Snowflake AI Agent: Automate Board Metrics Briefs in Slack
Turn a Snowflake KPI query into a reviewable board metrics brief, with an agent writing the narrative and Slack delivering it on schedule.

TL;DR: what you'll build. A finance or RevOps lead hand-builds the weekly board metrics brief from Snowflake every cycle, on a deadline. A Snowflake AI agent can automate it: approved SQL runs against curated warehouse tables, results are compared with prior periods, and a concise narrative goes to an executive Slack channel after review. A Major app keeps the query, metric history and approvals in governed state, while the agent handles interpretation and exceptions.
Why automate the weekly board metrics brief with a Snowflake AI agent?
The weekly or monthly board metrics packet is usually built by hand. A financial analyst or RevOps lead pulls revenue, pipeline, churn and headcount numbers from Snowflake into a spreadsheet, writes two sentences of commentary on what moved and why, and posts it to an executive Slack channel before board prep starts.
The work is repetitive, happens on a deadline, and breaks in small ways every cycle:
- A view gets renamed.
- A number comes in late.
- Someone forgets to flag that pipeline dropped because of a data refresh rather than a real change in bookings.
Who owns this problem
This is squarely a business operations problem that sits across finance and RevOps. It belongs in the function hub for AI agents for finance.
The owner is usually a finance or analytics operations lead, not a data engineer. That is exactly why the fix should not be "write a Python script."
Separating repeated work from judgment work
The repeated work is deterministic and belongs in software: pulling the same KPI set from Snowflake, comparing it to the prior period, and posting a draft for review.
Deciding whether a number's movement is a story worth telling, and handling exceptions, is a reasonable job for an agent, as long as a person signs off before it reaches a board-facing channel.
Reason once, run forever
Major's wedge here is "reason once, run forever." A Major app persists the query definitions, run history and approval state, so the model does not rebuild the warehouse connection or re-derive metric definitions on every invocation.
It reasons about this week's numbers once. The deterministic plumbing around it (scheduling, retrieval, formatting and routing) runs the same way every time.
Which Snowflake data and APIs does the workflow use?
Snowflake is the source of truth for the numbers, and Slack is where the reviewed brief lands.
Snowflake
Revenue, pipeline, churn and headcount typically live in curated analytics views built on top of raw transactional tables.
The workflow reads through the Snowflake SQL API, a REST API built to submit SQL statements, check status, and retrieve or cancel results from outside the Snowflake UI or client libraries.
- Objects and endpoints: the statements endpoint accepts a SQL statement and returns a handle you poll for status. A separate call fetches or cancels results once the statement completes. It supports standard queries plus most DDL and DML, and a documented statement API list covers how paginated result sets come back when a query returns many rows.
- Auth: key-pair JWT authentication, where a user's private key signs a bearer token valid for up to one hour, or OAuth. Neither credential should live in application code; Major's credential proxy holds it and injects it per call.
- Webhooks vs polling: there is no native webhook push for query completion in the documented SQL API. The workflow has to poll the statement status endpoint until the query finishes.
- Rate limits: Snowflake returns HTTP 429 on excess request frequency or concurrency. The documented guidance is exponential backoff with jitter, not a fixed retry interval.
- Gotcha from the docs: the JWT is time-boxed to roughly an hour. A long-running or heavily throttled run has to re-mint its token mid-poll rather than assuming one token survives the whole cycle.
Polling is like checking a restaurant pager by hand: Snowflake will not buzz you, so you keep glancing, waiting a little longer between each glance.
Slack
Slack is the delivery and review surface. The workflow posts through the Slack Web API, specifically the chat.postMessage method.
- Objects and endpoints: chat.postMessage delivers the narrative to a channel. The
conversations.listmethod helps if the app needs to resolve a channel ID dynamically rather than hardcoding it. - Auth: an OAuth 2.0 bearer token scoped to a Slack app installed in the workspace, sent in the
Authorizationheader. - Webhooks vs polling: Slack's Events API can deliver event subscriptions to an HTTPS endpoint. This workflow is purely outbound, so no inbound webhook is required unless you later want to catch a reviewer's reply in-thread.
- Rate limits: Slack's documented rate limits assign chat.postMessage to a special tier that generally allows one message per second per channel, plus a workspace-wide ceiling. Events API deliveries cap at 30,000 per workspace per app per 60 minutes. Exceeding a limit returns HTTP 429 with a
Retry-Afterheader. - Gotcha from the docs: a short burst above the per-second posting limit is tolerated, but Slack gives no guarantee the extra messages will display. If the app posts a draft and a correction in quick succession, it should space them.
What governance actually requires here
Board metrics touch internal financial data, but for a cross-industry use case like this there is no sector-specific regulation such as HIPAA or SOC-controlled PHI in play. The governance need is correctness and access control.
Because this content later presents itself to a board, it has to be the output of approved SQL against approved views, posted by a bot token that cannot see channels it should not.

Snowflake, Major and Slack architecture for an automated board metrics brief
Split the work across app, agent and workflow
Keeping the three parts separate is what makes the brief trustworthy.
