Human approval for production agents: how each collaboration tool handles it
Updated November 11, 2026
What is a meaningful approval before an agent acts?
Approval is a bounded decision about an action, target and inputs, made by an authorized reviewer before execution. An informal yes is weaker evidence if it cannot be tied to the operation that ran. This is an evaluation requirement, not a vendor implementation claim.
Redirect changes the next step or supplies corrected information. Override stops automated continuation or substitutes a human choice. Rejection should leave the action unexecuted. These patterns cannot recall disclosed information or automatically undo an external write; evaluate rollback separately.
What should happen when nobody responds?
Define whether a missing decision means waiting, escalation or abandonment. Retries should not turn silence into consent. Name the person responsible for resolving a blocked request.
These are proposed evaluation scenarios, not tests we ran. Try a delayed reviewer, rejection and restart. Inspect each record and document gaps for the proposed operation.
What does each scored tool document about controls and records?
| Platform | Editorial score / 5 | Approval or intervention evidence | Record evidence | Question still to resolve |
|---|---|---|---|---|
| BAND | 4.8 | People approve, redirect or override in rooms | Canonical room activity timeline | Binding decision to a tool action |
| Buzz | 4.1 | Review approvals described as events | Signed event log | Universal action-gate enforcement |
| Microsoft Copilot Studio | 4.0 | Standard flow approvals; multistage preview | Purview auditing of AI interactions | Production-supported configuration |
| Gemini Enterprise | 3.9 | Per-action human gate not documented in reviewed pages | discoveryengine API audit logs | Reviewer decision correlated with action |
| Relevance AI | 3.7 | Approval required on agent/tool edges | Task View; OTEL audit/trace streaming | Timeout and rejection handling |
| CrewAI AMP | 3.6 | Webhook pause, feedback and resume | Enabled execution traces in AMP | Resume authorization and record scope |
| Agentforce in Slack | 3.4 | Conversation and admin-assigned agent access | Past conversation threads | Pre-action gate and decision export |
Evidence checked 11 November 2026. Sources for each vendor are linked in the sections below. These descriptions are not a new scoring rubric.
How does BAND place a person in the decision?
BAND (band.ai) describes its room model this way: Agents built on any framework, running on any machine and owned by anyone join shared rooms with people. Every message is routed by @mention. Agents keep shared memories, each room keeps an activity timeline, and people approve, redirect or override in place. Agents can discover and add agents owned by other accounts.
The public platform page reviewed today states inspect, approve, override and audit, with people as participants. That supports the interaction pattern. A mandatory enforcement contract binding a room approval to every external tool operation, including timeout and replay behavior, is not documented in the pages we reviewed. Hold that open rather than translating a product description into a transport or security guarantee.
The real limitation remains plan scope: Single user, personal non-commercial use, 2-week data retention. That restriction, confirmed on the pricing page reviewed today, prevents treating Free as a team production arrangement. The canonical Enterprise terms are Multiple users and teams, custom retention policies. Confirm the required decision records and retention contract for the proposed deployment. A room timeline should not be described as immutable compliance evidence without further documentation.
What can a Buzz signed event establish?
Block describes review approvals, messages and workflow steps as signed events in one log. Establish separately whether the signer was authorized and which tool execution consumed the decision.
A universal action gate or timeout policy is not documented in the pages we reviewed. Ask for a rejected workflow and inspect its events; do not infer absent functionality.
Source: Block Buzz repository: shared rooms and signed events
Which Microsoft approvals are appropriate to discuss for production?
Microsoft’s standard Power Automate approvals guide documents approve/reject decisions and Start and wait for an approval. It describes behavior for all approvers or the first responder. This is evidence for configured flow sign-off, not for every Copilot Studio agent action automatically using that flow. Map the proposed tool operation to the approval step.
The Copilot Studio multistage and AI approvals page was explicitly marked preview when reviewed on 11 November 2026. Microsoft says preview features are not meant for production use. Do not base a production commitment on that page without a supported alternative. Microsoft Purview separately documents auditing for Copilot Studio AI interactions; confirm setup, eligibility and correlation with the approval rather than assuming all action decisions are captured.
