A new arXiv paper sets out a blueprint for agentic data governance: AI agents argue with citations, human stewards decide and every edit is kept on file.
Overview
AI agents are getting good at the unglamorous side of data work: tagging records and linking fields to the documents that explain them. But who checks their work, and how would anyone know later what was checked? That is the core question of agentic data governance, and a new paper on arXiv, Contestable Agents for Data Governance, posted on October 22, 2026 by Humza Akhtar (MongoDB) and Brian Freeman (Trane Technologies), proposes an answer: make the AI show its reasoning, let a human expert challenge it piece by piece, and keep every step on file.
The examples come from heating, ventilation and air conditioning (HVAC) equipment, but the lessons apply to AI data governance wherever knowledge lives in both databases and documents.
The problem: approved, but by whom and why?
At an HVAC manufacturer, connected air handlers stream readings from hundreds of data points per unit, while the meaning of each point sits in manuals that vary by model and firmware. Warranty systems hold claim codes; technicians describe what broke in free-text notes. These sources live in different tools with different owners, and only a few experts know how they connect.
Mistakes are expensive. A mixed-air temperature sensor labeled as discharge air could make a monitoring system read normal operation as a failed heating coil and send a technician out with the wrong diagnosis. Because labels get copied to every unit of the same model, one error can spread across a whole fleet.
Agentic data governance hands this labeling to AI agents, but people tend to approve automated suggestions with one click and little scrutiny. The record then shows that someone approved, not what they checked. And when a reviewer fixes a mistake in an email or ticket, the agent never sees it and repeats it.
What “contestable governance” means
The authors’ answer, drawn from research on contestable AI, has three parts:
- Agents argue, with receipts. Specialist agents each examine different evidence, such as the point name or the manual, and submit scored, cited arguments for or against each option.
- A human edits the reasoning. A data steward can accept, reject, modify or add individual arguments. Scores update after every change, and nothing reaches the data catalog until the review ends.
- Everything goes on file. A governance record keeps the evidence, every argument (rejected ones included), each human edit with its reason, and the outcome. An auditor can pull the full history with one database query.
A versioned glossary helps too, because abbreviations change meaning between models: “DISCH” points to different spots on different air handlers, and reviewers can challenge its entries.
Saving expert time
Not every decision needs a debate. To keep agentic data governance affordable, the authors propose triage: a simpler model labels clear cases and only uncertain ones are argued out in full. Clear cases still go to a steward, who can escalate any of them.
A promotion gate controls what the system learns: only verified decisions can become precedents for the next unit of the same model, and demoting one flags every decision that relied on it.
Seven requirements for agentic data governance
The paper sums up the approach in seven requirements, a ready-made checklist for any team that lets agents touch its data:
- Cited evidence. Every proposal names the fields and documents it relies on.
- Contest before commit. A reviewer can challenge a proposal before it is saved.
- Clear attribution. Each decision records the agent, the model version and the named reviewer.
- Append-only history. Past entries are added to, never edited or deleted.
- One-query lineage. A decision’s full trail, from source evidence to final label, comes back in a single query.
- A gate for precedent. Only verified decisions can guide future cases.
- Measurable interventions. Reviewer edits are stored as data, so teams can see, for example, which agent gets overruled most.
Example 1: a sensor that wasn’t what its name said
The open reference implementation, built with CrewAI agents, GPT-4o from ChatGPT maker OpenAI, Voyage AI embeddings and MongoDB, uses a synthetic site with 247 data points, 18 of them flagged as ambiguous. It walks through one, AHU-1-DISCH-T, which needs a standard label from Brick Schema, a shared vocabulary for building equipment.
A “String Matcher” agent sees only the name and favors discharge air. A “Documentation Reader” agent finds section 4.2 of the manual, which places the sensor in the mixed-air chamber. After the debate, discharge air leads by a hair, 0.045 to 0.044. The steward rejects the name-only challenge, lifting mixed air to 0.821, then adds an argument citing the manual, taking it to 1.758. The record names both agents and the reviewer and keeps the rejected reasoning.
Example 2: a warranty claim that blamed the wrong part
In the second case, a warranty claim lists a supply-air temperature fault and a replaced sensor, so “sensor drift” looks obvious. But telemetry shows a damper that doesn’t move when told to, a technician notes a loose damper linkage and says the new sensor didn’t help, and a service bulletin on sensor drift only covers serial numbers ending 1000–1999. This unit ends in 2407.
Before review, sensor drift leads, 0.907 to 0.839. The steward rejects the bulletin argument as out of scope and adds the technician’s evidence. Sensor drift falls to −1.016, and the unit is tagged with damper linkage failure. The original claim code stays untouched.
What the paper doesn’t show
The authors flag the limits:
- Both cases use synthetic data, and the scores are illustrative and uncalibrated. They show how the system responds to edits, not how accurate it is.
- Triage, the promotion gate and append-only history are still proposals, not yet in the code.
- No one has measured the effect on review time or decision quality.
As the authors note, an auditable record makes a decision traceable. It doesn’t make it correct.
Putting agentic data governance into practice
-
Ask your AI tools for receipts.
Every proposed label should cite the fields and documents behind it.
-
Let reviewers challenge reasons, not just results.
An approve button records a click. Editing individual arguments records judgment.
-
Store corrections where agents can see them.
A fix buried in a ticket is a fix the agent will repeat.
-
Gate what becomes precedent.
Decide who promotes verified decisions, and make demotion possible.
Bottom line
This is a design paper, not a benchmark: a clear blueprint for agentic data governance that keeps humans in charge, plus open code to try part of it. If agents already write to your data catalog, whether commercial tools or open-source AI agents, start by checking your setup against those seven requirements. Proof that the approach saves time and catches more errors is still to come.
