GES-C02 tests Cortex AI functions more than anything else: "Snowflake Gen AI Functions" is the largest of the four domains at 38%, and governance of those same functions adds another 29%. Together that is about two thirds of the 55 questions on the SnowPro Specialty: Gen AI exam, so knowing which AI_ function fits which job, and who is allowed to call it, is most of the preparation.

 

This entry looks at the exam through that lens. It sorts the functions by the task they solve, walks through the privilege model, explains the three document parsing modes, and ends with a four-week plan and five short answers. The exam figures were read on Snowflake's certification page on 17 November 2026, and the function details come from Snowflake's own documentation.

 

GES-C02 Cortex AI functions cover for the Snowflake SnowPro Specialty Gen AI exam: 55 questions, 85 minutes, 750 of 1000 to pass, Functions 38% and Governance 29%

What does the GES-C02 exam look like?

It is a 55-question, 85-minute exam with a scaled pass mark of 750 out of 1000. The SnowPro Specialty: Gen AI certification page gives the full format:

  • Exam version: GES-C02.
  • Questions: 55, multiple choice and multiple select.
  • Time: 85 minutes, which is about 93 seconds a question.
  • Passing score: 750 on a scaled score from 0 to 1000.
  • Fee: $225, or $180 for candidates registering in India.
  • Prerequisites: none.
  • Language: English.
  • Delivery: online proctoring or an onsite testing center.

Snowflake adds that the exam may include unscored items. They are not marked on the form, they do not affect your score, and extra time is already built in for them.

 

The candidate Snowflake has in mind has at least one year of Gen AI work with Snowflake in an enterprise setting. SQL and data engineering knowledge is assumed, and Python helps.

How are the 55 questions spread over the four domains?

The four domains are Snowflake for Gen AI Overview at 18%, Snowflake Gen AI Functions at 38%, Snowflake Gen AI Governance at 29%, and Snowflake Document Processing at 15%.

 

Snowflake publishes percentages, not question counts. If you apply the weights to 55 questions, you get roughly 10 overview questions, 21 on functions, 16 on governance and 8 on document processing. That is our own arithmetic, not a Snowflake figure, but it makes the point clearly: functions and governance together are about 37 of the 55 questions.

 

GES-C02 square infographic for the Snowflake Gen AI exam: Cortex AI functions grouped by task (generate, classify and filter, extract and summarize, embed and compare, documents and media) with the four domain weights

Which Cortex AI function fits which task?

Most functions questions are really "which function would you use here?" questions. The Cortex AI Functions overview lists each function with one line on what it does. Grouping them by task makes them much easier to remember:

  • Generate: AI_COMPLETE produces a completion from text or an image with the LLM you choose. Snowflake calls it the function to use for most generative AI tasks.
  • Classify and filter: AI_CLASSIFY sorts text or images into categories you define. AI_FILTER returns True or False, so it can sit inside a SELECT, a WHERE clause or a JOIN ... ON condition.
  • Extract and judge: AI_EXTRACT pulls information out of text, images or documents, in several languages. AI_SENTIMENT returns the sentiment of a text.
  • Summarize: AI_SUMMARIZE works on one input. AI_AGG and AI_SUMMARIZE_AGG work across many rows of a column, and the documentation notes that both are not subject to context window limits.
  • Embed and compare: AI_EMBED returns an embedding vector for text or an image. AI_SIMILARITY scores how similar two inputs are.
  • Clean and convert: AI_REDACT removes personal information from text, and AI_TRANSLATE translates between supported languages.
  • Media and documents: AI_TRANSCRIBE turns staged audio or video into text with timestamps and speakers. AI_PARSE_DOCUMENT reads staged documents.

Three helper functions round out the list. TO_FILE points to a file in a stage, AI_COUNT_TOKENS counts tokens before you send a prompt, and PROMPT builds prompt objects for AI_COMPLETE and similar functions.

 

A useful habit is to read each scenario for its shape. One row in, one answer out points to a single-input function. A whole column in, one answer out points to an aggregate function. A yes or no inside a query points to AI_FILTER.

When should you use SQL functions and when the REST API?

Use the SQL functions for batch work and the REST API when a person is waiting for the answer. Snowflake says the functions are optimized for throughput and suit processing many inputs, such as text from large tables. For interactive use where latency matters, it points to the REST API: the Complete API for simple inference, the Embed API for embeddings and the Agents API for agentic applications.

 

That split is easy to test in a scenario question. A nightly job that labels a million support tickets is a SQL job. A chat window in an app is an API job.

Who is allowed to call Cortex AI functions?

A user needs two things: the USE AI FUNCTIONS account-level privilege, and one of the CORTEX_USER or AI_FUNCTIONS_USER database roles. The privileges and model access page explains the rest, and much of the 29% governance domain sits on these rules:

  • Defaults are open. Both USE AI FUNCTIONS and CORTEX_USER are granted to the PUBLIC role by default, so every user can call the functions until an administrator changes that.
  • ACCOUNTADMIN controls it. Only ACCOUNTADMIN manages the account-level privilege, for example by revoking it from PUBLIC and granting it to a chosen role.
  • Per-function grants exist. USE AI FUNCTION followed by a function name, such as AI_COMPLETE, allows just that one function.
  • The blanket privilege wins. The two grants have an OR relationship. A role that holds USE AI FUNCTIONS can call every function, so revoking a per-function grant from it changes nothing.
  • Roles, not users. SNOWFLAKE.CORTEX_USER cannot be granted directly to a user; you grant it to a role.
  • An exception to remember. A user with USE AI FUNCTIONS but without CORTEX_USER can still use AI_AGG and AI_SUMMARIZE_AGG.

