Guide, 10 Nov 2026, read 10 Nov 2026
Approval and audit for deal agents: what each vendor documents
Three questions that separate vendors
Every vendor says its agents are safe. Three questions test the claim. Does a person approve a write before it lands? Does the agent act with that person's permissions, or with a broader service account? Can you see afterward what the agent changed? A page that answers all three is rare, and a page that answers none is common.
Rox's own guide gives a usable rule. It says approval should be required where an error reaches a customer, changes a system of record, or commits money or terms, and that low-risk research can run alone with sampling. Hold every vendor in the table below to that rule.
What the vendors document
| Vendor | Approval as documented | Source type |
|---|---|---|
| Aligned | Seller Agent applies changes on seller approval; confirmation for irreversible or removing actions and anything that notifies the buyer side; adding content needs none | Vendor help center |
| Attention | New organizations start with every Super Agent action off; an enabled write action shows an Approval required card with a preview | Vendor docs |
| Day AI | Every CRM field update held for review; the CRM agent stops before closed-won and closed-lost | Vendor pages |
| Dock | Every write action through the MCP has an approval step; the MCP is in beta | Vendor MCP page |
| GetAccept | The AI asks permission and the seller allows or denies before a change | Vendor MCP page |
| Gong | Governed harness with permissions and oversight; the page describes no per-action seller approval step | Vendor page |
| HubSpot | Users can approve suggested CRM updates and AI-written follow-ups | Vendor help center |
| Momentum | None found on 10 Nov 2026 for an approval step before an agent writes to the CRM | Vendor pages checked |
| Rox | Most agent actions require user confirmation | Vendor docs |
| Salesforce | Critical write actions need human confirmation; suggestive mode lets sellers approve field updates | Vendor pages |
| Sybill | For connectors an admin has selected, tools run with no approval prompt; control sits at connector selection | Vendor release notes |
Reading the table
Attention's default is the strongest posture in the table: off for every role until an admin turns an action on. Sybill states the opposite design openly. Once an admin approves an MCP connector, the tools run automatically, so the approval is one decision at setup, not one per action. Neither is wrong. They suit different teams.
Gong is a case of what a page leaves open. Gong says its Deal Reviewer suggests deal updates and automatically updates the CRM, and the page does not say whether a seller must approve the write. Gong's help center separately says the Deal Reviewer drafts notes that the account executive reviews. Ask Gong which setting applies to your account.
Permissions travel with the person
Several vendors tie the agent to the user. Attention says its MCP server enforces the user's existing permissions. Rox says its MCP server runs actions with the signed-in user's permissions and needs no API key. Day AI says each team member connects Salesforce or HubSpot with their own login, so the agent works within that person's CRM permissions. Salesforce says agents inherit field-level security, sharing rules and audit trails.
Aligned says Seller Agent files and conversations are not used for model training. Aligned's help center also says write skills through its MCP need the mcp:write permission and are confirmed by the assistant.
Audit: the weakest part of most pages
Few pages describe a log you can read. Day AI says the audit trail records what ran, and that record changes made through its MCP are tracked and audited. Salesforce says actions that reach a customer are logged. Rox says its agent shows its sources, data pulled and actions queued, and tags retrieved data so users can trace conclusions to sources.
Aligned's help center says clients never see the Seller Agent and no marker shows that AI made a change. That keeps the buyer view clean, and it also means a reviewer cannot tell from the room alone which edits came from the agent. Ask for the in-app history before you rely on it.
A short test for your trial
- Turn the agent on for one rep and one deal, and write down every action it takes in a week.
- Try an action that should need approval, such as notifying a buyer, and confirm it stops.
- Check whose permissions the action ran under.
- Find the place where the change is logged, and read a week of it.
- Confirm what happens to your data: Aligned, for one, says its Seller Agent files and conversations are not used for training.
Where the index stands
The criterion AI that acts on the deal carries 35 of 100 weight points on this site, and approval is part of what we read for it. See the method for how a vendor claim is capped and the platform pages for each vendor's evidence.
Sources
- Rox, human in the loop guide, read 10 Nov 2026
- Aligned help center, AI Seller Agent, read 10 Nov 2026
- Attention docs, access and permissions, read 10 Nov 2026
- Day AI, controlling agent CRM access, read 10 Nov 2026
- Day AI, CRM auto-updates, read 10 Nov 2026
- Dock, MCP, read 10 Nov 2026
- GetAccept, MCP, read 10 Nov 2026
- Gong, AI agents, read 10 Nov 2026
- Gong, revenue AI, read 10 Nov 2026
- Gong help center, AI Deal Reviewer, read 10 Nov 2026
- HubSpot knowledge base, deal progression, read 10 Nov 2026
- Momentum, approvals, read 10 Nov 2026
- Rox docs, human-agent interaction, read 10 Nov 2026
- Rox docs, AI safety, read 10 Nov 2026
- Rox docs, MCP, read 10 Nov 2026
- Salesforce, headless sales, read 10 Nov 2026
- Salesforce, sales AI, read 10 Nov 2026
- Salesforce, sales agent, read 10 Nov 2026
- Sybill, release notes, read 10 Nov 2026
- Day AI, MCP, read 10 Nov 2026
- Attention docs, MCP authentication, read 10 Nov 2026
- Aligned help center, MCP capabilities, read 10 Nov 2026
Last reviewed 10 Nov 2026.