Series
The 360 View
Part 1 of 2 — the leadership read. Part 2 is the technical read: Anthropic's methodology, Snowflake Cortex Sense, and our context engine, side by side, so your data team can check our work.
For a decade the work was moving data: out of the server room, into the cloud, into one warehouse, secured and governed. Companies spent fortunes and mostly got there.
The data is valuable. But it is not the gold. The gold is what your people built on top of it: the accumulated decisions and judgment that turn a warehouse of raw tables into insights a business will actually bet on. Call that layer what it is, because every organization is trying to build one now: your brain. And here is the problem. You moved the data to a centralized, secure, accessible place, but the brain never migrated anywhere. It is your real intellectual property, and right now it is unreadable, unowned, and about to be quietly enclosed inside someone else's product.
This is a piece about that enclosure: where your brain actually lives, who is moving to capture it, and what it takes to keep it yours before the window closes.
You connected everything to your AI except the one thing that runs the business
Look at what you have already wired into your AI over the last two years. Your wiki. Your Drive and your SharePoint. The decks you send clients. Your tickets, your Slack, your meeting notes, your legal folder. Every one of those became something a model could read, and you did the work to connect them, because you understood that your institutional knowledge is worth more when your AI can actually use it.
You left one out. The biggest one. And it is the only one that does not just record what your people said, it records what your business actually decided.
Your SQL.
Every one of those is a record of intention. A wiki is what someone meant to write down. A deck is what you wanted a client to believe. Slack is what people said in the moment. Put together, they capture the why and the what of every analysis your business ever ran: why the question mattered, and what the answer turned out to be. What none of them capture is the how, the steps that turned the question into a number someone was willing to bet on. All useful, all partial, because people document what they remember, and they remember badly. Your query history is different in kind. Every query your team ever ran is a decision, in executable form, proven by the fact that it ran and someone trusted the answer. It is the working definitions your business actually operates on, not the ones the handbook claims, along with all the hard-won caution your veterans apply on instinct and never wrote down. Years of your organization's real judgment, recorded by accident, one query at a time, by the people who understood the business well enough to interrogate it.
Weigh the two piles and the imbalance is stark. Everything you connected is the part of your organizational brain that talks about the work. The how, the part that actually produces the deliverable, is the other 80%, and today it lives in exactly two places: your query history, unread, and your best analysts' heads, unwritten. Which is why every answer your AI gives still has to pass through a person to become something the business can act on.
That query history is the richest body of institutional knowledge you own. Almost nobody treats it as one. That is the whole opportunity, and as of this year, the whole fight.
A brain that remembers is not a brain that acts
Here is what the brain you have already built gets you in practice, without the how.
Someone asks your AI how the team built the customer cohort for last year's board analysis. It finds the deck. It summarizes the approach beautifully: why the question mattered, what the segments were, what the topline said.
Then it stops. The actual steps, the query logic that produced those segments, the exclusions someone fought over in week three, the join that silently drops records unless you handle it, none of that was ever captured. So the answer gets handed to an analyst, and the analyst does what analysts have always done: reads the deck, interprets it, rebuilds the logic from scratch, and hopes the first version comes out clean. It usually does not.
You can have all the nice knowledge in the world. If your brain does not know how to turn a question into a query, it is a brain that can remember and cannot act. That is not a brain. That is a search box in front of the same bottleneck.
The market just settled whether this matters. The fight now is whose it is.
For two years, "give your AI context about your business" was a niche idea almost nobody was buying. In 2026 it became the entire agenda, in the span of a few months.
In February, Snowflake shipped a tool that auto-builds semantic definitions from your SQL and BI assets, and thirty-plus companies signed onto an open standard for those definitions, the Open Semantic Interchange. On June 3rd, Anthropic published how it runs its own analytics on Claude: accuracy went from 21% to above 95% purely by giving the model structured business context. Their words, not ours: analytics accuracy is "a context and verification problem, not a code generation issue." A day earlier, at their Summit, Snowflake announced Cortex Sense, a context engine that mines your query history automatically. Databricks is moving the same direction.
