For AI service providers · Guide
How to handle AI agent approvals for client work
Assign a reviewer and a tested approval rule before allowing an agent to change client systems. For each supported tool, inspect the proposed action, approve or reject it and verify the external result. Keep persistent trust decisions separate from one-time approvals and document what happens when no reviewer is available.
By Agentic Workers Team · Published · Updated
Responsibilities and evidence
Agree on the owner and retain a safe evidence reference for each responsibility.
- Service scope
- Provider drafts; client accepts. Evidence: agreed output and stop conditions.
- Account ownership
- Client and provider confirm owners. Evidence: account inventory without secret values.
- Configuration
- Provider sets up the intended workspace and agent. Evidence: reviewed settings.
- Action review
- Named reviewer checks supported actions. Evidence: request and decision.
- Acceptance
- Client or delegated reviewer checks the result. Evidence: accepted artifact.
- Exit
- Account owner and provider remove access. Evidence: handoff and denied-access check.
1. Build the approval matrix
List the action, client account, reviewer, allowed scope and timeout response. Include external messages, record changes and paid actions relevant to your service. Mark unsupported paths as manual; do not assume an instruction creates an enforced approval barrier.
2. Test all intended decisions
Use a harmless supported write in a test destination. Verify that it waits, that rejection leaves the destination unchanged and that approval executes only the intended action. If you enable persistent trust, test and document its scope separately.
3. Retain enough evidence to review
Record the request, decision, reviewer, resulting run and external result in your operating record. Review trusted actions when scope changes. Ordinary run records are not a claim of immutable audit compliance.
Example provider workflow. Adapt it to the client’s agreed scope.
- 01
Step 1: Build the approval matrix
- 02
Step 2: Test all intended decisions
- 03
Step 3: Retain enough evidence to review
Fit and limitations
- Good fit
- Providers applying this process to an agreed client service.
- Poor fit
- Treating this checklist as an automatic product feature or a substitute for a client agreement.
- Known limits
- These are operating recommendations. Test each supported product path in the target workspace and plan. Examples are not performance, pricing or security benchmarks.
Sources and evidence
Documentation establishes the scope of a claim. It does not replace an acceptance test in your client’s environment.
| Source | What it supports | Checked |
|---|---|---|
| Agent tools and review guidanceProduct documentation | Supported tool actions can require approval. Test the exact tool and runtime; coverage is not universal. | |
| Read an agent run recordProduct documentation | Inspect available tool calls and execution details within the relevant workspace. A completed run is not proof of an accepted result. |
Questions providers ask
Who is responsible for the result?
The provider names an operator; the client names the person who can accept the deliverable. The platform does not replace those responsibilities.
What evidence should we retain?
Keep the source date, workspace and agent, expected output, review decision and artifact reference. Exclude credentials and unnecessary personal data.
When should we stop and review?
Stop when the account, recipient, output or permissions differ from the agreed scope, then resolve the exception before continuing.
Set up one client service
Set up the client workspace, connect its tools and review the first deliverable. Use a tested delivery path, such as Telegram, so the client can receive work in chat. Track operating costs and review effort before pricing a recurring service.
Sign up now