The context your warehouse can't reach
The two big data platforms have converged on the same diagnosis, and I think it’s the right one. Databricks spent its summit arguing that AI has a context problem rather than an intelligence problem, and that the context problem is what explains why agents perform badly. Snowflake is selling the Agentic Enterprise on the same premise, that enterprises will run fleets of agents and the platform that grounds them wins.
Both have built seriously against it. Semantic models and metric definitions so everyone agrees what a given metric measures. Glossaries and domains so an agent gets an authoritative definition. Lineage, query history, catalogs, governance over what an agent may touch. Ontologies that improve themselves from what the platform observes. This is the most sustained engineering anyone has done on grounding.
Consider what those context sources have in common. Pipelines, query logs, table lineage, metric definitions, governed objects. Every one of them is structured, and the self-improving ones improve from what the platform can see, which is the warehouse.
That is the boundary.
Where the meaning gets decided
Consider what happens when a number changes meaning.
Somebody in a thread says churn should exclude the accounts we migrated. Finance agrees in a reply. It’s decided, but the semantic model learns about it whenever a person remembers to update the governed object, which could be days or weeks later, or worse, never. Until then every agent grounded in the catalog answers with the old definition, governed and consistent but wrong, which is a worse failure than giving an obviously broken answer, because nothing about it looks wrong.
The same thing is true of everything a company decides rather than computes. Who owns renewals now. Which account is quietly shrinking. Why the Q3 target moved. What we agreed to do about the complaint. None of it starts life as structured data, and a lot of it arrives from outside: a customer mentions their expansion plans on a call, their new billing entity shows up in an email signature, the account that’s pulling back never says so outright, and you only see it by reading five separate threads together.
This is operational knowledge, and it sits between the two kinds of knowledge companies already keep. Structured records live in the systems built for them. Canonical knowledge, the mission and the positioning and the playbooks, changes rarely and survives in curated documents. Operational knowledge is neither. It is born in text, it goes stale in weeks, and where it lands, if it lands anywhere, is a wiki.
That only happens if somebody writes it up, and almost nobody does. Not because people are lazy. Because it is nobody’s job, it has no deadline, and the person who knows already knows. Which is why wikis rot.
So the platform can tell an agent what happened with total authority, but it structurally can’t tell the agent why, or what the customer said last week. Both live in text the platform has no access to.
Retrieval is not the same as knowing
The obvious objection is that unstructured data is already on the roadmap. Snowflake bought Datavolo in 2024 for exactly this, a pipeline layer built on Apache NiFi that moves documents and other unstructured sources into the platform. Snowflake’s own description of the deal was that it deepened their service in the “bronze layer” of the data lifecycle, which is the raw landing tier. Cortex Search then retrieves from what lands there.
Databricks arrives at the same place by a different route. Unity Catalog volumes hold files of any format alongside its tables, under the same governance, and its retrieval tools read them.
Those tools solve part of the problem, getting the text in and putting it somewhere governed. But they don’t affect the definition of churn that was discussed in email. What both platforms have afterward is a searchable corpus of messages, and querying a corpus is not the same as having a current answer.
Retrieval over raw messages resembles archaeology. You excavate, you find something that matches the query, and nothing guarantees that what you found is the most recent version. The thread where the price was floated and the thread where it was agreed both match a query about pricing, and the one that’s easier to retrieve is not reliably the one that’s true now. Ask again next month over a corpus that’s grown, and you can get a different answer with the same confidence.
Retrieval over maintained conclusions is recall. The distinction is the corpus underneath, not the retrieval. One is an undifferentiated record of everything anyone ever said. The other is a smaller set of statements that someone has decided are currently true, each with a history behind it that a person can correct.
An authored artifact does not solve it either, as the data platforms already know from their own glossaries. A semantic model is maintained by hand, and hand-maintained artifacts rot the way wikis rot, for the same reason the transcription never happens.
The layer that’s missing
What sits between those is something that reads the communications continuously, proposes what has changed, and asks a person to confirm it. Not a search index over the mail, and not a document someone owns.
That’s what we build. A brain, which is a set of pages holding what a company knows, assembled from the messages it already has across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, kept current by workflows that write back to it as they run, with a person approving or editing each update and each company deciding how long to keep that approval step in the loop.
This enables joining to data in the warehouse. When finance redefines churn in a thread, the brain reads that thread and can propose the semantic-model update the same day, for a person to approve. That is a write path the glossary does not have. The warehouse stays the authority on what happened. The layer above it records why it happened, and when that reason was last confirmed.
If agent quality is a context problem, the context that lives in communications is the unsolved half of it.