The workflow: the deterministic backbone
A weekly schedule kicks off a run, for example Monday morning before board prep, with timing configurable by the operator.
The workflow submits the same fixed set of approved SQL statements to Snowflake's SQL API, polls the statement status endpoint with backoff until each query resolves, then writes the raw results into the app's database as a new report run.
If a query errors or a KPI is missing, it stops that metric's path and flags it rather than guessing. It then hands off to the agent for drafting and waits on a human approval step before it may call Slack's chat.postMessage.
The agent: one narrow job
The agent reads the current and prior period's values, both already fetched and stored by the workflow, and writes two or three sentences describing the direction of each metric relative to its threshold.
It never calculates a number itself or queries Snowflake directly. If a value is missing, inconsistent with its historical range, or the query returned an error code, it says so plainly and does not invent an explanation.
Deciding whether "pipeline down 4%" merits a flag or is within normal week-to-week noise is a language task, not a SQL task.
The app: system of record and review surface
The app stores:
- Metric definitions.
- Every run's results, with their query status and timestamps.
- Every narrative draft, with its approval state.
The finance owner reviews drafts in the app, not in Slack. This keeps the executive channel clean of half-finished output and gives a durable audit trail independent of Slack's own message history.
Build the Snowflake-to-Slack brief on Major
The app's data model comes from three tables:
- Metric definition table: name, sqlorview, owner, unit, threshold.
- Report run table: period, querystatus, results (JSON), fetchedat, provenance.
- Narrative draft table: text, evidencerefs (links back to report run rows), reviewer, approvalstate.
- Connect Snowflake with a key-pair or OAuth credential scoped to a read-only role against the curated KPI schema, and attach Slack as an app installed in the workspace with
chat:writescoped to the board-prep channel, with both credentials held by Major's credential proxy and never embedded in agent prompts or app code. You should see both connectors listed as connected, with the Snowflake role limited to the curated KPI schema and the Slack app limited to the board-prep channel. - Build the app by modeling the three tables in the app's managed database and giving Major the app prompt below, which asks for a run history list, a draft review screen with the source Snowflake values next to the agent's narrative, and an approve or reject action that gates the Slack post button. You should see a run history list and a review screen where the post button stays disabled until a draft is approved.
- Add the agent with the agent prompt below, which restricts it to approved metric values from the current and prior period's report runs, asks for direction and threshold status for each metric, requires explicit flags for missing or inconsistent values without an inferred cause, and caps the narrative at three sentences. You should see a short narrative that cites only stored values and calls out gaps in plain words.
- Set the workflow trigger as a weekly schedule (for example, every Monday at 7am local time, adjustable per operator) so the workflow submits the approved SQL statements, polls for completion with exponential backoff, writes the report run, and invokes the agent to populate a narrative draft in the pending state. You should see a new report run row with query status and fetched-at timestamps, followed by a pending narrative draft linked to it.
- Test the failure paths before the first real Monday by forcing a query error, a 429 response from Snowflake or Slack, a partial result set, and a manual rerun in the same period, and confirm the workflow flags the failed metric, backs off with jitter, honors Slack's Retry-After header, and reuses the existing report run rather than creating a second draft. You should see flagged gaps instead of silent zeros, one draft per period, and no duplicate or premature Slack posts.
- Roll out with permissions by adding the human-in-the-loop step, where the workflow halts after drafting until the finance owner approves, edits or rejects the draft in the app, and by limiting app access through SSO and role-based access control to the finance owner and designated reviewers. You should see the post to the restricted channel fire only after an approved state change, with the approver and timestamp recorded in the audit trail.
The app prompt for step 2:
"Build an internal app with tables for metricdefinition, reportrun and narrative_draft. Show a run history list, a draft review screen with the source Snowflake values next to the agent's narrative, and an approve/reject action that only enables the Slack post button once a draft is approved."
The agent prompt for step 3:
"Use only these approved metric values from the current reportrun and the prior period's reportrun. Describe the direction of each metric and whether it is inside or outside its threshold. If a value is missing or inconsistent with history, flag it explicitly and do not infer a cause. Keep the narrative to three sentences."

Human approval gate between query results and the Slack post
Governance and IT control
Governance here comes down to who can see and approve drafts, what each credential can touch, and whether the record survives.
Access and audit
Access to the app runs through Major's SSO and role-based access control, so only the finance owner and designated reviewers can see draft narratives or trigger an approval.
Every run, query, draft edit and approval is logged with a timestamp and actor in the app's audit trail. That matters when a board member later asks why a number read a certain way in a given week.
Scoped credentials
The Snowflake credential is scoped to a read-only role on the curated KPI views, nothing broader, and held in Major's credential proxy rather than in the workflow definition or agent prompt. The Slack bot token can post only in the one approved channel.
Like a hotel key card that opens your room and the gym but not every door, a lost credential does limited damage.
Durable records and versioned changes
The managed database holds report run and narrative draft history, so the record of what was approved outlives any Slack message, channel archive or change to Slack's retention policy.
Deploys of the app and workflow are versioned. A change to the metric definitions or the agent's prompt is reviewable before it goes live, rather than silently changing what gets reported next Monday.
