Turn customer service policies into clear agent actions

What belongs in an AI customer service operating policy?

Request and scope

Name the workflow, products, customer groups, and channels it covers.

Required context

List the information and identity checks needed before a decision.

Eligibility rules

Define timing, order status, exclusions, and any monetary limits.

Permitted action

Specify exactly which change the agent may make.

Approval and exceptions

State who approves exceptions and what the customer should be told while waiting.

Success evidence

Define how to confirm the action succeeded in the connected system.

Failure handling

Explain what happens if the request cannot be completed, including how to avoid duplicate changes.

Owner and version

Assign a person to maintain the policy and record its effective date.

Example policy: a delivery address change

This example is a starting point for discussion, not a default policy for your store.

The important boundary is between a customer’s request and permission to act. A request provides intent; the policy and order context determine the next step.

Copy this policy worksheet for your team

Complete one worksheet per workflow. Use explicit values for windows and limits instead of phrases such as “recently purchased” or “a small refund.” Keep the completed worksheet alongside the source policy it implements.

Workflow name:
Policy owner and effective date:
Source policy:
Customer request and scope:
Required identity and order checks:
Eligibility rules and exclusions:
Permitted action and limits:
When approval is required:
Exception owner and next step:
Evidence of successful completion:
Failure and retry behavior:
Customer confirmation:
Test cases and expected outcomes:

How do you test whether the agent follows the policy?

For each workflow, write a small set of cases before enabling it. Include an eligible request, an ineligible request, an ambiguous request, and a connected-system failure. State the expected action and the expected customer message separately.

For an exchange, an eligible order with stock available should produce a different result from an excluded item or an unavailable replacement. A system timeout should not produce a success message. Review both the conversation and the final record in the system that owns the transaction.

When a policy changes, rerun the relevant cases. A new return window or an updated exclusion can change the correct decision even if the customer’s wording is identical.

How does Concierge use operating policies?

Concierge answers customer inquiries and makes the needed changes according to your operating policies. The current pilot applies policy checks to supported email workflows, with human review for exceptions and owner approval for eligible unfulfilled cancellations.

During early-access onboarding, bring your current policies and the workflows you want to cover. Review the Shopify workflow examples and the AI agent evaluation guide to plan the discussion. Confirm the specific integrations and action permissions before enabling a workflow.