Data Speaks for Itself: Agentic Data Quality – The Beginning of a Beautiful Friendship

Data Speaks for Itself is a TDAN column published every quarter.

In an earlier post, I talked about how high-quality data is necessary for creating quality language models (LMs), but in turn, LMs have the potential to help improve data quality. Data quality and AI can form a virtuous cycle: DQ helps AI, and AI helps DQ. With the development of agentic AI, the cycle can be even stronger.

The reason is that agentic AI is becoming a game changer. It extends generative AI from simple prompt-and-response interaction to the ability to deploy agents that can autonomously observe, analyze, evaluate, make decisions, and coordinate with other agents. Perhaps most importantly, these agents can access and use external tools and data. Agents can even create their own helper agents!

The impact of applying agentic AI to data quality, i.e., agentic data quality (ADQ), will be enormous. ADQ has the potential to automate the detection of data anomalies, autonomously remediate certain data errors, and minimize the number of human-in-the-loop interactions, dramatically increasing data quality management efficiency and effectiveness. ADQ pipelines have the potential to dramatically reduce processing costs while improving data quality that leads to better decision-making.

We are now entering an age where we can replace many rule-based, batch-oriented validation routines with real-time agents that can detect data anomalies like missing value, invalid values for the item domain, misfielded item, and out-of-range values. Not only that, but these agents can decide which specialized agent should try to correct the problem. If agents are unable, or not allowed, to make the correction, they can escalate the problem to a person to resolve the issue. Through the escalation process, agents can learn new solutions for their specialization. Agents can also log and report all their decisions and actions for later analysis and audits.

Properly deployed and governed agents can be like having an army of additional data stewards in your organization. Agents can increase the level of automation in both data anomaly detection and correction, lower the volume of manual decisions and interventions thus speeding up processing. Having agents learning new ways to remedy problems reduces the reduces the time and effort needed to release new code as required with rule-based systems.

While all this sounds wonderful, there are definitely some issues that need to be addressed before an ADQ miracle can happen in your organization. Perhaps the most important one is accuracy and trust. According to a Dataiku survey (March 2025), 59% of data leaders have already faced a business issue or crisis stemming from AI hallucinations or inaccuracies in the past year. This is why AI governance is necessary to enforce the safe and reliable use of AI for you and your stakeholders. Because AI agents can decide and act autonomously, they require governance just as much as, or perhaps even more than, employees.

The good news is that many strategies have already been developed to detect and remedy hallucinations. For example, LLM-based audits. This is where one LLM assesses the output of another independent LLM. Another is real-time monitoring and logging to detect model “drift.” And don’t throw away the rule-based checks yet. Validation rules can still be used to quicklycatch data formatting errors and many types of invalid values produced by agents. Finally, extensive simulations should be run before deployment along with strong adversarial (Red Team) testing to try and intentionally break agents.

While LM performance is constantly improving, there are two main drawbacks to scaling agentic AI applications. The first is model inference time, and second is error cascading. One problem of agentic scaling is that LM inferencing time is on the order of seconds instead of milliseconds for traditional rule-based systems. This makes scaling to process tens of millions of records a challenge. The key strategies here are to run on local models, not remote API calls. Whenever possible, employ graphic processing units (GPUs) and large memory spaces. And don’t abandon rules! Combining deterministic rules and inferencing where appropriate can create efficient hybrid solutions. Another strategy is to employ distributed parallel processing where the worker nodes employ small language models (SLMs) or distilled LLMs. Many applications such as master data management (MDM) and semantic search use the power of LMs through embedding models where text records are transformed into high-dimensional vectors for faster processing.

The second problem of error cascading is easy to overlook. While the ability to build multi-agent solutions helps solve a broader range of problems, it also carries the risk that agentic pipelines that appear to work wonderfully in a small-scale POC can fail dramatically after scaling to enterprise level. The problem is the potential for cascading errors. Cascading error is similar to the old telephone game, where a message is passed person to person through a chain of people. Often the final form of the message is completely different than the original message because of small changes that occur at each exchange. Scaling agentic systems requires taking care to monitor and assess each agent’s output to ensure that it hold up the stress of enterprise-level processing.

There are many other architectural issues for ADQ such as modification, maintenance, reusability, interoperability, vulnerability, and security, but these will have to wait. However, the parting thought for this column is governance. ADQ and agentic AI in general have elevated the need for governance to a new level. Because agents can make decisions and take actions, they must be subject to the same governance policies as employees. New AI frameworks continue to make it easier for non-technical employees to build and deploy agents. Rogue agents can cause tremendous damage to organizational security, regulatory compliance, integrity, and reputation.

Human oversight is still important for critical decisions. Strict, least-privileged access to data by agents must be enforced to prevent unauthorized data exposure and agents must be limited to well-defined tasks. It is also important to maintain an inventory of agents and document their actions, ensuring decisions are understood. There should be clear human ownership, responsibility, and accountability for each agent.

We can never be sure of what the future holds, but I am hopeful that ADQ will turn out to be one of our best friends.