The Slack post is the one action in this pipeline that needs explicit human approval before it executes. Every other step is deterministic and auditable but does not require a person in the loop.
Failure handling and edge cases
Data and connection failures
- Snowflake query errors or timeouts. If a statement errors or never completes within the polling window, the workflow marks that metric's report run as failed, so the agent sees a flagged gap, not a silent zero.
- Rate limiting. Both Snowflake and Slack return HTTP 429 under load. The workflow applies exponential backoff with jitter on Snowflake's status polling and respects Slack's Retry-After header rather than retrying immediately.
- Partial syncs. If three of four KPIs come back but one errors, the app still assembles a draft noting which metric is missing. The reviewer decides whether to post partial results or wait.
- Stale data. If the underlying Snowflake views have not refreshed since the last run, the app flags the data's actual timestamp next to the narrative so the reviewer can judge freshness.
Duplicate runs
A second trigger in the same period, such as a manual rerun plus the scheduled one, should not produce two drafts awaiting approval.
The workflow checks for an existing report run in the current period before creating a new one, like a mail sorter checking whether this week's folder already exists.
Agent limits and the fallback
The agent must never calculate a KPI value, post to Slack directly, or fabricate a cause for a flagged anomaly. Those are deterministic or human decisions.
If the agent cannot produce a coherent draft, or Snowflake's connection is down past the retry window, the workflow routes the run to the finance owner as a manual-review case. It does not fail silently.
What to measure
This workflow should make existing metrics visible rather than invent new ones:
- Cycle time, from scheduled trigger to approved Slack post, to see whether board prep is waiting on this step.
- Exceptions caught, meaning how often a metric comes back flagged as missing or out of threshold before it reaches the reviewer.
- Manual touches, meaning how many drafts get edited versus approved as-is, which tells you whether narrative quality is holding up.
- Query failure rate, meaning how often Snowflake statements error or time out, which points at upstream data pipeline health.
- Approval turnaround, meaning how long a draft sits pending, since that is now the only manual step left.
Before you go live
- [ ] The Snowflake credential uses key-pair JWT or OAuth, sits in Major's credential proxy, and maps to a read-only role on the curated KPI views only.
- [ ] The Slack app has chat:write scoped to the single board-prep channel, and the bot cannot read or post elsewhere.
- [ ] Every metric definition has an owner, unit, threshold and approved SQL or view recorded in the app.
- [ ] The workflow re-mints the Snowflake JWT mid-poll if a run approaches the one-hour token limit.
- [ ] Backoff with jitter is in place for Snowflake 429s, and Slack posts honor the Retry-After header.
- [ ] A rerun in the same period reuses the existing report run instead of creating a second pending draft.
- [ ] The Slack post fires only after an approved state change, and the approval is logged with actor and timestamp.
- [ ] Stale data shows the views' actual refresh timestamp next to the narrative, and agent failures route to the finance owner for manual review.
Start building on Major
Build a Snowflake metrics app that keeps KPI runs, narrative drafts and approvals in governed state, then routes the reviewed weekly brief to the right Slack channel.
Related articles
- Greenhouse BambooHR Integration: Automate New-Hire Onboarding
- Automate Claim Document Signing in Guidewire ClaimCenter with DocuSign
- Automate Marketo Lead Audiences in Google Ads with AI
About the author
Ritvij Gautam is Head of Growth at Major.
Last reviewed November 7, 2026 by Ritvij Gautam.
Frequently asked questions
- Can an AI agent query Snowflake and post results to Slack?
- Yes. A workflow can run approved SQL against Snowflake's SQL API on a schedule, then post the result through Slack's chat.postMessage once a reviewer approves the draft. The agent only writes the narrative; it never has direct Slack posting rights or unscoped warehouse access.
- How do you keep an AI-generated board report accurate?
- Fix the metric definitions and SQL in advance, fetch values deterministically through Snowflake's SQL API, and show the agent's narrative next to the source values during review. If a value is missing or inconsistent with history, the agent flags it instead of guessing, and a human signs off before anything reaches the board channel.
- What Snowflake API does a reporting agent use?
- The Snowflake SQL API: a REST API for submitting SQL statements, checking execution status, and retrieving or canceling results. It authenticates with key-pair JWT or OAuth and has no native webhook for query completion, so the workflow polls for status.
- Can the report run automatically every week?
- Yes, through a scheduled trigger on a Major workflow that fires weekly, runs the approved queries, drafts the narrative, and waits for approval before posting. Snowflake's own CoWork product documents a separate scheduled automation path; this is a distinct architecture built around Major's app and workflow layer.
- Does this require a dedicated data engineer to build?
- No. The SQL, metric definitions and Slack channel are configured once by the finance or analytics operations owner. Major's connectors handle the Snowflake and Slack authentication, and the app's managed database stores run history without custom pipeline code.
- What happens if the finance owner rejects a draft?
- The draft is returned with reviewer notes and never reaches Slack. The workflow can regenerate a new draft from the same source values, or the owner can edit it directly in the app before approving.