Human Approval Rules
Purpose: This is the master gate for everything the agent produces. It defines the one distinction that governs the whole system (draft versus send), and names who approves what. When any other document conflicts with this one, this one wins. Dave owns it personally.
The one rule that matters
The agent drafts. Humans send.
The agent may produce any text, summary, or analysis the Approved AI Use Cases list allows. But nothing the agent produces reaches a customer, a prospect, the public, or the CRM until a named human has reviewed it and taken the action themselves. There are no exceptions, no thresholds under which the agent may act alone, and no "low-risk" category that skips review. This posture is a feature Summit sells on, not a limitation to engineer around.
Draft versus send, precisely
| Draft (agent does) | Send (human does) | |
|---|---|---|
| Follow-up text or email | Writes it | Rep reads, edits if needed, hits send |
| Review response | Writes it | Jenna posts it |
| CRM stage change | Recommends it with reasoning | Rosa or the rep moves the card |
| GBP post | Writes it | Jenna publishes |
| Customer schedule update | Drafts the message | Pete or Rosa confirms the date first, then sends |
| Price for a customer | States the vault's range only | Tara quotes after site visit or photo review |
If an action is not clearly a draft, treat it as a send and route it to a human.
Who approves what
| Output | Approver | Backup |
|---|---|---|
| Sales messages (follow-up, reactivation, no-show) | Marcus Webb | Dave |
| Intake, scheduling, and customer-update messages | Rosa Delgado | Marcus |
| Anything with scheduling or crew commitments | Pete Sandoval | Dave |
| Marketing content, reviews, ads, GBP | Jenna Fields | Dave |
| Estimates, pricing language, discounts over 5% | Tara Nguyen | Dave |
| Complaints beyond a first acknowledgment | Dave Kowalski | none; Dave decides |
| Anything touching warranty, legal, or insurance wording | Dave Kowalski | none; Dave decides |
The full chain for non-agent decisions lives in Approval Rules and the Escalation Matrix; this table is the agent-specific slice of it.
Approval mechanics
- The agent finishes a draft and labels it: what it is, who the approver is, and what source docs it drew from.
- The approver reads the whole draft, not just the first line. Editing is normal and expected; the agent's draft is a starting point.
- The approver takes the action in the real system (GHL, GBP, email) themselves.
- If the approver is out, the backup approves. If both are out, the draft waits. A waiting draft is never an emergency; emergencies go through the Emergency Job SOP phone process, which is human end to end.
Special cases
- Emergencies: the agent can help Rosa or the answering service find triage info fast, but the $450 dispatch-fee conversation and the dispatch decision itself are always human.
- Discounts: the agent may state the rule (over 5% needs Tara or Dave) but may never approve, imply approval, or draft a message that promises a discount that has not been approved.
- Anything the agent is unsure about: unsure means no. The agent routes it to the approver with a note rather than guessing.
What approvers watch for
- Firm prices or dollar figures outside the vault's published ranges
- Promised dates the crews have not confirmed with Pete
- Warranty, legal, or insurance-coverage language of any kind
- Tone drift from the Brand Voice Guide
- Facts that do not match the vault (flag these to Rosa so the vault gets fixed)
Related
Want documentation like this for your business?
Every Company Brain engagement produces documents at this level of detail — captured from your team, organized, and wired into an agent.