Daan Bakboord: Reading Snowflake Summit Through the Semantic Layer

Snowflake Horizon Context is the announcement that did not get the standing ovation. Cortex Sense did not either. The agents did — the demos were polished, the vision was clear, and the keynote earned its applause. But Daan Bakboord has been to five Summits. He knows which layer the important decisions live on. And this year, from Amsterdam, he watched Snowflake split the semantic layer in two and tried to work out what that actually means for the data teams now being asked to build on top of it.

Daan Bakboord — Snowflake Data Superhero since 2018, Chapter Lead of the Dutch Snowflake User Group, independent consultant at DaAnalytics, Consulting Partner at Bravinci — joins Allocating Intelligence to read this Summit out loud. Not the keynote version. The version that matters in production.

Five Summits: The Arc From Warehouse-Only to the Agentic Enterprise

Daan's first Summit was San Francisco, 2019. Snowflake was, in his framing, "just a database." The founders still had a whiteboard story. The pitch was elastic compute and separation of storage and compute — genuinely novel at the time, but modest in scope compared to what the platform has become.

Five years later, Snowflake is positioning itself as the substrate for the agentic enterprise. Every keynote moment — Cortex Agents, the Polaris Catalog integrations, the AI-native application framework — is built on the claim that you should run intelligence where your data already lives, rather than moving the data to wherever your intelligence is being built.

That framing is correct. And it is also the most important thing to stress-test. Daan's value in this conversation is that he has watched every iteration of Snowflake's positioning since 2019 and has the consulting track record to know which announcements turned into durable architecture and which ones became footnotes. That perspective shapes how he reads this year's announcements — not skeptically, but carefully.

Horizon Context and Cortex Sense: Why Snowflake Split the Semantic Layer in Two

The most architecturally significant announcement at this year's Summit — the one that deserves more analysis than it got in the coverage — is the split of Snowflake's semantic layer ambition into two distinct products: Horizon Context and Cortex Sense.

Horizon Context sits in the governance engine. It is the layer where business definitions, metric relationships, and data context get encoded and enforced. The promise is enforcement at query time — not just documentation that lives in a wiki and gets ignored, but semantic rules that are applied as queries execute against the warehouse.

Cortex Sense sits closer to the AI layer. It is the interface through which Snowflake's Cortex models — and, via MCP, external AI agents — access the semantic context that Horizon Context defines. The idea is that an AI working inside the Snowflake ecosystem should be able to reason about your data using your business definitions without anyone having to re-explain those definitions in the prompt.

Daan works through why this split happened and what it means in practice. The honest answer is that it reflects a real architectural tension: governance-layer semantics and inference-layer semantics have different requirements, different latency tolerances, and different failure modes. Merging them into a single product was always going to force uncomfortable tradeoffs. Splitting them is the right call — but it introduces its own complexity for teams trying to maintain alignment between the two layers.

Enforcement at Query Time: Real or Keynote Theater

The enforcement question is where Daan is most precise, and most useful. Semantic definitions that exist only as metadata — definitions that a BI tool reads and a human might override — have limited value as a governance mechanism. The claim that Horizon Context enforces semantics at query time is a stronger claim. It means that the definition of "revenue" is not advisory. It is applied.

Whether that holds up under the conditions actual enterprise data teams operate in — mixed-tool environments, legacy pipelines still writing to the same tables, external BI tools that bypass Snowflake's query path entirely — is a different question. Daan does not dismiss the announcement. He probes it. The answer is that enforcement at query time works reliably within the Snowflake ecosystem and becomes more complicated as you move further from it.

That boundary matters enormously for implementation decisions. Teams building entirely within Snowflake get the full benefit. Teams with a multi-cloud architecture, or with significant tooling outside the Snowflake perimeter, need to think carefully about which parts of their semantic governance they can actually enforce and which parts they are still relying on human discipline to maintain.

The Lock-In Question and Exit Strategy as a Design Principle

The lock-in conversation is one of the most practically useful exchanges in this episode. It is not framed as a warning against Snowflake — Daan is genuinely bullish on the platform and has been since 2018. It is framed as a design principle that should be present in every implementation regardless of which platform you are building on.

The core argument: if you build your semantic layer inside Snowflake — your metric definitions, your business logic, your governance rules — and then something changes (pricing, competitive alternatives, regulatory requirements, an acquisition), how long does it take you to move? If the answer is "we would have to rebuild from scratch," that is not a lock-in complaint. It is a design failure.

Exit strategy as a design principle means making architectural choices that preserve optionality. Not because you plan to leave, but because the ability to leave is itself a form of negotiating leverage and organizational resilience. Daan applies this to the semantic layer specifically: encoding business logic in a format that is portable — or at minimum, documented and reproducible — is not a hedge against Snowflake. It is good data architecture practice.

AtScale, Honeydew, and the Snowflake Ventures Question

The Open Semantic Interchange announcement and Snowflake's Ventures investments in AtScale and Honeydew sit in an interesting tension. Both AtScale and Honeydew offer semantic layer capabilities that compete, in some configurations, with what Horizon Context is trying to do natively. Snowflake investing in both while building its own semantic layer infrastructure raises a question that Daan addresses directly: what does "open" actually mean when the platform defining the standard also has a stake in the companies building on it?

His read is that the OSI is real as a technical effort and limited as a competitive neutralizer. The interchange format solves the portability problem at the definition layer. It does not solve the enforcement problem, the query-time integration problem, or the organizational problem of getting teams to maintain semantic definitions in a shared format rather than duplicating them in every tool. Those problems require more than a standard. They require organizational commitment — and that is not something Snowflake or the OSI can mandate.

