By Ben Wilson, [TITLE: Principal Consultant at phData?]
Twenty-nine percent of Snowflake's public study guide for the SnowPro Specialty: Gen AI certification is governance. Only the AI functions themselves get more weight. When I started studying, that ratio looked odd. After a summer of running my own agents, it looks right.
I passed in late September. [BEN: share score? 854/1000.] This post is about what the studying taught me that applies to agents in production on Snowflake, and what changed in the product while I was studying. It says nothing about the exam's content; that is covered by the candidate agreement, and it is also beside the point.
The public map
Snowflake publishes four domains and their weights for this certification (exam code GES-C02):
- Snowflake for Gen AI overview: 18%
- Snowflake Gen AI functions: 38%
- Snowflake Gen AI governance: 29%
- Snowflake document processing: 15%
If you lead a data team, read that list as a budget. The guide puts 29% of its weight on governance: who can call which model, where the data goes, and what it costs.
The building blocks that matter for agents
Snowflake's AI product family is called Cortex. Four parts of it did most of the work in my notes.
AI functions in Structured Query Language (SQL). Functions such as AI_CLASSIFY, AI_EXTRACT, AI_FILTER, AI_REDACT and AI_PARSE_DOCUMENT run a model inside a query. For agents, this is the batch layer: classify ten thousand prior-authorization notes or redact personal information before anything else reads them, under the same access rules as the table. (Hypothetical example.)
Cortex Search. A managed retrieval service. Per the docs, it combines vector search, keyword search and semantic reranking. If you define a primary key, refreshes process only the rows that changed. This is the agent's path into unstructured text: policies, clinical guidelines, contracts.
Semantic views and Cortex Analyst. A semantic view stores logical tables, facts, dimensions, metrics and the relationships between them. Cortex Analyst turns a question into SQL against that view. This is where your metric definitions live, which was the whole argument of my first post.
Cortex Agents. Agents plan the work, call tools, and then reflect on the results to decide the next step. Their built-in tools include Cortex Analyst for structured data and Cortex Search for unstructured data. They can also call custom tools built from stored procedures and user-defined functions, and remote Model Context Protocol (MCP) servers.
The lesson I took from putting those four side by side: the agent is the thin part. The search index, the semantic view and the access rules are the thick part, and they are what your data team already knows how to build.
Governance: the 29%
Three facts from the governance material change what I'd check first in any account.
Access is open by default. Calling AI functions needs the USE AI FUNCTIONS account privilege and one of the CORTEX_USER or AI_FUNCTIONS_USER database roles. By default, both USE AI FUNCTIONS and CORTEX_USER are granted to the PUBLIC role. Unless someone revoked them, every user in your account can call models today.
Data can cross regions. The CORTEX_ENABLED_CROSS_REGION parameter lets Snowflake process a request in another region when your region doesn't host the model. For organizations created after March 9, 2026, the default is ANY_REGION. Other commercial accounts default to a same-cloud geography such as AWS_US or AZURE_EU, which is still wider than your home region. Snowflake's docs say your data stays stored in your home region and the inference payload passes through the processing region only while it runs. For a health system, that is a question for your privacy officer, not a default to inherit.
Cost has several meters. AI functions bill by the tokens they process. Cortex Search has separate meters, including warehouse compute for refreshes, embedding tokens, serving compute and storage. Usage views in ACCOUNT_USAGE break this out per service. Pick the views before the first invoice, not after.
What changed since I studied
The product moved faster than any study guide can. Items I'd now add, each checked against current docs:
- Model access moved to role-based access control (RBAC). The old account-level allowlist,
CORTEX_MODELS_ALLOWLIST, is being deprecated; starting August 2026 you can no longer change it to a new value. Models now appear as objects inSNOWFLAKE.MODELSwith application roles you grant like any other privilege. - Agents got their own identity. The
SERVICE_AGENTuser type went generally available in July 2026. When an agent acts for a user, Snowflake can mark the session "agent-active," and masking, row access and other policies can check for it withIS_AGENT_ACTIVATED. For a third-party agent, that marking happens only when it connects through an OAuth integration set withIS_AGENTIC = TRUE; a connection that uses only a token is not marked. TheQUERY_HISTORYviews gained anagent_typecolumn. - A gateway for models. Cortex AI Gateway entered preview on September 15, 2026: one endpoint for model traffic, a trace table of agent activity, per-user quotas and a usage view.