And it is not only accuracy. AtScale published a benchmark from a Tier 1 bank that ran the same five questions two ways: $17.93 per run without a semantic layer, $0.0008 through one. Same questions, same warehouse, a twenty-thousand-fold difference, because an AI with context asks the database for the answer instead of rummaging through it. Context does not just make AI right. It makes AI affordable enough to hand to everyone.
The vendors themselves have stopped hedging. IBM, Databricks, Snowflake, and ThoughtSpot spent this year's conference circuit calling this layer essential infrastructure, no longer a nice-to-have; one enterprise customer now calls it "the operating system for data access." And when AtScale's CTO listed what actually differentiates an AI agent, he named the fuel in order: deep metadata, query history, and semantic context. The whole market is converging on the same sentence we opened with: the knowledge that matters is the kind your teams already wrote.
Here is what that wave means, said plainly. The question is no longer whether this layer matters. The most sophisticated data teams on earth just spent millions settling that one for everyone. The only question left is who owns it, and that is the question this piece is about.
A 30-second glossary, if this is not your world
Query history: the log of every SQL question anyone ever asked your data. Your company's analytical diary, written by accident, one entry at a time, for years.
Semantic layer: a dictionary of official metric definitions, so people and AI compute the same number for "revenue."
Context layer: everything around that dictionary that makes it trustworthy, where a definition came from, who uses it, how the competing versions differ, the filters and joins your veterans apply on instinct.
OSI (Open Semantic Interchange): an open, vendor-neutral file format for these definitions. Think PDF, but for business logic: any platform can read it.
Agent: AI that does not just chat, it does the work, writes the SQL, runs it, returns the answer.
Snowflake built a real translator. It only ever speaks Snowflake.
Snowflake built an automatic business translator for its own AI assistants. Turn on Cortex Sense and it studies how your analysts have queried data, plus your table descriptions and your Power BI and Tableau dashboards, and it teaches Snowflake's AI what your company means by "revenue" or "churn." It works. Setup takes about a day. It took their agents from 47% to 83% accuracy on their own benchmark.
We are not going to tell you it is bad, because it is not. If all you want is for Snowflake's agents to answer better, and you are certain you will never run anything but Snowflake, it may be all you need.
Here is the catch. Everything Cortex Sense learns lives inside Snowflake, works only with Snowflake's agents, and, as announced, cannot be exported, moved, or read directly by you. The metrics your team fought over and finally agreed on. The filters your best analyst applies without thinking. The join nobody documented. Years of judgment, written by your people, on your payroll, distilled and then handed back to you as a slightly smarter search box.
Data gravity was the old lock-in. Moving petabytes is expensive, so you stay, and everyone made peace with that. This is a different kind of gravity, and it is worse, because the thing being held is not your storage. It is your company's brain.
What a brain does for you, once it is yours
Set the vendors aside for a moment, because the real reason to care is not what someone might do to your knowledge someday. It is what that knowledge could be doing for you today.
Once your query history is mined into something real, structured, portable, and yours, it stops being fuel for one chatbot and becomes infrastructure for the whole company. The same brain that grounds an AI assistant also does a job in almost every corner of the business:
It keeps the knowledge when the people leave. It onboards a new analyst in weeks instead of months, because the proven patterns and the quiet gotchas are finally written down. It insures you against your best analyst walking out the door with the only working copy of how the business runs in their head. And this is not hypothetical:
Five analysts. Three months. One hundred thousand queries. 88% of the metrics we found were known to exactly one person. When that team's leader asked the finished brain who his most indispensable analyst was, it named someone who had asked to transfer teams the day before. He had no idea.
It saves you real money now. Your history shows which tables nobody has touched in a year, so you stop paying to maintain them, and which expensive query runs a thousand times a week, the one worth materializing or caching. When you face a migration or a warehouse consolidation, it scopes the job by revealing the ten percent that is actually used instead of forcing you to move everything blind.
It makes you governable and audit-ready. When someone asks how a board number was really produced, or which cohort definition made it into a regulated submission, your query history is the record of what was actually computed, by whom, and how often.
And it grounds all of your AI, not one vendor's. The same knowledge feeds your Claude, your Cursor, your internal copilots, and whatever you adopt next, because it is not trapped inside any single one of them.
