AI Tools, Policies, And Approvals

Why tools make the AI agentic

Without tools, AI can only produce text. Tools let it search approved knowledge, look up orders, read customer context, create a support ticket, update explicit customer fields, or transfer a chat to a human.

Built-in tools

• lookup_order: reads order status, items, and shipping data for this organization.
• get_customer_profile: reads approved CRM fields needed for the response.
• search_knowledge_base: searches reviewed documents and approved learned resolutions.
• transfer_to_human: transfers the current conversation to a human team.
• create_support_ticket: creates or returns an open support ticket for follow-up.
• update_customer_profile: updates only fields the customer explicitly supplied or corrected.

Policy controls

For every tool you can configure:

• Enable tool: makes it available to the AI.
• Always require approval: pauses execution until an authorized manager approves.
• Calls per minute: prevents runaway or repeated calls.
• Allowed channels: limits the tool to selected WhatsApp, Facebook, Instagram, or other channels. Leaving it empty allows all channels.

Approval lifecycle

1. AI requests a sensitive action with redacted arguments.
2. The request appears in the pending approvals area.
3. A manager reviews the tool, customer context, and displayed values.
4. Approving executes the action once; denying prevents it.
5. The original full arguments remain encrypted at rest.
6. A background job sends the customer a concise result after successful execution.
7. Failed customer responses appear in Operations and can be retried safely.

Recommended policy

• Keep approval mandatory for transfers, ticket creation, and profile updates.
• Keep lookup and knowledge search read-only.
• Start with 30–60 calls per minute per lookup tool and lower limits for actions.
• Restrict action tools to the channels where the operating team is ready.
• Review the immutable audit trail after policy or permission changes.

Custom tools

Use Custom Tool Builder only for public HTTPS endpoints. Do not point tools at private IPs, localhost, credential-bearing URLs, or endpoints that accept unvalidated customer instructions.

Detailed Operating Playbook

Primary owner: AI manager with operations and security review.

Team workflow

• Enable only required tools.
• Require approval for every side-effect action.
• Set per-minute and channel policies.
• Review pending approvals and immutable audit events.

Success signals

• Read-only lookups return verified data.
• Actions do not execute before approval.
• Approved actions execute only once.
• Failed customer responses can be retried from Operations.

When to review or escalate

Review this workflow when channels, policies, or workload change. Escalate to a manager when numbers look inconsistent or the same issue repeats more than once in the same day.