Overview
Snowflake Gen AI Governance is 29% of GES-C02, the second-largest domain, so plan for about 16 questions. It has four objectives. Objective 3.1 covers model access controls: restricting which models can be used through the account allowlist, RBAC and application roles; data safety through cross-region inference settings, Cortex Guard guardrails and AI_REDACT; reducing hallucination and bias; and authenticating REST API calls.
Objective 3.2 covers granting and revoking privileges for Cortex Analyst, Cortex Search, Cortex Agents and Snowflake Intelligence, and the SNOWFLAKE database roles CORTEX_USER, CORTEX_ANALYST_USER, CORTEX_AGENT_USER and CORTEX_EMBED_USER. Objective 3.3 covers cost: limiting agent token use, Cortex Search cost components, minimizing AI function tokens, tracking compute pools, usage history views and object tagging. Objective 3.4 covers AI observability, including evaluation metrics, run comparisons, tracing, logging to event tables and the TruLens SDK.
How it's tested
Domain 3 questions put you in the administrator's seat. A security, compliance or budget requirement is stated, and you choose the parameter, role, grant, view or tool that meets it with least privilege. Watch for defaults, account-level versus object-level controls, and which usage view holds the data being asked about.
-
Lock down access
A policy limits who may call Cortex or which models are allowed, and you choose the role grant, revoke or allowlist setting.
-
Find and cut cost
Spend is rising and you identify the usage view to inspect or the setting that reduces token or compute consumption.
-
Evaluate quality
An app gives poor or ungrounded answers and you choose evaluations, tracing or guardrails to measure and reduce the problem.
A bank wants only its data science role to use Cortex LLM functions and only two approved models. Because CORTEX_USER is granted to PUBLIC by default, the admin revokes it from PUBLIC and grants it to the data science role. Setting the CORTEX_MODELS_ALLOWLIST account parameter to the two models enforces the model limit. Revoking warehouse usage would block far more than Cortex.
Objectives 4
Official objectives from the Snowflake outline, in outline order.
3.1 Set up model access controls.Model restrictions, data safety and REST API authentication.
- What it means
- This objective covers restricting model use through application roles, RBAC and the account-level model allowlist parameter; protecting data with cross-region inference settings, Cortex Guard guardrails and AI_REDACT; reducing hallucinations and bias; and authenticating calls to Cortex REST APIs.
- What to understand
- CORTEX_MODELS_ALLOWLIST limits which models can be called across the account, and model-specific access can be managed with RBAC. Cross-region inference is disabled by default and should stay off when data must not leave the region. Cortex Guard filters unsafe model output, and AI_REDACT removes personal data before it reaches prompts or logs. Grounding answers with retrieval and clear prompts reduces hallucination. REST API calls authenticate with methods such as key-pair JWT, OAuth or programmatic access tokens.
- How it shows up
- A compliance rule about models, residency or personal data is given and you choose the control that enforces it.
- Common confusion
- Assuming cross-region inference is on by default, or that a guardrail redacts personal data in inputs.
3.2 Grant and revoke Role-Based Access Control (RBAC) and privileges.Cortex database roles and object privileges.
- What it means
- This objective covers granting and revoking the individual privileges needed for Cortex Analyst, Cortex Search, Cortex Agents and Snowflake Intelligence, and the SNOWFLAKE database roles CORTEX_USER, CORTEX_ANALYST_USER, CORTEX_AGENT_USER and CORTEX_EMBED_USER.
- What to understand
- CORTEX_USER allows calling Cortex AI functions and is granted to PUBLIC by default. CORTEX_EMBED_USER is narrower and covers embedding functions only. Querying a Cortex Search service needs USAGE on the service plus its database and schema. Cortex Analyst needs access to the semantic view or the stage holding the YAML model and SELECT on the underlying tables. Agents need USAGE on the agent and on the tools it calls. Grant database roles with GRANT DATABASE ROLE.
- How it shows up
- A user gets a privilege error or must be limited to a subset of Cortex features, and you pick the exact grant or revoke.
- Common confusion
- Granting ACCOUNTADMIN or broad roles instead of the specific Cortex database role and object privileges.
3.3 Manage, monitor, and optimize Snowflake Cortex costs.Cost drivers, usage views and optimization for each Cortex service.
- What it means
- This objective covers limiting token use by agents, understanding Cortex Search cost types (warehouse, embedding, serving and indexing), minimizing tokens in AI functions, tracking Snowpark Container Services compute pools, reviewing usage history views and applying object tagging for cost attribution.
- What to understand
- AI functions bill by tokens, and some count both input and output, so shorter prompts, smaller models and task-specific functions save money. Cortex Analyst bills per message. Cortex Search charges for the warehouse that refreshes the index, embedding tokens, and serving based on indexed data size, so suspend unused services and set sensible target lag. Compute pools bill per node while active, so set auto-suspend. ACCOUNT_USAGE views such as CORTEX_FUNCTIONS_USAGE_HISTORY show consumption, and tags attribute it to teams.