None of that is possible from inside a locked box. A single-purpose engine can bolt on an export button tomorrow and it still will not become a company-wide knowledge asset, because it was only ever built to feed one thing. That is the real cost of the locked box. Not that you might get trapped someday. That you are leaving nine of ten uses on the floor today so you can have the one the platform picked for you.
The three questions that settle it
You do not have to take our word for any of this. Ask anyone who wants to mine your query history three questions:
Can I export what you build?
Can another platform's agent read it?
Can I read it myself, line by line?
For a platform's built-in context engine today, the honest answers are no, no, and no, which means the knowledge it builds can only ever do the one job the platform wants. For what we build, the answers are yes, yes, and yes. That is the whole difference, and everything about how we position ourselves flows from those three yeses.
How we are different: everything we build, you keep
You own the output. We produce files you keep: an open, vendor-neutral semantic model, plus versions that plug straight into Snowflake, Databricks, and dbt. If we disappeared tomorrow, nothing would happen to your context layer. It does not run on our servers or expire with a contract. We built it that way on purpose, because knowledge you cannot keep is not really yours. And this is not a gap a platform closes with a feature. Even if one adds an export button tomorrow, a copy of a model built to feed their own agent is not the same as a brain built from the first line to be yours and to work anywhere. The difference is not the file. It is what the thing was designed to do.
You can prove every answer. Every definition we hand back carries its evidence: this is how the metric was calculated, in these queries, by these people, this many times. A black box cannot do that. In pharma, finance, and healthcare, where a definition can end up in a regulatory submission, "trust the ranking" does not survive an audit, and "here is the query that proves it" does. And when your teams disagree, and they will, we hand you every competing version side by side so your people decide which one is true, instead of a machine quietly picking the most popular. They disagree more than anyone expects: in that same three-month history we found 182 metrics with more than one definition, some as complex as cohort logic, some as simple as a patient count, one divided by four and by two hundred with no explanation anywhere.
You can use it anywhere. A platform engine sees that platform. We read any SQL environment, Snowflake, Databricks, BigQuery, Redshift, on-prem. If you run more than one system, and almost everyone does, there is knowledge a single-platform tool will never see.
You keep it in your building, and it is lighter than it sounds. We read the text of your queries, not your data, so no rows ever leave your environment and no warehouse compute gets burned. The work is ordinary lightweight processing that runs in a single container on hardware you already have, because the heavy lifting is deterministic decomposition of query text, not a model grinding through a hundred thousand queries one by one. That is the difference between a fast local run and a six-month security review of a third party ingesting your data.
We go deeper, and the how is in Part 2. We treat your query history the way modern AI treats any pile of unstructured knowledge. On its own it is a mess: thousands of one-off queries, each a single answer to a single question, almost never repeated word for word, which is why reading it at the surface tells you little about what is reusable. So we do not read it at the surface. We decompose it into its smallest meaningful elements, analyze them, and study the patterns across your entire history, which is how we surface the proven, battle-tested logic a whole-query view slides past. That claim deserves proof, and Part 2 gives it, next to exactly how Anthropic says this should be done and what Snowflake ships today.
We will be fair about the giants
Cortex Sense is good. One-day setup, automatic refresh, no maintenance, a real accuracy jump for teams living entirely inside Snowflake. If all you want is for their own agents to answer better, and you are sure you will only ever be a Snowflake shop, and your auditors are fine with a black box, it may be enough.
We think most serious data organizations want more than that. They want to own their knowledge outright, prove where every number came from, use it across whatever platforms they run today and tomorrow, and put it to work in the nine other places knowledge like this belongs. That is the part the platforms cannot hand you, and it is the part we do.
The knowledge is already written. The question is who keeps it.
The context is already yours. It was written by your people, in your queries, over years. You connected your documents, your tickets, your decks, and your conversations to your AI because you understood that knowledge is worth more when everything can use it. Your SQL is the same, only richer, because it is the how behind every why and what you already connected: the record of what your business decided, not just what it said.
It has been doing one job, badly, for years. It can do ten. It just has to be yours first.
Part 2 is the technical read: how Anthropic says a context layer should be built, what Snowflake Cortex Sense does under the hood, and how our approach compares, so your data team can validate everything claimed here.
We are building the context infrastructure to make your analytics memory durable, traceable, and reusable. Implementing AI in your analytics team? Let's talk.