
Right now, every vendor wants to sell you an ontology. Each one will say it holds your business context, and each one will be telling the truth about its own platform. Your CRM defines “customer” one way, your billing system another, and your data warehouse a third.
But your agent doesn’t work in only one platform. It reaches into several, with access to all of them, and picks up whichever definition each one hands it. The question is whose context goes with it.
We’re tired of context too
We’ll admit it. We’re tired of hearing about context, and we suspect you are too. A first draft of an ontology takes minutes now. What takes months is getting Sales and Finance to agree on which definition to use.
We’ve seen that pattern for years helping customers build data products. Publishing the data is the easy part. Getting the business to agree on what it means, and keeping that agreement current as the business changes, is where the real effort goes. Every ontology a vendor generates for you runs into the same wall.
Then each platform asks you to rebuild that hard-won agreement in its own tool and its own format. Every rebuild drifts a little from the last, and the business ends up defending several versions of the same definitions. That’s what wears people out. It’s also what your agents inherit.
What autonomy takes
Most enterprises we talk to have agents running. Very few let them run without a person checking each step, and the checking is rational. In VentureBeat’s October 2026 Pulse survey, 78% of respondents with agents working on company data had traced at least one confident but wrong answer to missing or conflicting business definitions in the past six months. In 42% of the most significant cases, the wrong answer reached a customer or partner, landed in a financial, regulatory or board report, or forced a decision to be reversed.
Conflicting definitions are exactly what all that rebuilding produces. Picture an agent that answers a question across PostgreSQL, Snowflake and Salesforce. Along the way it meets three versions of your business, one per platform. Then an auditor asks the analyst to stand behind the result. Which definition of “flagged” did the agent use? Did it stay within what she’s allowed to see? What exactly did it read, and in what order?
Autonomy depends on answering all three questions, and each one breaks at a platform’s edge. Each platform can vouch for its own step, and nobody can vouch for the whole task.
Whose context goes with the agent
The fix for both problems is the same. Agree once. Approve once. Then let that one version go wherever the agent goes.
Context will stay as distributed as the data it describes, so it has to sit above the platforms. That starts with harvesting what your teams have already built. The definitions in your catalogs, semantic layers and transformation code are agreements someone fought for. They should be harvested from where they live, so nobody retypes them for each new tool. Someone accountable reviews what comes in and approves it.
Then the approved version has to show up at query time. When an agent needs to know what flagged means, it should get the certified definition, the right joins and the rules that apply, at the moment it runs, in whichever system it’s working in. The business stops rebuilding its definitions for each platform, and the analyst can give the auditor one answer.
That’s when teams can start taking people out of the loop.
Launch Point 2026
We’ve spent the past year building toward this. On Wednesday, November 18, at 12 p.m. ET (9 a.m. PT / 5 p.m. GMT), we will share what we have built at Starburst Launch Point 2026. Join us to see what happens when access to data and context are solved for enterprise AI.

