In an earlier column, I talked about the virtuous cycle of data quality (DQ) helping artificial intelligence (AI), and conversely, AI helping DQ. An emerging tool for the latter is agentic data quality (ADQ), a way to apply large language models (LLMs) that has the potential to automate much of the data quality assessment, correction, and reporting now being performed by data stewards and rule-based systems.
While ADQ promises to bring this greater efficiency to data quality management (DQM), it is not perfect, and as we all know, AI makes mistakes. That brings me to the topic for this post: the ADQ governance principle that “AI agents should only be allowed to change values in reversible data fields!” While this principle should be applied to all automated data changes, it is especially important for ADQ because of the scale and speed at which DQ agents can change data.
A reversible field is one where either:
- The original value is preserved and can easily replace the changed value, or
- The changed value can be algorithmically reversed, or
- The change is deterministic and has low risk of causing a serious problem if incorrect.
Data Quality Accelerator
Learn how to build, sustain, and measure a data quality initiative – September 30 – October 1, 2026.
Here are some examples:
| Field | Change | Reason |
| Change “john smith” → “John Smith” | Yes | Original text is preserved or easily reconstructed. |
| Trim leading/trailing spaces | Yes | Safe, deterministic correction. |
| Standardize phone format (501)555-1212 → 501-555-1212 |
Yes | Only presentation changes. |
| Change “Arkansas” → “AR” | Yes | Standardization with little ambiguity. |
| Change “Gogle” → “Google” | Maybe | If confidence is very high and original value is retained. |
| Change customer last name “Smyth” → “Smith” |
No | Could overwrite legitimate data. |
| Replace missing salary with estimated salary | No | Creates new information that may be incorrect. |
In enterprise DQM, AI driven corrections can sometime be incorrect. Restricting agents to only change the values in reversible fields provides:
- Auditability: You can see what was changed.
- Recoverability: Mistakes can be undone.
- Trust: Users know the original data hasn’t been lost.
- Regulatory compliance: Many industries require preservation of original data values.
Suppose agents in an ADQ system are cleaning customer records. Then instructions for agents might include:
Allowed automatic corrections:
- Capitalization
- Date format normalization
- Phone number formatting
- ZIP code formatting
- Removal of extra spaces
Require human approval:
- Changing a person’s name
- Changing an address
- Merging duplicate customers
- Correcting company names
- Inferring missing values
Organizations in the financial services industry often use the following guidelines:
- Safe, reversible transformations can be fully automated
- High-confidence semantic changes suggested by AI are allowed with human approval
- Low-confidence or high-impact changes require manual review
The goal is that we all want trustworthy and reliable data curation processes. Applying the principle of reversibility in AI governance can help us reach that goal.
Build your AI governance skills in 2026.
DATAVERSITY’s training programs cover AI governance, data governance, and compliance for data practitioners.