- How it shows up
- Spend has jumped on a specific service and you pick the view to inspect or the change that lowers cost.
- Common confusion
- Assuming Cortex Search costs only warehouse credits and ignoring serving and embedding charges.
3.4 Use Snowflake AI observability tools.Evaluation, tracing and logging for Gen AI apps.
- What it means
- This objective covers AI observability in Snowflake: running evaluations with metrics, comparing runs or app versions, tracing each step of a request, logging to event tables, and instrumenting applications with the TruLens SDK.
- What to understand
- AI Observability uses LLM-as-judge metrics such as context relevance, groundedness and answer relevance to score RAG and agent responses. Comparing runs shows whether a new prompt, model or retrieval setting is better. Traces record each step, such as retrieval and generation, with latency and token counts, and are stored in an event table. The TruLens SDK instruments Python apps so these traces and evaluations are captured.
- How it shows up
- An app returns weak or ungrounded answers and you choose how to measure quality or locate the failing step.
- Common confusion
- Relying on usage history views for answer quality, when they report consumption rather than accuracy.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
SNOWFLAKE.CORTEX_USER
Database role for Cortex AI functions.
- Remember
- Granted to PUBLIC by default.
- Tested as
- Access control questions.
- Trap
- Forgetting to revoke from PUBLIC.
-
CORTEX_EMBED_USER
Narrow role for embedding functions only.
- Remember
- Use for least privilege on embeddings.
- Tested as
- Least privilege questions.
- Trap
- Granting full CORTEX_USER instead.
-
CORTEX_ANALYST_USER and CORTEX_AGENT_USER
Roles scoped to Analyst or Agents.
- Remember
- Separate from general function access.
- Tested as
- Feature-level access questions.
- Trap
- Assuming CORTEX_USER covers all cases.
-
CORTEX_MODELS_ALLOWLIST
Account parameter listing permitted models.
- Remember
- Values like all, none or a list.
- Tested as
- Model restriction questions.
- Trap
- Trying to restrict models per warehouse.
-
Cross-region inference setting
Account parameter controlling routing.
- Remember
- Disabled by default.
- Tested as
- Residency questions.
- Trap
- Enabling it despite residency rules.
-
Cortex Guard
Guardrail filtering unsafe LLM output.
- Remember
- Option on AI_COMPLETE.
- Tested as
- Safety questions.
- Trap
- Thinking it redacts personal data.
-
AI_REDACT
Removes personal information from text.
- Remember
- Apply before prompting or logging.
- Tested as
- Data safety questions.
- Trap
- Confusing it with masking policies.
-
Cortex Search costs
Warehouse, embedding, serving and indexing.
- Remember
- Serving scales with indexed data.
- Tested as
- Cost questions.
- Trap
- Counting only warehouse credits.
-
Usage history views
ACCOUNT_USAGE views for Cortex consumption.
- Remember
- Break down by function, model and query.
- Tested as
- Monitoring questions.
- Trap
- Expecting real-time data.
-
Object tagging
Tags on objects for attribution.
- Remember
- Supports chargeback by team.
- Tested as
- Cost allocation questions.
- Trap
- Using comments instead of tags.
-
AI Observability
Evaluations, tracing and logging for apps.
- Remember
- Uses LLM-as-judge metrics.
- Tested as
- Quality questions.
- Trap
- Using cost views to judge accuracy.
-
TruLens SDK
Instrumentation library for Gen AI apps.
- Remember
- Captures traces and evaluations.
- Tested as
- Observability questions.
- Trap
- Confusing it with a model.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Cortex database roles
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| CORTEX_USER | Call Cortex AI functions | Broad; on PUBLIC by default | Leaving it on PUBLIC |
| CORTEX_EMBED_USER | Embedding functions only | Narrowest function role | Granting CORTEX_USER for embeddings |
| CORTEX_ANALYST_USER | Use Cortex Analyst | Scoped to text-to-SQL | Forgetting table SELECT grants |
| CORTEX_AGENT_USER | Use Cortex Agents | Scoped to agents | Forgetting tool privileges |
Safety controls
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Model allowlist | Limits callable models | Account-wide setting | Expecting per-user effect |
| Cortex Guard | Filters harmful output | Acts on responses | Using it for PII |
| AI_REDACT | Removes personal data | Acts on text you pass in | Confusing with masking policies |
Monitoring cost vs. quality
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Usage history views | Credits and tokens consumed | Answers how much was spent | Using them to judge accuracy |
| AI Observability | Evaluations and traces | Answers how good the output is | Using it for billing |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Grant only, never revoke
- What candidates pick
- Grant CORTEX_USER to the approved role and stop there.
- Why it looks right
- Granting feels like enough.
- Why it's wrong
- Because PUBLIC holds CORTEX_USER by default, every user keeps access until it is revoked from PUBLIC.
