Data quality has always been limited by human attention. Someone has to profile new sources, write the rules, read the alerts, chase root causes and fix records. Most data teams know exactly which checks they should have; they simply never get to them. In 2025 and 2026, a new generation of tools started to change that equation. Large language models can read a table’s schema and sample values and suggest validation rules. AI agents can group related alerts, trace them through lineage, explain the likely cause and draft a fix. Data quality and MDM vendors now ship agents that resolve issues and fill missing values under configured rules, with audit trails.
The opportunity is real, but so is the risk of automating bad judgment at scale. This playbook explains where AI genuinely helps across the data quality lifecycle, where it should not be trusted alone, and how to introduce it in stages with guardrails and measurable outcomes.
Why Data Quality Needs a New Operating Model
The pressure on data quality comes mostly from AI itself. Gartner reported in February 2025 that 63% of organizations either lack the right data management practices for AI or are unsure whether they do, and predicted that through 2026 organizations will abandon 60% of AI projects that are not supported by AI-ready data. Its recommendations included making metadata active rather than passive and testing and monitoring data continuously. Both are hard to do manually across thousands of assets.
Three gaps show up in almost every assessment:
- Coverage gap. Rules exist for a small share of critical fields. New sources go live with few or no checks because nobody had time to profile them.
- Triage gap. Monitoring produces more alerts than people can investigate, so real incidents hide among false positives.
- Remediation gap. Even when issues are found, fixing them depends on a few overloaded stewards and engineers, so backlogs grow.
AI can help with all three, but differently. Rule suggestion widens coverage, alert intelligence narrows triage, and agents with write access speed up remediation. The risk profile rises in that same order, so the controls must too.
Where AI Fits in the Data Quality Lifecycle
A useful way to plan is to map AI capabilities to each stage of the lifecycle and assign a level of autonomy to each.
Profiling and Discovery
LLMs are good at interpreting what a column probably contains from its name, description and sample values: that “cust_dob” is a date of birth, that a ten-character field matching a pattern is likely a tax identifier, or that “ctry” uses ISO country codes. Combined with statistical profiling, this speeds up classification of sensitive data and the identification of reference data domains. Autonomy here can be high, because the output is a suggestion that feeds later steps.
Rule Generation
Given a profile, documentation and examples, a model can propose rules: valid ranges, allowed values, cross-field dependencies (“ship_date is on or after order_date”), uniqueness constraints and freshness expectations. The value is in breadth. A model can draft fifty candidate rules for a new source in minutes. The risk is that it encodes what the data currently looks like rather than what the business requires, effectively certifying existing errors. Rules should be reviewed and approved by an owner before they become blocking.
Anomaly Detection and Alert Triage
Statistical and machine learning detectors find deviations; LLMs and agents help explain and group them. An agent can cluster 40 alerts from one morning into three incidents, trace each through lineage to a likely upstream cause, check recent deployments, and write a short summary for the on-call engineer. This is one of the highest-value uses because it saves time without changing data. Actian’s 2026 paper on agent readiness explicitly recommends automating triage with agents that validate, diagnose, route and explain issues.
Root Cause and Impact Analysis
With access to lineage, query logs and change history, agents can answer “what changed upstream?” and “who is affected downstream?” far faster than a human clicking through tools. Their conclusions should be presented as hypotheses with evidence, not verdicts.
Remediation and Enrichment
This is where agents write data: standardizing addresses, filling missing attributes, merging duplicates or correcting codes. Vendors are moving fast here. Profisee’s 2026.R1 release, announced in March 2026, introduced data quality and enrichment agents that resolve issues and fill missing values according to rules and context fields administrators configure, with full audit trails. Ataccama describes agents that triage match proposals, resolve clear cases and route judgment calls to stewards, with risky or permanent changes requiring human approval. Autonomy should be lowest here and increase only with evidence.
Let AI widen coverage and shorten triage freely, but make it earn the right to change data. Suggestions are cheap to review; silent corrections at scale are expensive to undo.
