data. Krishna Subramanian, president and COO at unstructured data management vendor Komprise, said data teams should also classify unstructured data at scale to prevent misuse by agents while enabling them to discover and query the right data for specific applications.

7. Supporting bidirectional data integration

AI agents are fundamentally changing data integration. Conventional integration for analytics applications is a one-way pipeline: data moves from source systems into a data warehouse or other repository. "AI agents break that model because they don't just consume integrated data," said David DuChene, senior manager of data and AI professional services at SHI International. "They generate new data, discover latent relationships and write enriched context back into the data estate."

This means data integration must become bidirectional and, in many applications, continuous. To enable that, DuChene said data teams should embed data governance into the integration architecture and develop an entity resolution layer that creates a unified view of related records across different data sources. He also recommended taking a domain-by-domain approach rather than an enterprise one. Doing so will help organizations find AI-driven business value "in weeks rather than years," he said.

George Lawton is a journalist based in London. Over the last 30 years, he has written more than 3,000 stories about computers, communications, knowledge management, business, health and other areas that interest him.