- Evaluation in SQL. Cortex AI Function Evaluation entered public preview on September 21, 2026. It measures output quality, cost and token usage against labeled datasets.
- Document parsing consolidated. The legacy
PARSE_DOCUMENTfunction will be deprecated by the end of 2026.AI_PARSE_DOCUMENTis the path forward. - A managed MCP server. Snowflake now hosts an MCP server that can expose Cortex Agents, Cortex Analyst, Cortex Search, SQL execution and custom tools to outside agents.
What studying didn't teach me
A certification proves you know where the controls are. It doesn't tell you which ones your organization has set.
[BEN: two or three sentences from your own building, not the exam. Candidates: what surprised you when you ran Cortex against your own synthetic or public data; what you'd have to see in a sandbox before trusting an agent with PHI; where the study material and the live product disagreed on features (not on exam items).]
The other gap is operations. Studying covers what each feature does. It doesn't cover what happens when an agent stalls at 2 a.m., or when a definition changes and nobody tells the agent. That is the part I'm writing about here.
Your turn
If you checked your account today, is CORTEX_USER still granted to PUBLIC, and what is your cross-region setting?
LinkedIn cut (~150 words, no link in body)
29% of Snowflake's public study guide for the SnowPro Gen AI certification is governance.
When I started studying, that looked high. After a summer running my own AI agents, it looks right.
Three checks I'd run on any Snowflake account before an agent goes live:
- Who can call models? By default, the PUBLIC role holds both the USE AI FUNCTIONS privilege and the CORTEX_USER role.
- Where does inference run? New organizations since March 2026 default to cross-region inference set to ANY_REGION; other accounts default to a same-cloud geography such as AWS_US. For patient data, that's a privacy-officer question.
- Can you tell an agent from a person? Snowflake can mark agent sessions, and your masking policies can check for it. Third-party agents are marked only through an OAuth integration flagged as agentic.
The study guide gave me the map. The product kept moving: model access moved to role-based grants, and a model gateway entered preview in September.
What's the first governance setting you'd check?
Ben to decide
- Score. Share 854/1000, or say "passed" only?
- Exam-content line. The intro says the post won't discuss exam content. Keep that sentence (signals integrity), or cut it as unnecessary?
- [BEN] slot in "What studying didn't teach me". This is the post's only first-person build evidence. Without it, the post reads as a doc summary.
- Title. Working title vs the two alternatives in frontmatter.
- "Access is open by default" framing. Accurate per docs, but some accounts revoke it at setup. Keep "unless someone revoked them," or soften further.
- Preview features. AI Gateway and Function Evaluation are previews. Keep, or hold for a later post.
- Prior-authorization example. Labeled hypothetical; keep or swap.
- Byline title confirmation.
