Natural-language questions over the revenue warehouse, answered only through governed metric definitions and with the SQL always shown — so “what's our CAC payback by segment” returns the number finance would sign, not a fluent guess.
The GTM Analytics Copilot answers natural-language questions over the revenue warehouse only through governed metric definitions, and always shows the SQL, so the answer is a number finance would sign rather than a fluent guess. It is one of fifteen workflows in the Enterprise GTM Agent Stack, an Opsmarshal series on GTM stack architecture.
How the flow runs — live
Questions outside governed definitions get “define it first,” not improvisation. Refusal is a feature.
Every GTM team runs on numbers that disagree: marketing's CAC, finance's CAC, and the board deck's CAC are three numbers. Self-serve BI made querying easy and consistency worse; LLM-over-SQL without governance makes confident wrongness instant.
The copilot answers only through the semantic layer. Ambiguous questions get a clarifying choice — “pipeline-influenced or closed-won attribution?” — instead of a silent assumption. Every answer shows its SQL and cites its definition version.
| Decision | Choice | Rationale |
|---|---|---|
| Refusal is a feature | Ungoverned questions get “define it first,” not improvisation | The failure mode of analytics copilots is confident answers to ungoverned questions. Refusing cheaply beats retracting expensively. |
| Read-only, PR-governed definitions | The copilot can't create metrics; humans version them | Metric sprawl is how the three-CAC problem started. Definitions change through review, and the copilot always reflects the current merged truth. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
An ambiguous question gets one plausible interpretation of several, and the wrong number ships to a board deck. Mitigation: Ambiguity detection forces a clarifying choice — assumptions are never silent, ever.
The warehouse schema changes under the layer; queries succeed and mean something different. Mitigation: Contract tests between layer and warehouse run on every deploy; drift pauses the copilot automatically.
Teams cite copilot answers in board materials with no audit trail. Mitigation: Every answer carries definition version + timestamp; exports include full lineage by default.
Opsmarshal builds workflows like this one inside client stacks — see GTM stack architecture and AI enablement and agentic workflows, or book a call to walk through your own.
Revenue Operations and AI Agentic Workflows for B2B companies in transformation. We build the machine. You run the company.