Data Sovereignty in the Benelux and What Bravinci's Dividos Actually Solves

The data sovereignty conversation is the most geographically specific part of the episode and one of the most practically relevant for European data teams. EU regulation — GDPR, the Data Act, and the emerging AI Act requirements — creates constraints on where data can live and how it can be processed that do not exist in the same form for US-headquartered organizations.

Daan works in the Benelux market. His clients include organizations in regulated sectors where the question of whether data can legally be processed on US-based cloud infrastructure is not theoretical. It is a compliance requirement that shapes architecture decisions before any technology evaluation begins.

Bravinci's Dividos is positioned as a best-of-breed sovereign stack for organizations operating under these constraints — a way to run modern data infrastructure without the data sovereignty compromises that come with defaulting to the major US hyperscalers. Daan explains what the stack looks like in practice and where the tradeoffs are. The short version: sovereign compliance is achievable, and it does not require giving up modern tooling, but it does require architectural choices that prioritize portability and jurisdiction-awareness from the start.

Is the CDO Ready to Run Agents on Monday

The answer is no. Daan is clear about this, and clear about why — and the reason is not model capability. It is context.

The blockers for running AI agents reliably in enterprise data environments are not resolved by better foundation models or more capable agent frameworks. They are resolved by having clean, governed, trustworthy context for the agent to operate on. That means defined metrics. Enforced business logic. Data lineage that is accurate, not aspirational. Governance that is encoded in the architecture, not maintained by human discipline.

Most CDOs do not have that foundation. They have data. They often have a lot of it. They may have a semantic layer in some form. But the gap between "we have a semantic layer" and "we have a semantic layer that an AI agent can trust enough to act on without a human checking every output" is the gap that decides whether agentic data actually works in production. Closing that gap is the work. The agents are already here. The context is not.

Whether BI Tools Survive When Everyone Asks the Agent

The BI survival question gets a nuanced answer. The framing that BI tools simply disappear — replaced wholesale by natural language interfaces to AI agents — misunderstands how enterprise data consumption actually works. Most of the value that BI tools deliver is not in the query interface. It is in the governed, structured presentation of business logic to people who need to make decisions with it.

That function does not go away when agents arrive. What changes is who — or what — is doing the querying. If the agent is well-governed, if it is drawing from a trustworthy semantic layer, if its outputs can be audited and explained, then the BI layer may shift from "where humans go to query" to "where humans go to verify what the agent returned." That is a different role. It is not no role.

The BI tools that do not survive are the ones that have been substituting for a semantic layer rather than building on top of one. If your business logic lives in the BI tool — in Looker LookML, in a Tableau calculated field, in a Power BI measure that three people maintain — then the agent era does not just threaten your BI tool. It surfaces the architectural debt you have been carrying.

What Success Looks Like in Three to Five Years

The closing frame of the conversation is the one that makes all the preceding detail useful. Daan's version of success in three to five years is not "we ran the most agents" or "we adopted the most Snowflake features." It is: we fixed the foundation first, and now the intelligence we are allocating on top of it is trustworthy.

That means a semantic layer that is maintained, not abandoned. Governance that is enforced, not aspirational. Context that an agent can act on without a human in the loop for every decision. And architecture that is portable enough that the decisions made today do not foreclose options that do not exist yet.

That is a less exciting pitch than the agentic keynote. It is the one that produces results.

Episode Chapters

• 0:00 — Cold open
• 0:49 — Conference season and what excites Daan most this year
• 2:09 — Who is Daan Bakboord: from Oracle to Snowflake Data Superhero
• 3:24 — Five Summits, and why this year was optional
• 4:38 — 2019, warehouse-only, and the founders' whiteboard
• 5:32 — Why everything is suddenly about context and agentic
• 7:10 — When Snowflake first talked semantics: Semantic Fuse and OSI
• 8:43 — Why split into Horizon Context and Cortex Sense
• 9:09 — Semantics in the governance engine: real or theater
• 10:47 — What actually changed about semantic views
• 12:42 — Semantics in the warehouse, and the lock-in question
• 16:52 — AtScale, Honeydew, Snowflake Ventures, and competing semantic layers
• 19:09 — Honest consulting: helping clients choose, not selling complexity
• 20:46 — Is the Open Semantic Interchange real or for show
• 21:00 — Exit strategy as a design principle
• 23:02 — Data sovereignty in the Benelux and the EU regulation reality
• 24:24 — Bravinci's Dividos: a best-of-breed sovereign stack
• 27:32 — The six-to-one ratio, agentic spend, FinOps, and carbon
• 28:12 — Is a CDO ready to run agents on Monday
• 33:46 — Do BI tools survive the agent era
• 34:00 — Which announcements still matter in a year
• 35:47 — Rapid fire
• 40:26 — Trusting the private preview to GA process
• 41:08 — What success looks like in three to five years

About Daan Bakboord

Daan Bakboord is a Snowflake Data Superhero since 2018 and Chapter Lead of the Dutch Snowflake User Group. He works as an independent consultant at DaAnalytics and as a Consulting Partner at Bravinci, where he advises organizations across the Benelux on data architecture, semantic layer strategy, and AI-ready data infrastructure. He has attended five Snowflake Summits in person and brings a practitioner's eye to platform announcements — focused on what holds up in production, not what earns applause in a keynote.

Connect with Daan on LinkedIn: linkedin.com/in/daanbakboord

About Allocating Intelligence

Allocating Intelligence is a podcast about where intelligence should live in the modern data and AI stack. Hosted by Wesley Nitikromo — founder of Unwind Data, formerly co-founder of DataBright (acquired). Each episode explores one layer of the Intelligence Allocation Stack: data foundation, semantic layer, orchestration, and AI agents.

Subscribe for new episodes and conversations with the people building the next generation of enterprise data and AI infrastructure.

Links: Allocating Intelligence · Unwind Data · Wesley on LinkedIn · Listen on Spotify