Snowflake Intelligence: ask questions and verify answers

Snowflake Intelligence is the business-data assistant capability that current Snowflake documentation describes as Snowflake CoWork↗ (opens in a new tab). You ask questions in ordinary language about configured company data, then check the evidence before using the answer. The useful outcome is a verified finding with a clear definition, time period, and source. fenbs can keep that work record so another assistant knows what you checked and what still needs an answer.

Start with a question you could explain to a colleague. “How did sales change?” leaves too many choices open. “Compare completed US orders last month with the previous month, excluding canceled orders” gives the investigation a boundary. The quality check begins with the question, before you read a single generated sentence.

Where should you start if you know the Intelligence name?

Use the current CoWork documentation rather than an old screenshot to find your way. For the people preparing the assistant, Snowflake documents Snowsight → AI & ML → Agents → Create agent, followed by Tools, Orchestration, and Save in its getting-started tutorial↗ (opens in a new tab). That is a setup sequence for an agent, rather than a reason to invent an Intelligence button for every business user.

Ask your data team which configured assistant covers the information you need. Request one example question with an answer they have already checked. Run that example first, then make a small change to the date range or business unit. This gives you a reference point when your own question produces a surprising result.

Keep this workflow separate from exposing Snowflake tools to another AI app. The Snowflake MCP server guide covers that connection direction. Here the job is to ask and verify a business question in the business-facing application.

How do you ask a question that can be checked?

1. Name the measure in plain language, then state the definition you expect. If revenue means invoiced revenue, say so.
2. Set the reporting period and comparison period explicitly. Include the reporting timezone when it could change the result.
3. State the population: region, product line, customer group, or other agreed boundary.
4. Ask for the source references, assumptions, and calculation behind the answer. Treat anything missing as a follow-up question.
5. Review one narrow result before broadening the request into an explanation or recommendation.

For example, ask: “Using the approved definition of completed orders, compare the two periods by region. Identify any missing data and explain which date field controls the grouping.” This is a suggested question, not a promise about a particular configured source. If the assistant cannot identify the definition, stop and find its owner.

Separate measurement from interpretation. First establish whether a number changed. Then ask what evidence could explain it. A plausible story about seasonality does not establish that seasonality caused the change. Keep explanations tentative until someone checks the supporting records.

How do you verify the answer before sharing it?

Read each source as evidence for a particular statement. Does it describe the same period, population, and measure? A reference about overall orders does not automatically support a claim about completed orders. You should be able to point from the sentence to the part of the source that supports it.

• Check the definition against the approved business wording, including exclusions and treatment of missing values.
• Reconcile a sample total with a report or calculation your team already trusts.
• Inspect whether the comparison uses matching periods and the same level of detail.
• Mark unsupported explanations as questions rather than carrying them into a presentation as facts.

A simple review sheet can have four lines: question, checked result, supporting reference, and remaining uncertainty. Include who checked it and when. This is enough to make a useful handoff without collecting an entire transcript. The AI at work guides use the same principle: preserve the part that changes the next action.

Why might the assistant give an incomplete or unexpected result?

Check access and configured tools before assuming the data is missing. Snowflake documents that Cortex Agents needs tools to query account data and that callers need the relevant agent and tool privileges. Its Cortex Agents guide↗ (opens in a new tab) also warns that generated answers and citations still need review. A refusal is not an invitation to ask for broader privileges without understanding the task.

Next, narrow the question. Compare one region, one period, or one measure with a known result. If that works, add complexity one piece at a time. Write down the first change that makes the answer diverge. A concrete example is much easier for a data owner to investigate than “the AI gives bad numbers.”

If you are responsible for the assistant itself, building and testing Cortex Agents is the next workflow. For code changes, use the separate Cortex Code guide. Neither task should be hidden inside a business user’s attempt to verify a report.

What should you keep after the conversation?

Keep the confirmed metric definition, source location, reporting boundary, and unresolved question. Save an approved interpretation separately from a generated suggestion. Record what would make the finding stale, such as a changed definition or a new reporting period. This helps the next person decide whether to reuse it or check it again.

Do not treat an uploaded file as permanent shared knowledge. Snowflake’s CoWork guide↗ (opens in a new tab) describes uploads as thread-bound, memory as specific to CoWork, and conversation sharing as a snapshot. Turning off Reference saved memories does not delete saved memories. Those distinctions matter when deciding where the lasting work record belongs.

A short saved conclusion in fenbs gives the next assistant the checked definition and next step; the handoff workflow explains how that record carries work forward. Keep restricted source data in its approved system, and save only the references and conclusions your organization permits elsewhere.

How to keep checked business findings with fenbs?

Keep the answer you checked, the definition behind it, and the question still waiting on someone. Save them together so you can carry on from the same point on any device and in another connected AI.

1. Create a free account.
2. Add a task named after the question you need answered. Write the measure, date range, and person who can confirm the result.
3. Connect Snowflake CoWork to fenbs through the connections listed at Integrations, with your administrator setting up and approving the remote connection first.
4. Save the checked conclusion, its approved source reference, and anything still uncertain. Save the definition a person agreed so the next AI knows which meaning to use.
5. Before stopping, save where you left off and the next check. Your next connected AI can read that record on any device and continue from there.

Snowflake’s MCP connector documentation↗ (opens in a new tab) describes the administrator setup and user authorization needed for remote tools. fenbs is useful for keeping the work between sessions; it does not replace Snowflake’s data controls, calculate your approved metrics, or verify an answer for you.