Market briefing · July 2026
The whole industry just agreed with us.
In 2026, every major data platform independently arrived at the same conclusion: AI doesn't fail because the model is weak. It fails because nothing has told the model what your business means. Here is what they said, what they shipped, and what they still leave you to build.
Four competing platforms. One conclusion. Same twelve months.
The convergence
Four CEOs, one sentence.
These companies compete ferociously and agree on almost nothing. In 2026 they agreed on this.
"AI doesn't have an intelligence problem. It has a context problem."
"The model is not your unique advantage. It's when you combine models with your data that things begin to shine."
"AI doesn't become valuable when a model gets smarter. It becomes valuable when it's connected to your business."
The ontology exists "not simply to organize data, but to represent the complex, interconnected decision-making of an enterprise."
The number that matters
Snowflake benchmarked a raw MCP connection. It scored 23%.
ADE-Bench is a data-engineering benchmark built by Benn Stancil with dbt Labs — real projects, deliberately vague prompts, scored on whether the tests actually pass. Snowflake ran it and published the results.
A frontier coding agent reaching the warehouse through an external MCP connector completed 23% of the work. The same class of agent, given native access and a governed context layer, cleared 72%.
- The protocol was never the problem. Same model, same data, same questions. The only variable was whether anything had modeled the business underneath.
- Snowflake also measured its own context layer. On complex enterprise queries, accuracy went from 47% without it to 83% with it.
- Read it with your eyes open. This is a vendor running a benchmark that favours its own architecture. See the caveats below — we'd rather you trust the argument than the scoreboard.
Snowflake's published ADE-Bench run, June 2026. Not independently replicated.
What they built, not what they said
Every one of them shipped a semantic layer this year.
Marketing copy is cheap. Roadmaps aren't. The clearest evidence that a modeled layer is mandatory is that four competing platforms each spent a year of engineering building one.
Horizon Context & Cortex Sense
A governed semantic foundation — metric definitions, glossary, lineage — plus a runtime layer that assembles context from query history, metadata, and existing dashboards. Snowflake Intelligence was rebranded CoWork.
Summit 2026Genie Ontology & Unity Catalog Metrics
A self-improving knowledge graph that learns the business from Databricks data, dashboards, queries, and 50+ connected apps. Marketed on accuracy and lower token cost — the same economics argument we make.
Data + AI Summit 2026OneLake & cross-domain semantic models
Positioned as "your foundation for an AI-ready data estate." Metrics and relationships defined once, reused across every agent and Copilot. Fabric reframed as an operating system for data, context, and action.
FabCon 2026Knowledge Catalog
Automatically extracts semantics into a dynamic context graph that, in Google's words, grounds AI agents in enterprise truth and "drastically reduces hallucinations."
2026One company, one thesis
The two biggest names in ingestion and transformation merged on 1 June 2026 under a single banner: "the Data Infrastructure for Trusted AI Agents." Roughly $600M ARR, 100,000+ data teams.
Merger closed June 2026The ontology argument
The sharpest version of the case: a decision model, not a data model. Palantir's argument is that plain retrieval over a database produces little more than a better search engine, because acting on data requires knowing how decisions actually get made.
OngoingThe gap they all leave
They agree you need the layer. None of them build it for you.
Read the fine print on every one of these launches and the same word appears: authored, curated, defined, prepared. Horizon Context is where humans and governance teams curate business semantics. Unity Catalog Metrics is authored. Microsoft requires the star schema first. Palantir's ontology is famously a services engagement.
So the industry has now certified, publicly and at enormous expense, that the modeled layer is mandatory — and then hands you an empty one and a project plan.
- That project is the reason you don't have this yet. Not budget. Not conviction. The six months and the people.
- Synopsis derives the layer instead of asking you to write it. We learn your metric definitions from the documents your company already trusts, then reconcile until our numbers match the ones you report.
- Same conclusion, no project. You get the thing all four of them say you need, without the year of engineering they each assume you have.
Every serious platform requires a modeled layer. The only question is who builds it.
Worth knowing
Two market moves from the last eight weeks.
Fivetran and dbt Labs are now one company
The merger closed 1 June 2026, explicitly positioned as building the data infrastructure for trusted AI agents. When the ingestion and transformation layers consolidate and point themselves at agents, the category thesis is no longer speculative.
Progress Software is acquiring Domo's data platform
Announced 22 July 2026: $400M in cash for substantially all of Domo's AI and data platform assets, serving 2,400+ businesses, expected to close by 30 November 2026. Progress frames it as "context and control for AI."
The AI vendors are partners, not opponents
Anthropic ships first-party Claude connectors for Snowflake, Databricks, and BigQuery, and appeared on stage at Snowflake Summit. The debate was never model-versus-warehouse. It was always about what sits on the other end of the connection.
How to read the numbers on this page
We would rather you trust the argument than the scoreboard, so here is what's wrong with the scoreboard:
- Every accuracy figure here is vendor-run. Snowflake's ADE-Bench result has no published independent replication.
- The test conditions differ. Snowflake's agent had native access to its own controls while the comparison agent used an external connector — an analyst reviewing it called that "a different test condition, not merely a capability gap."
- Others score higher. dbt's live ADE-Bench leaderboard has shown a third-party agent above Snowflake's own result.
- Lab beats production. Research on enterprise agentic systems suggests roughly a 37% gap between benchmark and production performance — which applies to everyone, including us.
The reason to believe the thesis isn't any single benchmark. It's that four companies with opposing interests, looking at their own customers' failures, all shipped the same answer.
The industry agrees. The project is still the problem.
Bring the question your team has never been able to get answered. Twenty minutes, on your data, and you'll hear it out loud.
