The Steward That Never Sleeps

What an AI agent can and cannot govern in modern data stewardship

INNOVATIONS

OF THE WORLD

FOR TODAY'S BIG THINKERS
Global village Globe

As Featured In:

Global Innovation Spotlight

Global village Globe

As Featured In:

Global Innovation Spotlight

A spreadsheet doesn't know who is touching it. Neither does a database, not really, not until someone asks it a question it can't answer honestly. That has always been the quiet trouble sitting under data governance: chasing down who owns a number and whether it can be trusted by the time it lands in a board deck. For years the job belonged to people with clipboard patience, hunting the truth across systems that rarely agreed with each other. Now a different kind of worker has entered the room. One that reads access logs instead of sipping coffee before the 9 a.m. stand-up. The question worth asking isn't whether it can help. It's how far that help actually reaches.

Something has shifted, too, in how companies vet outside expertise for their information. A firm hired as a data analytics service provider used to be judged mostly on modeling skill and turnaround time, and that yardstick has quietly gotten longer. Clients increasingly want the same partner to also police the pipeline, watching for definitions that drift and access that quietly widens, the sort of thing that otherwise surfaces only after a customer complains. The expectation now stretches toward a provider of data analytics services who also stands watch over what it touches, not just a vendor that hands back a model and walks away. Whether that trust is earned depends entirely on what the agent doing the watching can actually see.

The Old Job, Handed to a New Worker

A data steward, in the traditional sense, spends the day arbitrating small disagreements. Marketing calls someone a customer the moment they fill out a form; finance waits until money changes hands. Someone has to decide which definition survives in the dashboard everyone else reads, then defend that decision when a director asks why last quarter's figures don't match this quarter's story.

Unglamorous work, mostly, built from patience and a decent memory for exceptions. Agentic AI has walked straight into that job description. Systems built on newer language models can scan a data catalog, flag a column that has drifted outside its documented range, and route the exception to the right team before a human notices anything is wrong. They can watch for access patterns that look off, say, someone pulling 10 times their usual volume of records at 2 a.m., and revoke a credential on the spot. None of that used to happen in real time. Most of it surfaced weeks later, in a retrospective, after the damage was already done.

Where the Machine Earns Its Keep

Give an agent a narrow, well-documented task, and it tends to do it faster and more consistently than the team that used to do it by hand. Cataloging is a good example. So is lineage tracing: the unglamorous business of proving a number in a report actually came from where someone claims it came from. A handful of jobs now fall naturally to software rather than staff:

  • Classifying new data as it lands, based on sensitivity and existing policy
  • Flagging schema drift and unauthorized access close to the moment it happens
  • Reconciling conflicting records across systems that were never built to talk to each other
  • Keeping an audit trail detailed enough to survive a regulator's questions

That shift is showing up in the data, not only in the pitch decks of any data analytics services provider trying to win new business. Databricks' 2026 State of AI Agents report found that multi-agent systems grew by 327% in under 4 months across its customer base, and that organizations with formal AI governance in place moved roughly 12 times more projects into production than those without it. Governance, in other words, isn't the brake pedal executives once assumed it would be. It looks more like the thing that lets the car go faster without leaving the road.

The Line It Cannot Cross

None of which means the agent understands what it's protecting. Ask it whether a "customer" in the CRM means the same thing as a "customer" in the support ticketing system, and it will answer with whatever definition it was given, confidently, even where that definition is wrong for the question being asked. Judgment calls that once lived in a steward's head, built from years of watching exceptions pile up, don't transfer cleanly into an outside analytics partner's policy document.

The gap shows up in survey data, too. Deloitte's State of AI report, based on more than 3,000 leaders across 24 countries, found that roughly 75% of companies expect to be using AI agents at least moderately within the next 2 years, while only 21% currently have a mature model for governing those agents.

A separate research note from the Cloud Security Alliance sharpens the point. Drawing on an EY and AIUC-1 consortium survey, it found that only 38% of organizations monitor AI activity end-to-end across prompts, tool calls, and outputs, with just 17% keeping any continuous watch on how agents talk to other agents.

An agent that cannot be watched is an agent operating on trust alone. Not a comfortable place to leave sensitive data.

Building the Partnership, Not the Replacement

Stop waiting for agents to replace human judgment. The right architecture looks like a sharp junior analyst reporting to a veteran who knows when the numbers are lying.

Leading firms like N-iX are designing the missing infrastructure between raw databases and autonomous systems. That means building strict policy boundaries, real-time drift monitoring, and escalation triggers that hand off ambiguity to people. When you evaluate a data analytics service provider right now, technical capability is table stakes. You're paying for their wisdom — knowing precisely where to kill automation before it creates a liability. That distinction will dictate who survives as AI tasks drift further from anything written in a compliance manual.

Conclusion

An AI agent can watch far more than any single steward ever could, and it rarely tires of watching. What it cannot yet hold is the kind of context a person carries after years inside a business: a working sense of which exceptions matter and which don't. The honest answer to whether an agent can be a data steward is a partial one. It can keep the ledger, but someone still has to read it. Businesses that pair the two, rather than choosing between them, are the ones most likely to trust what their data tells them.

Other INNOVATE® Ecosystems