A Staged Adoption Playbook
The safest path is to introduce AI in stages, each with a clear exit criterion before moving on.
Stage 1: Assist (Weeks 1–6)
- Select one domain with known pain, such as customer or product data feeding an AI assistant or a key report.
- Use LLM-assisted profiling to classify fields, detect sensitive data and identify reference data domains.
- Generate candidate rules and have data owners review them in a structured way: accept, edit or reject, with a reason.
- Track the acceptance rate. If owners reject most suggestions, improve the context you give the model (documentation, glossary terms, examples of good rules) before scaling.
Stage 2: Triage (Weeks 6–12)
- Connect an agent to monitoring alerts, lineage and change history with read-only access.
- Have it group alerts into incidents, propose a likely cause with evidence, and identify affected consumers.
- Ask on-call engineers to rate each summary as useful, partly useful or wrong. This becomes your quality signal.
- Measure time from alert to confirmed root cause, before and after.
Stage 3: Supervised Remediation (Months 3–6)
- Pick narrow, reversible fix types first: standardizing formats, mapping known synonyms to reference codes, filling attributes from an authoritative source.
- Let the agent propose changes as drafts or tasks with its reasoning attached; a steward approves in bulk or individually.
- Log every change with before and after values, the rule or reasoning used, and the approver.
- Promote a fix type to auto-apply only after it has a sustained record of approvals with negligible reversals, and keep sampling audits in place.
Stage 4: Governed Autonomy (Ongoing)
For a small set of well-understood, low-risk fix types, agents act automatically within limits: a maximum number of records per run, specific fields only, and automatic rollback if downstream checks fail. Everything else stays supervised. Review the boundary quarterly.
Reference Architecture: How the Pieces Fit Together
AI-assisted data quality works best as a thin intelligent layer on top of the deterministic machinery you already have, not as a replacement for it. A practical architecture has five components.
- A deterministic quality engine. Rules run where the data lives, in the warehouse or lakehouse, in transformation code, or in a data quality platform. This is where bulk validation happens cheaply and reproducibly.
- Active metadata. Schemas, profiles, glossary terms, lineage, ownership and past incidents, kept current automatically. This is the context the model needs; without it, suggestions are generic.
- A model gateway. A controlled point through which agents call LLMs, with masking of sensitive values, logging of prompts and responses, cost limits and the ability to swap models.
- Agent tools with scoped access. Agents act through defined tools, for example “get profile,” “propose rule,” “create steward task,” “apply standardization,” exposed via APIs or an MCP server that enforces the same permissions as the user interface.
- A human workflow. Stewards and owners review suggestions and approve changes in the tools they already use, with reasoning attached and decisions recorded.
Two design choices matter more than model selection. First, send models metadata, profiles and small masked samples rather than full datasets; this keeps costs predictable and limits privacy exposure. Second, capture every human decision on a suggestion, accepted, edited or rejected, as structured feedback. Over time, that feedback becomes the most valuable asset in the program, because it encodes what your business actually considers good data and lets you improve prompts, evaluation sets and rule libraries.
On model choice itself, general-purpose frontier models handle rule drafting and incident summaries well, while smaller models deployed in your own cloud tenancy are often sufficient, and cheaper, for classification and standardization tasks at volume. Keep the gateway model-agnostic so you can change the mix as prices and capabilities shift.
Guardrails That Make AI-Assisted Quality Safe
The controls below are what separate a credible program from an experiment that someone will eventually switch off.
- Scoped permissions. Agents operate under the same roles and permissions as human users, never with broad database credentials. Read-only by default; write access is granted per fix type.
- Deterministic rules for high-stakes decisions. For decisions with financial or regulatory impact, such as merging customer records, the decision logic should be defined and reproducible. AI can suggest and explain, but the rule decides.
- Explainability by default. Every suggestion or change carries its reasoning and the evidence used, visible in the tool stewards already use.
- Reversibility. Keep history so any automated change can be rolled back, individually or by batch.
