Agents Have Great Expectations. Here’s How to Prepare the Estate.

TL;DR

Most Snowflake estates were architected to answer questions people already knew to ask. But agents, turns out, have great expectations. And, unlike Pip, they will not wait patiently for someone to prepare the estate. Intelligence infrastructure requires four things a reporting foundation was never designed to provide: data that reaches the people and systems who need it at operational speed, compute that scales without mystery spend, a semantic layer that translates business context into governed queries, and access controls that reflect how decisions are made. Organizations with all four are building a compounding advantage. The distance between where most estates are and where they need to be is smaller than it looks. This post maps it, gap by gap.

What this post covers:

• Why reporting foundations and intelligence foundations are structurally different builds
• The four gaps and what each one costs operationally
• What a production deployment of an intelligence foundation looks like
• How to score your own estate in 15 minutes

Why Most Snowflake Estates Were Built for a Different Job

Snowflake estates were built carefully. Years of engineering work went into getting data ingested, modeled, governed, and surfaced through reports and dashboards that keep the business informed. That foundation is real, it works, and for the job it was designed to do, it’s excellent.

But the job description has changed. AI agents operating across supply chains, revenue workflows, and operational systems need a foundation that can do something different: return governed, semantically grounded answers to questions nobody has thought to ask yet, at the speed of execution, with access controls that reflect organizational reality rather than IT architecture.

The gap between what a reporting estate provides and what an intelligence estate requires is not a Snowflake problem. It’s an architectural one, and most organizations are running into it at exactly the moment they are being asked to move fastest. MIT's NANDA initiative found in 2025 that 95% of organizations deploying generative AI saw zero measurable return. Not low return. Zero. The failure is almost never the model. It is data readiness, workflow integration, and the absence of the foundational architecture the model needs to operate inside (MIT, The GenAI Divide: State of AI in Business, 2025).

The four gaps below are where that foundation breaks down. Every one of them is addressable without replacing the underlying Snowflake investment. The sequencing matters more than the speed.

Gap 1: Snowflake Data Accessibility for Agentic Workloads

The reporting foundation delivers insight to people who know what to look for. An intelligence foundation delivers insight to systems running decisions in real time. Those are different distribution problems, and they require different architectural answers.

In a reporting architecture, a data engineer builds a model, a dashboard surfaces it, and a business user pulls a report when they need it. Governed, functional, and measured in days. Agents need answers in milliseconds, structured for direct consumption, without a human intermediary assembling context from three different tools. When data accessibility is architected for human query patterns rather than machine consumption, agents slow down, work around the gaps, or return answers that lack the context to be trusted.

The operational cost is compounding. Teams spend engineering cycles building workarounds for access paths that were never designed for agent consumption. Decisions that should execute in systems execute in meetings instead. The data exists, but the path to the data isn't designed for the speed or the volume that agents require.

What fixed looks like:

• Data models structured so agents can query directly without a human assembling context first, which means defining the access paths, the filters, and the business rules the agent needs before deployment, not after it stalls
• Pipelines running at operational cadence rather than dashboard cadence, the difference between data that is current when a decision executes and data that was current when someone last opened a report
• Access paths that treat agent queries as first-class requests with their own identity, permissions, and audit trail, not as a subset of human query behavior that happens to run faster

Gap 2: Snowflake Cost Efficiency at Agentic Scale

Reporting workloads are predictable. A dashboard refreshes on a schedule. A query runs when someone opens a report. Warehouse costs scale in ways that a reasonable person can anticipate and budget.

Agentic workloads are different. Agents fire queries continuously, often in parallel, and often at unpredictable volume. An estate tuned for reporting costs will encounter compute bills that require explanation to a CFO who approved a very different number. Query patterns that made sense for dashboards become expensive at agent scale. Idle warehouses that were acceptable during reporting cycles become a cost line that compounds while nobody’s looking.

The broader context makes this more urgent. Enterprises now spend an average of $29.3 million annually on data programs—one of the largest line items in enterprise technology budgets—with data teams dedicating 53% of their engineering time to maintenance rather than forward work (Fivetran, Enterprise Data Infrastructure Benchmark Report 2026). Layering agentic workloads on top of an estate that is already maintenance-heavy without addressing the underlying cost structure is how organizations end up in conversations they did not anticipate.

What fixed looks like:

• Query optimization reviewed on a regular cycle with warehouse sizing matched to actual workload patterns rather than inherited from the reporting era
• Idle compute monitored systematically, with warehouses scheduled to match actual demand rather than running continuously against a billing clock
• Cost visibility instrumented well enough that an engineer can connect a spend spike to a specific workload or agent behavior within hours, because at agentic query volumes, the window to catch a cost problem before it becomes a budget conversation is short

Gap 3: AI and Agentic Readiness - The Semantic Layer Problem

This is the gap Sparq’s Jackson Stakeman, GM of The Shop, named from the Snowflake Summit floor: “data, data everywhere, and not a drop to drink.” The data exists in Snowflake. The semantic layer that would allow an agent to understand what "churn risk" or "approved vendor" or "revenue-critical account" means in business terms does not.

