Talking to your CRM
How the Voice MCP agent turns a spoken sentence into a CRM action — keyword fast-path, Haiku intent classification, spoken clarifying questions, and a confirm-then-commit executor.
When you speak to Voice MCP, the agent turns your sentence into a CRM intent through a short, layered pipeline. Cheap checks run first; the language model only runs when they fall through.
The routing pipeline
- 1
Keyword fast-path
Each utterance is first matched against keyword patterns for the common intents (deals, pipeline, and so on). These run instantly and for free, so a direct question like "how's my pipeline" never needs the model.
- 2
Intent classification
If no keyword matches, the utterance goes to an intent router backed by a single Haiku call. The model returns a structured result naming a tool plus a query spec.
- 3
Clarify when unsure
The router carries a confidence score. Below a 0.7 confidence threshold — or when the model explicitly flags the utterance as ambiguous — the agent asks a spoken clarifying question instead of guessing.
- 4
Fall back to ask_workspace
If routing produces no usable intent, the agent falls back to
ask_workspace, the catch-all that answers open questions against your workspace.
Why it asks you to repeat sometimes
A clarifying question isn't a failure — it's the agent declining to act on a low-confidence guess. It would rather confirm which "Acme" you mean than pull the wrong record.
Taking action, not just answering
Beyond read-only questions, the agent can perform actions — logging a note, sending a follow-up SMS or email, or booking a meeting. Actions run through a confirm-then-commit flow: the agent states what it's about to do and waits for your spoken confirmation before it commits the change.
Confirm before commit
Side-effectful actions are never fire-and-forget. The executor holds the pending action, reads it back to you, and only writes to the CRM once you confirm.
Meta questions
The agent also answers questions about itself and your workspace through a dedicated meta-query module:
- What can you do? — an introspection summary of its capabilities.
- What happened recently? — recent audit history.
- Who handles this? — routing and ownership questions.
- Knowledge lookups — answers pulled from your knowledge base.