- Blast-radius limits. Cap the number of records an agent can change per run and stop automatically on anomalies.
- Privacy controls. Decide which sample values may be sent to which model, mask sensitive fields, and prefer models deployed in your own cloud tenancy for regulated data.
- Evaluation sets. Maintain a labeled set of known issues and correct fixes to test agents whenever prompts, models or rules change.
Worked Example: Supplier Master Data Before an AP Automation Rollout
Imagine a manufacturer preparing to roll out an accounts payable agent that matches invoices to purchase orders and suppliers. The supplier master has around 30,000 records from three ERPs, with inconsistent names, missing tax IDs and duplicate entries.
In Stage 1, LLM-assisted profiling classifies fields across the three systems, recognizes that “VAT_NO,” “TaxRef” and “tin” hold equivalent identifiers, and proposes about 60 rules, from tax ID format checks by country to “payment terms must be one of the approved codes.” Procurement and finance owners accept roughly two-thirds, edit some, and reject a handful that simply encoded current bad practice, such as allowing blank bank country.
In Stage 2, a triage agent watches nightly loads. When an ERP upgrade changes a date format and triggers a burst of alerts, the agent groups them into a single incident, links it to the deployment, and lists the AP reports affected. Investigation time drops from an afternoon to minutes.
In Stage 3, an agent proposes standardized supplier names and country codes, and flags probable duplicates with evidence: same tax ID, similar names, same bank account. Stewards approve clear standardizations in bulk. Duplicate merges, which affect payments, remain under deterministic matching rules with steward approval. By the time the AP agent goes live, it reads from a cleaner supplier master with a documented quality baseline, and the team has data showing which fix types are safe to automate further.
The numbers in this example are illustrative, but the pattern is the important part: assist first, triage next, remediate under supervision, and automate only what has earned it.
Measuring Whether It Works
Track a small set of metrics from day one so the program is judged on outcomes, not novelty.
- Rule coverage of critical data elements, and the acceptance rate of AI-suggested rules.
- Alert precision: share of alerts or incidents that led to real action.
- Mean time to detect and resolve incidents on critical data products.
- Steward throughput: issues resolved per steward per week, and backlog age.
- Reversal rate of automated or approved AI changes, the key safety indicator.
- Downstream impact: incidents caught before consumers noticed, and quality scores of datasets used by AI systems.
Pitfalls to Avoid
- Certifying the status quo. Rules inferred purely from current data can enshrine existing errors. Always pair suggestions with business review.
- Letting models decide identity. Probabilistic AI judgments about whether two customers are the same person are hard to audit. Keep matching logic explicit and use AI to explain and prioritize.
- Skipping evaluation. A prompt or model update can quietly change agent behavior. Re-run evaluation sets on every change.
- Ignoring stewards. If agent tasks land outside stewards’ normal tools, adoption collapses. Integrate suggestions into existing workflows.
- Underestimating cost. Running LLMs over every row is expensive and unnecessary. Use them on metadata, samples and exceptions, and leave bulk checks to deterministic engines.
Outlook and Conclusion
Over the next two years, expect data quality tooling to look less like a rules engine with a dashboard and more like a supervised team of agents: one profiling new sources, one drafting rules, one triaging incidents, one proposing fixes, all reporting to human owners. Analysts already describe AI for data management and data management for AI as two sides of the same shift, and MCP-based access is making it easier for agents to query quality status and act within permissions.
The organizations that benefit will be those that treat AI as a force multiplier for accountable people, not a replacement for accountability. Start with suggestions and triage, where AI saves time without changing data. Measure acceptance, precision and reversals. Expand autonomy only where the evidence supports it. Done this way, AI-assisted data quality closes the coverage, triage and remediation gaps that have held programs back for years.
How Hanoi Technology Can Help
Hanoi Technology helps teams introduce AI-assisted profiling, rule generation and supervised remediation with the guardrails, evaluation sets and metrics that keep data quality accountable.