Without a semantic layer, agents return technically correct answers to technically formed queries. They cannot interpret a question phrased as a business user would phrase it. They cannot distinguish between a metric that means one thing in sales and something slightly different in finance. They cannot carry context across a multi-step workflow without that context being explicitly engineered into every step. McKinsey's 2026 AI Trust Maturity Survey of approximately 500 organizations globally found that as organizations move from AI experimentation to autonomous systems, the need for trusted business context, not just model capability, is the defining constraint on whether those systems can be trusted in production (McKinsey, State of AI Trust in 2026: Shifting to the Agentic Era).

This is the gap that Snowflake's Horizon Context announcement directly addresses. It’s also the gap that organizations that have already invested in a governed semantic layer have closed, and the compounding advantage from having done so is about to become significantly more visible.

A global identity and access management leader encountered this gap with hundreds of governed reports and models in Snowflake and insight delivery that remained entirely pull-based. Sales and customer success teams navigated multiple tools to assemble signals, slowing execution and introducing inconsistency into revenue-critical decisions. Sparq deployed Ask.IQ to translate natural-language questions into governed Snowflake queries, aligned business terminology with the underlying data through a semantic layer, and captured every interaction to establish a continuous feedback loop. The result: instant natural-language access to governed data, 40-plus use cases identified across the business, and 100% of user questions captured in a feedback loop that surfaced what the organization needed to know. The semantic layer was the unlock.

What fixed looks like:

• Business terminology mapped to underlying data models so that "revenue" means the same thing to every agent, every user, and every downstream system, not slightly different things depending on which team built the query
• A consistent semantic layer that agents and human users share, so that the governed understanding of the business is a single source of truth rather than a per-team interpretation
• A feedback loop that captures what questions are being asked and what answers are being acted on, so the foundation gets smarter over time, rather than requiring manual maintenance every time a business definition changes

Gap 4: Snowflake Governance and RBAC for AI Agents

The number one concern at the Snowflake Summit booth was governance and access control. Not because practitioners have suddenly developed an interest in compliance as an intellectual exercise, but because agents change what governance needs to do.

In a reporting architecture, governance means controlling who can see which reports. Role-based access control is mapped to job titles, managed in IT, and reviewed annually (if everyone is disciplined). It works because humans move slowly enough that manual oversight can keep pace.

Agents move fast. They query continuously, operate across systems, and make decisions that touch sensitive data at a volume no human review process can track in real time. Gartner predicted in February 2024 that 80% of data and analytics governance initiatives will fail by 2027, and that prediction was made before agentic workloads became an operational reality for most enterprises (Gartner, Predicts 2024: Data and Analytics Governance Requires a Reset). The governance initiatives that were already fragile are now being stress-tested by a class of query behavior they were never designed to handle.

When Snowflake announced Agent Identity tracking at Summit, the specific problem they were solving is the one practitioners have been naming for months: without granular, model-level RBAC that distinguishes agent actions from human actions, governance breaks down at scale. Access controls structured for human query patterns become a liability rather than a safeguard when agents start running. The operational cost is exposure: decisions made on data that should not have been accessible, audit trails that cannot be reconstructed, and trust in the system that erodes with every unexplainable output.

What fixed looks like:

• Role-based access controls that reflect actual job function and data sensitivity, applied at the role level rather than the domain level, because a wall around an entire schema protects the sensitive data and blocks the legitimate access that agents and analytics both require
• Lineage traceable from source to decision automatically, so that any AI-generated output can be reconstructed and explained without manual investigation
• Governance architecture that treats agent identity as a first-class concept, with agents operating under their own defined permissions, distinct from the human users whose access patterns governance was originally built around

How to Build a Snowflake Intelligence Foundation

The four gaps are addressable. None of them require replacing the Snowflake investment. They require sequenced architectural work on top of a foundation that, in most cases, is stronger than it appears from the outside. The organizations doing this work now are building an advantage that compounds every quarter, because every gap closed makes the next AI initiative faster, safer, and less expensive to deploy.

The sequencing matters. Data accessibility and governance are the right starting points because they unblock everything that follows. Cost efficiency work compounds once the query patterns are understood. The semantic layer, which is where the intelligence actually lives, is the final layer, and it is significantly faster to build when the governance and accessibility architecture underneath it is sound.

The best starting point is knowing where the estate stands today. Most data teams have a strong intuition about which gaps are largest. Fewer have a systematic read across all four.