The page also notes that USE AI FUNCTIONS does not currently apply to AI function queries run inside Snowflake native applications, and that services running with restricted caller's rights need grants on both the session role and the owner role.

What does a worked governance example look like?

Here is a small scenario, written for this entry and not taken from any exam. A company wants only its analytics team to summarize call notes with AI_COMPLETE, and nobody else to use any AI function.

 

A sound answer runs in four steps:

  1. Close the default. As ACCOUNTADMIN, revoke USE AI FUNCTIONS from PUBLIC. Plan the CORTEX_USER change too, since it is also granted to PUBLIC by default.
  2. Create a role. Make a role for the analytics team rather than granting anything to individual users.
  3. Grant only what is needed. Give that role the per-function privilege for AI_COMPLETE, plus the CORTEX_USER database role.
  4. Check the result. Use SHOW GRANTS on the role and confirm it does not also hold the blanket privilege, because that would undo the narrow grant.

The trap in a question like this is step 4. If the role keeps USE AI FUNCTIONS, the per-function grant adds nothing.

How does document processing work on GES-C02?

Document processing is the smallest domain at 15%, about 8 questions by our arithmetic, and it centers on reading files that sit in a stage. The AI_PARSE_DOCUMENT reference is the page to know:

  • Input: a FILE object, usually built with TO_FILE, pointing to a document in a stage. FL_FILTER_PAGES lets you parse only some pages.
  • OCR mode: text only. This is the default.
  • LAYOUT mode: text plus structure, including tables.
  • AGENTIC mode: text, layout and visual content.
  • Images: the extract_images option pulls out embedded images, and it needs LAYOUT mode.
  • Output: a JSON-formatted string.

Snowflake also notes that the function runs natively in selected regions and through cross-region inference in all regions. A question that mentions tables in a scanned invoice is asking you for LAYOUT, not OCR.

 

The VMExam short below sums up the same point about the two big domains in under a minute.

 

How should you study for GES-C02 in four weeks?

Plan on about 40 hours over four weeks, weighted the way the exam is. Split by domain weight, that is roughly 15 hours on functions, 12 on governance, 7 on the overview and 6 on document processing. That split is our suggestion, not Snowflake's.

  1. Week 1: overview and setup. Read the certification page's four overview points, including Snowpark Container Services and the Model Registry for open-source models. Get a trial account and run AI_COMPLETE on a small table.
  2. Week 2: functions. Run every function in the list once on your own data. Write one line per function on its input shape and output.
  3. Week 3: governance. Revoke and grant the privileges in a test account, then check them with SHOW GRANTS. Write down what each default grant allows.
  4. Week 4: documents and review. Parse one PDF in OCR, LAYOUT and AGENTIC mode and compare the JSON. Finish with Snowflake's own practice exam and the study guide.

For a last round of untimed review after the official material, the GES-C02 practice questions on VMExam are one option to find the domain where you still hesitate.

What are the common mistakes on GES-C02?

The most common mistake is to study the models and skip the rules around them. These are the traps the four domains point to:

  • Learning function names without input shapes. Single-row and aggregate functions are easy to confuse under time pressure.
  • Forgetting the PUBLIC defaults. Access is open by default, so "why can every user call AI_COMPLETE?" has a simple answer.
  • Trusting a narrow grant. A per-function grant does nothing while the blanket privilege is still held.
  • Mixing up parsing modes. OCR for text, LAYOUT for tables and structure, AGENTIC for visual content as well.
  • Ignoring the API side. Interactive, low-latency needs point to the REST API, not a SQL function.

How do you keep the SnowPro Gen AI certification?

Snowflake lists continuing education options for renewal on the certification page. You can pass one of the listed SnowPro exams, which include this Gen AI specialty, the Snowpark and Native Apps specialties and the Advanced series. Or you can complete an eligible instructor-led course, and the list includes the Snowflake Gen AI course.

 

If you are also studying for a VMware cloud exam alongside this one, this blog's sister notes on Blogger explain why the 2V0-17.25 exam tests only 2 of VMware's 5 sections, using the same weight-first approach.

Frequently asked questions

1. How many questions are on the GES-C02 exam?

GES-C02 has 55 multiple choice and multiple select questions, and you have 85 minutes to answer them.

2. What is the passing score for GES-C02?

You need a scaled score of 750 on a scale from 0 to 1000.

3. Which domain carries the most weight on GES-C02?

Snowflake Gen AI Functions is the largest domain at 38%, followed by Snowflake Gen AI Governance at 29%.

4. How much does the GES-C02 exam cost?

The registration fee is $225, and $180 for candidates registering in India.

5. Does GES-C02 have any prerequisites?

No. Snowflake lists no prerequisites, though it expects about a year of hands-on Gen AI work with Snowflake.