Verified product claims (checked 2026-10-02 via WebFetch)
| # | Claim | Source |
|---|---|---|
| 1 | AI functions incl. AI_CLASSIFY, AI_EXTRACT, AI_FILTER, AI_REDACT, AI_PARSE_DOCUMENT exist | https://docs.snowflake.com/en/user-guide/snowflake-cortex/aisql |
| 2 | Calling them needs USE AI FUNCTIONS + CORTEX_USER or AI_FUNCTIONS_USER | same page |
| 3 | USE AI FUNCTIONS and CORTEX_USER granted to PUBLIC by default | https://docs.snowflake.com/en/user-guide/snowflake-cortex/aisql-privileges-and-access |
| 4 | CORTEX_MODELS_ALLOWLIST being deprecated; from Aug 2026 can't set a new value; RBAC via SNOWFLAKE.MODELS objects + application roles | same page; https://docs.snowflake.com/en/sql-reference/parameters#cortex-models-allowlist |
| 5 | CORTEX_ENABLED_CROSS_REGION; ANY_REGION default for orgs created after Mar 9, 2026; other commercial accounts default to same-cloud geography (AWS_US, AZURE_EU) per "Default routing"; data stored only in home region, payload transient | https://docs.snowflake.com/en/user-guide/snowflake-cortex/cross-region-inference |
| 6 | Cortex Search = vector + keyword + semantic reranking; primary key enables incremental refresh | https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-search/cortex-search-overview |
| 7 | Semantic views store logical tables, facts, dimensions, metrics, relationships; usable with Cortex Analyst | https://docs.snowflake.com/en/user-guide/views-semantic/overview |
| 8 | Cortex Agents plan, use tools, reflect; built-in tools include Analyst + Search; custom tools from stored procedures/UDFs; MCP connectors | https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents |
| 9 | SERVICE_AGENT GA Jul 23, 2026; agent-active sessions; IS_AGENT_ACTIVATED usable in masking/row access policies; third-party agents marked only via OAuth integration with IS_AGENTIC = TRUE (token-only not marked) | https://docs.snowflake.com/en/release-notes/2026/other/2026-07-23-service-agent-user-type ; https://docs.snowflake.com/en/user-guide/agent-identity |
| 10 | agent_type column added to QUERY_HISTORY views (Jul 27, 2026) | https://docs.snowflake.com/en/release-notes/2026/other/2026-07-27-query-history-agent-type-column |
| 11 | Cortex AI Gateway preview Sep 15, 2026: endpoint, trace table, AI_GATEWAY_USAGE_HISTORY, per-user quotas | https://docs.snowflake.com/en/release-notes/2026/other/2026-09-15-cortex-ai-gateway |
| 12 | Cortex AI Function Evaluation public preview Sep 21, 2026: quality, cost, token usage vs labeled datasets | https://docs.snowflake.com/en/release-notes/2026/other/2026-09-21-ai-function-evaluation-preview |
| 13 | PARSE_DOCUMENT deprecated by end of 2026; AI_PARSE_DOCUMENT canonical | https://docs.snowflake.com/en/sql-reference/functions/parse_document-snowflake-cortex |
| 16 | AI functions incur compute cost based on tokens processed | https://docs.snowflake.com/en/user-guide/snowflake-cortex/aisql-cost |
| 15 | Cortex Search cost categories incl. warehouse compute, EMBED_TEXT tokens, serving compute, storage | https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-search/cortex-search-costs |
| 17 | ACCOUNT_USAGE has per-service Cortex views (e.g. CORTEX_AI_FUNCTIONS_USAGE_HISTORY, CORTEX_SEARCH_DAILY_USAGE_HISTORY, CORTEX_AGENT_USAGE_HISTORY, CORTEX_ANALYST_USAGE_HISTORY) | https://docs.snowflake.com/en/sql-reference/account-usage |
| 14 | Snowflake-managed MCP server exposes Agents, Analyst, Search, SQL execution, custom tools | https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp |
Unverified, cut or check
- Domain weights 18/38/29/15 and exam code GES-C02: from the official study-guide PDF in the vault (
source-materials/SnowProGenAIStudyGuideC02.pdf, a public download behind a free learning-portal login). Not re-fetched today; Ben should confirm against his copy. - Cut from draft: Cortex Guard (docs page for AI functions no longer mentions it; Cortex AI Guardrails GA May 14, 2026 was confirmed by release-note title only), TruLens/RAG-triad detail (current AI Observability page doesn't mention the RAG triad).
Deliberately left out
Anything from colleague exam notes, the colleague pass checklist, Ben's exam-experience note, the flashcards, and his comment that the exam differed from the study material. No question formats, timing or item topics.