Get the Snowflake Value Scorecard

15 minutes, 12 statements, a scored output that shows exactly where to focus first. Get it here →

Rewired is a publication from Sparq. Each edition examines what happens when AI enters production inside the performance engine, the operational systems where margin, throughput, and decision speed are effectively determined. Straight pattern analysis, economic stakes attached, written for operators accountable for outcomes. Subscribe to Rewired.

Frequently Asked Questions

What is the difference between a reporting foundation and an intelligence foundation in Snowflake?

A reporting foundation is architected to answer questions people already know to ask. Data is modeled for human query patterns, pipelines run on reporting cadence, and access controls are designed for dashboard consumers. An intelligence foundation is architected to serve systems operating in real time. It requires data accessible at machine speed, a semantic layer that translates business context into governed queries, access controls that reflect how decisions are made, and compute tuned for continuous agentic workloads rather than scheduled reporting cycles. The platform is the same. The architectural requirements are different.

What is a Snowflake semantic layer and why does it matter for AI agents?

A semantic layer is the translation layer between how a business talks about its data and how that data is physically structured in Snowflake. It maps business terminology — "churn risk," "approved vendor," "revenue-critical account" — to underlying data models, so that a query phrased in business language returns an answer that reflects business meaning rather than raw database logic. Without a semantic layer, agents can query Snowflake but cannot interpret questions the way a business user would phrase them. With one, agents operate from the same governed, consistent understanding of the business that human users rely on.

What is the difference between Snowflake Cortex and a semantic layer?

Snowflake Cortex provides the AI capability — the large language models and inference infrastructure that power natural-language queries and agentic reasoning inside the Snowflake platform. A semantic layer provides the business context that makes Cortex useful: the mapping between how your organization talks about its data and how that data is physically structured. Cortex can run without a semantic layer, but the outputs it produces will reflect database logic rather than business meaning. Together, they are what make natural-language access to governed data both possible and trustworthy. Horizon Context, announced at Snowflake Summit 2026, is Snowflake's native approach to centralizing that semantic layer across every person, tool, and agent operating in the platform.

What is Snowflake Agent Identity and why does it matter?

Agent Identity is the governance mechanism Snowflake announced at Summit 2026 that allows the platform to distinguish between queries made by autonomous AI agents and queries made by human users. Without it, governance systems treat all queries as equivalent — which means an agent operating at high volume across sensitive data inherits the same access controls and audit trail as a human analyst pulling a weekly report. With Agent Identity, organizations can define granular, model-level RBAC specifically for agents: what data they can access, what actions they can take, and how every interaction is logged. It is the specific infrastructure that makes governance architecture compatible with agentic workloads rather than a liability under them.

How do agentic workloads affect Snowflake costs?

Agentic workloads are structurally different from reporting workloads in ways that make reporting-era cost assumptions unreliable. Reporting workloads are scheduled and predictable — a dashboard refreshes, a query runs, a warehouse scales up and back down. Agents query continuously, often in parallel, at volumes that compound quickly. Warehouses that were sized for dashboard refresh cycles will over-provision at agentic query rates. Token costs for Cortex features that were acceptable at reporting volume can become a significant budget line at agent scale. The organizations managing this well are monitoring warehouse sizing and query patterns on a regular cycle, reviewing Cortex token costs against the value of each use case, and instrumenting cost visibility before agentic workloads go live — not after the first billing cycle requires explanation.

How does governance need to change when AI agents are querying Snowflake?

Traditional RBAC is designed for human users: it controls who can see which data based on role and job function, managed manually, reviewed periodically. Agents query continuously, at high volume, across sensitive data, faster than manual oversight can track. Effective governance for agentic workloads requires access controls that distinguish between human and agent actions at a granular level, lineage that is traceable from source to decision automatically, and audit artifacts that can reconstruct any AI-generated output without manual reconstruction. Snowflake's Agent Identity tracking addresses this directly. Organizations that invest in governance architecture before agents arrive are significantly better positioned than those retrofitting it after a production deployment has already introduced exposure.

How long does it take to move from a reporting foundation to an intelligence foundation?

The timeline depends on the current state of the estate, but the distance is typically shorter than organizations expect. The semantic layer, governance structure, and data accessibility architecture are all buildable on top of an existing Snowflake investment without replacing the underlying platform. The sequencing matters: data accessibility and governance are usually the right starting points because they unblock everything that follows. Cost efficiency work compounds from there. Agentic readiness — with the semantic layer as its foundation — is the final layer. A production-ready intelligence foundation, deployed on an existing Snowflake estate, is achievable in weeks with the right architectural approach. The Snowflake Value Scorecard is designed to surface where the largest gaps are so sequencing decisions can be made from evidence rather than instinct.