Atlan's Austin Kronz on why context is the one thing AI can't commoditize

Every enterprise shipping AI agents runs into the same question: the model reasons well, but how do you make it understand the way the business actually works. Austin Kronz, who leads AI and data strategy at Atlan, sat down with our co-founder and CEO, Gorkem Sevinc, to talk through what closing that gap actually takes.

Their conversation covered why "context" is the natural evolution of the data catalog in a machine world, what happened once AI took over the documentation nobody wanted to write, why context decays the same way data does, and why the chief data officer who becomes a ticket-taker gets left out of the next decision.

Here's what we learned from their conversation:

1. Context is the catalog's evolution for a machine that can't infer what nobody wrote down

Atlan didn't pivot away from data catalogs, it treats context as their evolution. A data catalog helped people discover, trust, and use data they didn't build themselves. An AI agent needs the identical thing, except none of the tacit knowledge that used to live in someone's head, a Slack thread, or an unwritten norm survives the handoff to a machine on its own.

🔸"Data catalogs were to people what context will be for AI."

2. Once AI could document a business, the resistance faded

The obvious barrier for most of Atlan's customers was that they didn't have documentation. Atlan built context agents to write that documentation, then pushed customers past the instinct to review every line before trusting it.

🔸"You don't need to be in the loop. You can be on the loop."

Atlan ran what it calls context sprints across roughly 250 customers, and the result was hard to argue with. Kronz estimates the sprints saved something like the equivalent of 10 years of collective working time, and 89% of Atlan's customers came back afterward and asked it to just rewrite everything a human had previously written. The apprehension about whether AI-written documentation would be good enough gave way once teams saw the output of a harness Atlan had spent 18 months building.

3. Context breaks into three parts, and two of them usually aren't written down anywhere

Kronz frames context as knowledge (what a term means and how a metric actually gets calculated), expertise (the tacit, procedural judgment a team builds over time, like knowing Europe's holiday calendar will dent this quarter's pipeline), and norms (how two teams can each correctly calculate the same metric differently, on purpose, because their goals differ). Documentation captures some of this. Plenty of it lives only in the way a company actually runs, not in what it wrote down about how it runs.

🔸"If you want to know how a company thinks they operate, go read their documentation. If you want to know how they actually operate, go to their systems."

4. Context decays exactly like data does, which means it needs its own lifecycle

Kronz draws a direct line between data drift and context drift. A skill built for an agent reflects the business as it existed the day someone wrote it, and the business keeps moving. Without active management of who built a given skill, when it was last touched, and whether it's still producing the right result, teams end up back where they started, waiting for someone to notice the skill went stale and fix it by hand.

🔸"Context, much like data, can drift and get stale."

The failure mode Kronz sees most often is a team hardcoding a new acronym or definition straight into a configuration file instead of updating the context layer. The fix doesn't travel the next time someone builds the same agent on a different platform.

5. Intelligence got commoditized, so context is the only differentiation left

Every company building agents today has access to roughly the same models its competitors do. What separates one company's agent from another's is the specific way that company runs its business, and Kronz is blunt that a semantic layer on top of wrong data can't produce a right answer.

🔸"Your secret sauce is your context. There's no semantic layer in the world that can give you an answer on bad data."

Sevinc's read on this pairs directly with Kronz's: trusted context and trusted data are two halves of the same problem, and treating an AI agent and a human analyst as equally important consumers of that context is what makes the combination hold together end to end.

6. The CDO who becomes a ticket-taker gets left out of what's next

Kronz has watched the economic buyer shift from a chief data officer with a governance mandate to a chief data officer with an AI mandate. Different parts of an organization are standing up their own agent development platforms, whether that's a team on Databricks Genie, one on Snowflake Cortex, or an engineer working locally in a code agent, and Kronz calls that sprawl inevitable. The unresolved problem is portability: context hardcoded into one platform doesn't move when the business builds its next agent somewhere else.

🔸"If you become the team that just says, 'submit a ticket, we'll get you the data,' you'll be left out of what's next."

Sevinc offered the clearest picture of what maturity looks like from his own field, a highly regulated financial services customer running on-premises with no internet access. Someone took a 170-page document from the Fed, ran it through an LLM, produced a mapping file, and generated the SQL to run the resulting data quality checks directly in their Teradata instance. That one experiment opened budget the team hadn't had access to before, because it let them stay on the loop instead of trusting the output blindly, and it pushed the team toward the next stage, deploying LLMs into production and then agents.

Context is everybody's job too

Kronz's prediction is that AI maturity across companies is skewed left today and will keep shifting right, not evenly, but inevitably. The companies that treat context as a governed, living layer rather than a one-time documentation project are the ones whose agents actually reflect how they do business.

Sevinc's question from earlier in the conversation is worth sitting with longer than the interview gave it: are you building the same trusted context for your AI agents that you'd insist on for a human analyst, or are you still building for one and hoping the other catches up?