Clue to watch for Only one team should be able to call Cortex.
-
Warehouse as the gate
- What candidates pick
- Revoke warehouse usage to block Cortex.
- Why it looks right
- Queries need warehouses.
- Why it's wrong
- That blocks all querying; the targeted control is the Cortex database role plus the model allowlist.
Clue to watch for Other SQL work must continue unaffected.
-
Guard redacts PII
- What candidates pick
- Use Cortex Guard to remove names and account numbers.
- Why it looks right
- Guard sounds protective.
- Why it's wrong
- Cortex Guard filters unsafe responses; AI_REDACT removes personal information from text.
Clue to watch for The concern is personal data, not harmful content.
-
Search costs warehouse only
- What candidates pick
- Shrink the refresh warehouse to cut Search spend.
- Why it looks right
- Warehouses are familiar cost drivers.
- Why it's wrong
- Serving cost grows with indexed data and runs while the service is active, so suspending unused services or indexing less data may matter more.
Clue to watch for Spend continues when no refresh runs.
-
Enable cross-region for speed
- What candidates pick
- Turn on cross-region inference for faster responses.
- Why it looks right
- More regions sounds like more capacity.
- Why it's wrong
- It routes data to other regions, which may breach residency rules; it exists for model availability, not for speed.
Clue to watch for A data residency requirement is stated.
-
Cost views for quality
- What candidates pick
- Query usage history to find why answers are wrong.
- Why it looks right
- Those views are familiar.
- Why it's wrong
- Usage views report consumption; groundedness and relevance come from AI Observability evaluations and traces.
Clue to watch for The problem is answer quality.
Answer framework
Apply these steps to any Domain 3 governance, cost or quality scenario.
- Identify the concern: who can use Cortex, which models, where data is processed, how much it costs, or how good the answers are.
- For access, start from defaults: revoke CORTEX_USER from PUBLIC if needed, then grant the narrowest Cortex database role and object privileges.
- For models and residency, use the CORTEX_MODELS_ALLOWLIST parameter and keep cross-region inference disabled unless policy allows routing data elsewhere.
- For cost, identify how the service bills, inspect the matching ACCOUNT_USAGE view, then reduce tokens, suspend idle services or pools, and tag objects.
- For quality, run AI Observability evaluations, compare runs and review traces in the event table before changing prompts, retrieval or models.
Practice question
Want more like this? The GES-C02 practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
This domain is 29% of the exam and rewards exact knowledge, so study it with the documentation open and a trial account ready. Begin with access control. Write down every Cortex database role, what it allows and its default grant, then practice revoking CORTEX_USER from PUBLIC and granting it to a custom role. Add the object privileges for Search services, semantic views and agents, and test them by switching roles until errors disappear.
Next, learn the account parameters for the model allowlist and cross-region inference, and contrast Cortex Guard with AI_REDACT using small examples. Review REST API authentication options.
For cost, build a table of how each service bills and which ACCOUNT_USAGE view reports it, then run a few functions and query the views to see the data. Practice tagging a warehouse or compute pool.
Finish with AI Observability: instrument a small RAG app with TruLens, run an evaluation, compare two runs and read the traces. Close with timed practice on least-privilege and cost scenarios.
FAQ
How much of the SnowPro Gen AI exam is Snowflake Gen AI Governance?
Snowflake Gen AI Governance is 29% of GES-C02, the second-largest domain, or about 16 of 55 questions. It has four objectives: model access controls, RBAC and privileges, cost management and AI observability.
Who can use Cortex AI functions by default?
Every user, because the SNOWFLAKE.CORTEX_USER database role is granted to PUBLIC by default. To restrict access, revoke it from PUBLIC and grant it to chosen roles.
How do I restrict which models can be used?
Set the account-level CORTEX_MODELS_ALLOWLIST parameter to the approved models, and use RBAC for finer model access where needed.
What drives Cortex Search cost?
The warehouse that builds and refreshes the index, embedding tokens for indexed text, and serving charges that scale with the size of indexed data while the service is active.
Where do I see Cortex usage?
In SNOWFLAKE.ACCOUNT_USAGE views for Cortex, such as CORTEX_FUNCTIONS_USAGE_HISTORY and service-specific views for Search and Analyst. These views have some latency.
What is AI Observability in Snowflake?
Tools to evaluate and trace Gen AI apps, using LLM-as-judge metrics such as groundedness and relevance, run comparisons, and traces stored in event tables, often captured with the TruLens SDK.
What is the difference between Cortex Guard and AI_REDACT?
Cortex Guard filters potentially harmful model output. AI_REDACT removes personal information from text, which helps keep sensitive data out of prompts, outputs and logs.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
You have covered Cortex roles, model allowlists, data safety controls, cost monitoring and AI observability. Check whether you can apply least privilege and cost control under exam timing.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the GES-C02 AI tutor Get any concept explained your way