AI agents may attract the attention, but governed business context will determine whether they can be trusted.
Earlier this summer, I joined Wesley Nitikromo on the Allocating Intelligence podcast to discuss Snowflake Summit 2026 and the announcements that I believe matter most beyond the keynote stage.
The conversation was not really about individual product features. It was about a more fundamental question:
Where should intelligence live in a modern data architecture, and where should the necessary context be managed?
Snowflake’s announcements around Horizon Context, Semantic Views and Cortex Sense offer an interesting answer. They also expose several architectural questions that organizations should resolve before putting AI agents into production.
From data platform to context platform
I attended my first Snowflake Summit in 2019. At the time, Snowflake’s proposition was relatively straightforward: separate storage and compute, use the elasticity of the cloud and make working with data simpler.
That foundation is still there, but the ambition has expanded considerably.
Snowflake is no longer positioning itself as merely the place where organizations store and process data. It increasingly wants to become the platform where data, business meaning, governance, applications and AI come together.
That shift makes context essential.
An AI agent can technically generate a query without understanding your organization. But it cannot reliably answer a business question unless it understands concepts such as revenue, active customers, product groups, reporting periods and organizational ownership.
Those definitions are rarely universal. They are specific to the organization and frequently disputed inside it.”
Without governed business context, a more capable model simply produces uncertain answers more convincingly.
Horizon Context and Cortex Sense solve different problems
One important part of the Summit story was the relationship between Horizon Context and Cortex Sense.
They should not be treated as interchangeable names for the same capability.
Horizon Context belongs to Snowflake’s governance and catalog direction. It is concerned with collecting and managing trusted business context: definitions, relationships, metadata, lineage and the controls surrounding them.
Snowflake’s native Semantic Views are an important part of this foundation. They allow business entities, dimensions, metrics and relationships to be represented as governed schema-level objects instead of definitions hidden inside individual reports or disconnected YAML files.
Cortex Sense is closer to the consumption and inference side. It helps AI experiences use relevant business context at runtime so that natural-language questions can be interpreted against the organization’s actual data and definitions.
The distinction matters.
Governance needs consistency, ownership, traceability and control. AI inference needs relevant context to be retrieved and applied quickly enough to answer a question. These requirements are related, but they are not identical.
A trustworthy AI architecture needs both.
A semantic layer is not a substitute for governance
It is tempting to think that introducing a semantic layer will resolve inconsistent definitions across an organization.
Technology can help, but it cannot decide which definition is correct.
If Finance and Sales use different definitions of revenue, putting both definitions into a semantic platform does not solve the disagreement. The organization still needs ownership, decision-making and a process for approving changes.
The technology becomes valuable after those responsibilities are clear. It can encode the agreed definition, make it discoverable, apply it consistently and show where it is being used.
That is why semantic governance is as much an operating-model challenge as a modeling challenge.
Before building an AI agent, organizations should be able to answer questions such as:
- Who owns each critical business metric?
- Which definition is authoritative?
- How are changes reviewed and communicated?
- Can an answer be traced back to its data and business logic?
- Do access policies remain effective when an AI agent executes the query?
If those questions cannot be answered, the organization is not yet facing an AI problem. It is facing a data foundation problem.
The multi-tool reality
Snowflake’s approach is especially compelling when Snowflake is the primary execution and governance environment.
Most production landscapes, however, are not Snowflake-only.
Business logic may also live in dbt models, Tableau calculations, Power BI measures, LookML, spreadsheets, notebooks and application code. Some data may never pass through Snowflake at all.
This creates a difficult architectural boundary.
A definition governed inside Snowflake does not automatically correct a conflicting definition maintained in a spreadsheet. Query-time controls inside one platform cannot govern a calculation that bypasses that platform entirely.
Open standards and integrations can reduce duplication, but they do not eliminate the need for coordination. Teams must still determine where definitions are mastered, how they are synchronized and what happens when two tools disagree.
The semantic layer should therefore be treated as part of a broader architecture, not as a box that can be installed to make governance complete.
Portability is a design principle
During the podcast, we also discussed vendor lock-in and exit strategy.
For me, an exit strategy does not mean expecting or planning to leave Snowflake. It means avoiding architectural decisions that make leaving practically impossible.
Business definitions are valuable organizational assets. If critical metrics, relationships and governance rules can only be understood inside one proprietary implementation, the organization has created a dependency that extends far beyond technology.
A responsible architecture should make those definitions documented, reproducible and, as far as practical, portable.
This creates optionality. It also improves the current implementation because definitions that must be explicit and transferable are usually better governed than definitions hidden in dashboards, code or institutional memory.
Portability is not an argument against choosing a platform. It is part of choosing and implementing that platform well.
Are organizations ready for AI agents?
The models and agent frameworks are moving quickly. Most organizations’ data foundations are not.
The main obstacle to deploying trustworthy agents is often not model capability. It is the absence of reliable context:
- Metrics are defined differently across departments.
- Lineage is incomplete.
- Documentation no longer reflects production.
- Access policies were designed for human users, not autonomous agents.
- Business rules exist inside the minds of a few experienced employees.
Giving an agent access to this environment does not remove the inconsistencies. It allows the agent to encounter them at greater speed and scale.
The sensible sequence is therefore:
- Establish ownership of important business concepts.
- Create and maintain trusted semantic definitions.
- Connect those definitions to lineage, security and governance.
- Test whether answers are accurate, explainable and reproducible.
- Only then increase the agent’s autonomy.
The agent is not the foundation. Context is.
What happens to business intelligence?
AI agents will change how people interact with business data, but this does not necessarily mean that BI platforms disappear.
Dashboards and reports provide more than a query interface. They offer a governed and repeatable presentation of information. They create a shared view that people can discuss, review and challenge.
Natural-language interfaces introduce a different interaction model. Instead of navigating a predefined dashboard, a user can ask a specific question and receive a generated answer.
That answer still needs to be verified.
BI may increasingly become the place where people validate, monitor and explain what an agent has returned. The role changes from being the only interface to becoming part of the trust and verification layer.
The BI tools facing the greatest pressure are those in which the business logic is trapped inside individual dashboards. AI makes that architectural debt more visible.
The announcement that matters after the applause
Agents made for the most visible Summit demonstrations. But the less spectacular work around semantic definitions, governance and context may prove more important.
Success will not be measured by how many agents an organization deploys. It will be measured by whether those agents can work with definitions people trust, produce answers that can be explained and operate within controls the organization can enforce.
That starts with the data foundation.
It continues with the semantic layer.
Only then does intelligence become something an organization can allocate responsibly.
Listen to the conversation
In the full episode, Wesley Nitikromo and I discuss:
- Snowflake’s evolution from cloud data warehouse to data and AI platform
- Horizon Context, Semantic Views and Cortex Sense
- Query-time governance in multi-tool environments
- Open Semantic Interchange and semantic-layer portability
- Vendor lock-in and exit strategy
- Data sovereignty in the Benelux
- Whether organizations are ready to run AI agents
- How the role of BI may change
Listen to or read Daan Bakboord: Reading Snowflake Summit Through the Semantic Layer on Allocating Intelligence.
You can also find the podcast post on LinkedIn.
If you are evaluating Snowflake, designing a semantic layer or preparing your data platform for AI agents, tell me what you are working on. We will start with the context.
Transparency note: This article was generated with the assistance of AI, based on my podcast conversation with Wesley Nitikromo. I reviewed and edited the content to ensure it accurately reflects my views and experience.
