Pick the first five questions
Start with the repeated questions that already create meeting friction instead of trying to catalog the whole warehouse.
Analytics governance field packet
Analytics Agent Source-of-Truth Register
Before an analytics agent answers recurring business questions, give each question a row: canonical metric, dataset, owner, grain, filters, freshness rule, query path, and review trigger. The register keeps confident answers from outrunning the metric layer underneath them.
An analytics agent and a dashboard return plausible but different answers.
A team is evaluating BI copilots, warehouse chat, or self-service analytics agents.
Recurring revenue, marketing, support, or finance questions need a named source and owner.
A metric definition changes often enough that stale reference docs would mislead the agent.
Leadership wants decision-grade numbers, but the current process still depends on tribal knowledge.
The team needs a practical governance artifact before a broader semantic-layer project.
How to use it
Start with one row, one owner, and one review trigger
Definitions before SQL
The first problem is deciding which definition the agent may use; SQL generation comes after the business meaning is pinned down.
Maintenance is the product
Tie each row to the same data-model, metric, dashboard, or monthly reporting changes that can make the answer stale.
Escalate outside the register
If a question has no row, stale freshness, or conflicting sources, the agent should refuse, label it exploratory, or route to a human.
Copyable worksheet
The register fields
Copy the table into a planning document or spreadsheet. Keep the first pass narrow: one production workload or five recurring questions before creating a broad catalog.
| Field | What to write | Why it matters |
|---|---|---|
| Business question category | The repeated question the agent is allowed to answer, such as active customers, conversion rate, or pipeline value. | Two teams can ask the same words and mean different queries. |
| Canonical metric and dataset | The approved metric definition and source table, view, semantic layer, or curated query path. | If three sources could answer, the agent needs the approved one, not a plausible one. |
| Owner | The named person accountable for the definition, source, and answer boundary. | A metric owned by a team alias has no one to page when the number looks wrong. |
| Grain | Per day, per account, per customer, per invoice, per opportunity, or another required grain. | The same metric at the wrong grain is a different, wrong number. |
| Required filters / exclusions | Test accounts, churned rows, internal traffic, refunds, inactive records, sandbox data, or other exclusions. | Filters that live in someone's head cannot guide an analytics agent. |
| Freshness SLA | How current the source must be before the agent may answer or cite it. | A confident answer from stale data is worse than a visible refusal. |
| Approved semantic-layer / query path | The blessed view, metric, skill, dbt model, dashboard source, or SQL template the agent routes through. | Maintained context is what separates self-service analytics from raw warehouse guessing. |
| Known gotchas | Edge cases, naming collisions, business exceptions, seasonal caveats, or historical definition changes. | Institutional memory needs to be written where the agent and reviewer can use it. |
| Last changed date | When the metric, source model, query path, or definition last moved. | A stale date is the fastest signal that a row needs review before a meeting. |
| Review trigger | Data-model PR, metric definition change, dashboard edit, monthly reporting close, anomalous answer, or leadership-deck use. | Maintenance is the product; the trigger makes it operational. |
| Example answer boundary | What the agent may answer directly, what it must label exploratory, and what it must escalate. | Exploratory answers and board-deck numbers should not share the same permission. |
Copy block
Paste this plain-text version into your next working doc
The register is intentionally portable. It should survive a meeting, a pull request, a wiki page, or a spreadsheet before it becomes a polished internal tool.
Analytics Agent Source-of-Truth Register Team / domain: Register owner: First five recurring questions: | Field | What to write | Owner / source | Last changed | Review trigger | | --- | --- | --- | --- | --- | | Business question category | | | | | | Canonical metric and dataset | | | | | | Owner | | | | | | Grain | | | | | | Required filters / exclusions | | | | | | Freshness SLA | | | | | | Approved semantic-layer / query path | | | | | | Known gotchas | | | | | | Last changed date | | | | | | Review trigger | | | | | | Example answer boundary / escalation rule | | | | | If the question is not on the register, the agent should: If freshness fails, the agent should: If two sources disagree, the agent should: Who reviews decision-grade answers:
Example row
A sample row, clearly marked as sample data
- Business question category
- Which customers are active this quarter?
- Canonical metric and dataset
- Active customer count from the billing-account mart, excluding test accounts and cancelled accounts after grace period
- Owner
- Revenue operations lead
- Grain
- Account per fiscal quarter
- Required filters / exclusions
- Exclude internal/test accounts, unpaid trials, cancelled accounts beyond grace period, and duplicate billing entities
- Freshness SLA
- Billing mart refreshed within 24 hours before the agent can answer
- Approved semantic-layer / query path
- analytics_metrics.active_customer_count_quarterly view or the matching curated agent skill
- Known gotchas
- Marketing dashboard includes trial users; finance report uses paid billing accounts only
- Last changed date
- 2026-07-03
- Review trigger
- Metric PR, billing-schema change, dashboard edit, monthly close, or leadership-deck request
- Example answer boundary
- Agent can answer with source and freshness note; leadership-deck use requires RevOps review
Next step
Bring five recurring metric questions to a source-of-truth review
BaristaLabs can help pick the first recurring questions, map the approved metric and dataset, define freshness and owner rules, and decide when an analytics answer needs human review before it leaves the room.
Map five metric questionsSource notes
Claims checked before publication
- Anthropic's July 2026 self-service analytics post still states that 95% of its business analytics queries are automated via Claude with roughly 95% aggregate accuracy, while warning that warehouse access alone can create a false sense of precision.
- The Anthropic post still identifies concept/entity ambiguity, data staleness, and retrieval failure as major error modes for analytics answers.
- The Anthropic post still says Claude's analytics accuracy without skills did not exceed 21% on internal evals, while curated skills pushed aggregate accuracy consistently above 95% and sometimes near 99% in specific domains.
- The Anthropic post still describes offline accuracy drifting from roughly 95% to roughly 65% over a month until skill maintenance became an engineering process tied to model changes and code review.
- Reddit discussion links from the source article were treated only as conversation signal during planning; they are not used here as authority for numeric or process claims.
Related resources
Keep the register connected to controls, receipts, and services
AI workflow controls
Decide what an assistant may read, surface, answer, escalate, log, and roll back before analytics access expands.
Open resourceAI workflow security review worksheet
Map source systems, fields, credentials, vendor exposure, retention, and unresolved questions before analytics data is exposed.
Open resourceApproval queue guide
Route decision-grade metrics to review when consequence, freshness, or ambiguity requires a person.
Open resourceAgent receipt template
Record the question, source, freshness, owner, answer, review threshold, and final use of each analytics run.
Open resourceProcess automation
Use the register in reporting, CRM, operations, finance, and handoff workflows before automating answers.
Open resourceCustom solutions
Build internal reporting tools and agent-assisted dashboards around an owned metric layer.
Open resource