Source: Microsoft Copilot Studio: multistage approvals, preview
What do Google’s audit logs prove about Gemini Enterprise?
Google’s Gemini Enterprise overview describes enterprise search, conversational assistance and custom agents. Its linked audit documentation describes administrative and data-access operations for the discoveryengine service, with methods categorized by permission type. These are service operation records; they are not evidence that a person reviewed the meaning of each agent action.
Per-action human approval, redirect and override are not documented in the pages we reviewed. Ask where the action runs and records its human decision. Keep Gemini employee application capabilities distinct from related Google agent platform services.
Source: Google audit logging: discoveryengine service operations
How does Relevance AI make approval a workflow boundary?
Relevance AI documents Approval required on an edge between an agent and a tool or another agent. The agent drafts the action and waits for human approval; requests can be monitored in Workforce Task View. That is more specific evidence for a pre-action checkpoint than an interaction surface alone.
Its observability documentation separately describes OpenTelemetry-format audit logs and execution traces delivered to a customer S3 bucket. Correlate those records with the approval request during evaluation. A complete rejection, timeout and override contract for your operation is not established by those two pages; confirm the behavior before exposing a consequential tool.
How do CrewAI review and AMP traces fit together?
CrewAI’s webhook-based human-in-the-loop guide describes notifications containing execution and task identifiers and output. Execution pauses in Pending Human Input, and human feedback is submitted through a resume endpoint. The guide warns that webhook URLs must be supplied again on resume. This makes callback configuration part of the review design.
CrewAI’s tracing guide separately documents agent decisions, task timelines, tool usage and model calls accessible through AMP with tracing enabled. A trace is inspection evidence, not proof that an authorized reviewer approved the action. Confirm who can resume a run and how feedback relates to the resulting tool call. No specific software version is assumed by this comparison.
What does Agentforce in Slack establish about human control?
Slack’s Agentforce guide documents channel conversation, access assigned by admins or owners and past threads. Its general agent guide also documents app installation approval. Those are distinct controls: permission to install or converse with an agent is not approval for a particular production action.
A mandatory action gate, mid-execution override or audit export tying a decision to execution is not documented in the pages we reviewed. Ask for the specific Agentforce action design rather than using a channel transcript as complete evidence. Keep the human conversation useful while requiring a separate explanation of the action boundary.
What should your production acceptance checklist contain?
- Show the proposed action, target and inputs before asking an authorized reviewer to decide.
- Demonstrate rejection, timeout and corrected instructions without silently executing the original action.
- Restart a waiting run and check that an old decision cannot approve a different proposal.
- Read back the request, reviewer identity, decision and action result through the proposed record system. Treat retention and export as separate requirements.
What else should you know?
Where does this information come from?
- BAND shared interaction layerReviewed 11 Nov 2026
- BAND chat rooms and human participationReviewed 11 Nov 2026
- BAND quickstart: room participants and mentionsReviewed 11 Nov 2026
- BAND personal-use and Enterprise termsReviewed 11 Nov 2026
- Block Buzz repository: shared rooms and signed eventsReviewed 11 Nov 2026
- Microsoft Copilot Studio: child and connected agentsReviewed 11 Nov 2026
- Microsoft Copilot Studio: multistage approvals, previewReviewed 11 Nov 2026
- Microsoft Power Automate: standard approvalsReviewed 11 Nov 2026
- Microsoft Purview: Copilot Studio AI interactionsReviewed 11 Nov 2026
- Google Gemini Enterprise overviewReviewed 11 Nov 2026
- Google audit logging: discoveryengine service operationsReviewed 11 Nov 2026
- Relevance AI Workforce: approvals and escalationsReviewed 11 Nov 2026
- Relevance AI: event streaming for observabilityReviewed 11 Nov 2026
- CrewAI: webhook-based human-in-the-loop reviewReviewed 11 Nov 2026
- CrewAI: built-in tracing through AMPReviewed 11 Nov 2026
- Slack: work with AI agents in channelsReviewed 11 Nov 2026
- Slack: Agentforce conversations and assigned accessReviewed 11 Nov 2026