Your agents don't stop to ask

Across the enterprise data estates we've worked in, we have yet to find one with automated end-to-end column-level lineage: sources, through the warehouse, through ETL and ELT, all the way out to BI. Not partial lineage, and not lineage that stops at the warehouse boundary, but the whole path, maintained without anyone maintaining it.

It reads like a technical footnote. It is evidence of an assumption baked into two decades of data infrastructure: that somewhere along the path, a person would know.

Humans were the error correction

For twenty years that assumption held up fine. Lineage gaps mattered less than they should have, because an analyst who pulled a revenue figure that looked wrong would stop, frown, and go and ask someone.

Governance living in a policy document was survivable, because a steward reviewed access requests one at a time and applied judgment the document never contained. A definition that had drifted got caught eventually, by someone who remembered how it used to behave.

Humans were slow. They were also self-correcting, and the correction cost nothing.

Agents are fast and they are not self-correcting.

Taking the human out of the loop removes the latency, and it also removes the pause where somebody noticed.

The cost of that pause turns out to be asymmetric. An access rule that was written down but never enforced becomes a compliance exposure the moment an agent can reach the data. A definition that drifts on a core metric no longer produces one wrong answer; it contaminates every answer downstream of it, across every agent, for as long as nobody happens to look.

So the failure mode has changed shape. Agents get things wrong with nobody present to catch it, in a system that had always assumed somebody would speak up.

This is now the consensus problem

Gartner^® puts it directly in Coolest Vendor Innovations in Data Management:

"For D&A leaders, the challenge is no longer simply providing access to data; it is making sure the data is understandable, governable and actionable for AI agents."

The same research is blunt about where the work currently sits:

"Most data governance processes today are manual, creating too much work. And now, leaders need to manage AI governance too. Organizations need a way to automate and streamline their data governance processes. Ways to partially address this challenge include policy-driven governance, automated governance, and data contracts."

And on the raw material the whole thing depends on:

"To make accurate decisions, AI agents need semantic context, business definitions, and lineage. While technical metadata is typically easier to manage, organizations often lack good business metadata for understanding context. The metadata should be treated as a first-class citizen right alongside the data itself, not as an afterthought."

We would put the diagnosis more sharply. Humans are not in the loop; they are the loop — the mechanism by which context stays current and rules stay applied. That was a workable design when demand arrived at human pace. It stops working when the consumers of your data outnumber and outpace the people maintaining it.

It is also why so many agent programmes clear a proof of concept and then fail to scale. The POC ran on one schema somebody curated by hand, and nobody can hand-curate the next two hundred.

What context built for agents requires

Automated

Context has to be ready before it is needed, which means it cannot be built by anyone. A live context graph continuously reconstructs lineage, usage, health, ownership and business logic across the whole stack, with real-time impact analysis that reflects the estate as it stands rather than as it stood at the last scan. Nothing to curate, nothing to stitch together, no rescans.

The semantic question resolves itself here as well. Semantic definitions belong to the graph, which reproduces them whenever the estate underneath changes. You still get the semantics. What you skip is the semantic work, and what you gain is definitions that cannot fall out of date while nobody is watching, because nothing is holding them in place by hand.

Scoped

When an agent underperforms, the instinct is to give it more. Dump the schema, widen the window, hope it copes. This makes accuracy worse and costs more. What an agent needs is the exact slice its task requires, preassembled from the graph, with the right logic and the right definition and nothing extraneous.

It also means nothing is missing. Knowing that a table exists and what its columns mean is not sufficient if the table is unhealthy. Scope means delivering the precise context that makes an answer trustworthy, including the parts that tell an agent when the data underneath it cannot be trusted.

Governed

Governance policy is set by humans, deliberately, once. That part is correct and should stay that way. Nobody should want an agent inventing its own access rules.

The problem is enforcement. When policy lives in a document, enforcement means a person applying judgment case by case, and that mechanism cannot survive contact with agent-scale demand. What has to become continuous is the application of the rules: labels certifying what is AI-ready, persona-based access determining who receives which context and with which capabilities, and exposure tracing that shows where sensitive data actually reached. Set once by people, applied constantly without anyone re-checking.

This matters more because the systems consuming the context are non-deterministic. Run the same process twice and you can get two different answers, which is precisely why the context feeding it has to be the part you can account for.

This is still data engineering

There is a lot of noise suggesting all of this requires a new discipline, with new job titles to match. We are sceptical, and we would be wary of anyone selling one.

In 2020, data engineers built warehouses and semantic models as the rails for BI tools. In 2026, data engineers build lakes and context platforms as the rails for AI agents. The job is the job it always was: get accurate, current information to the part of the business that is about to make a decision with it.

What has changed is the stakes. There are more consumers, they move faster, they reach further into the business, and none of them will stop to ask whether the number looks right.

That is still the work, and it cannot be done by hand any more.

‍

Gartner, Coolest Vendor Innovations in Data Management, Nina Showell, Anurag Raj, Sarah Turkaly, Xingyu Gu, Robert Thanaraj, Michael Simone, Jenna Goodrich, 29 July 2026.

‍Gartner does not endorse any company, vendor, product, or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner's business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.

GARTNER is a trademark of Gartner, Inc. and/or its affiliates.

‍