Define the job before enabling AI Agent
A useful launch plan specifies what AI Agent can answer, what it can change, and what it must hand to a person. Uploading an FAQ addresses only the first part. Small support teams also need an owner for escalated conversations and a way to judge the first real responses.
The current requirements include an active AI Agent subscription, a connected Shopify store, and a Lead, Admin, or Account Owner role. A Helpdesk subscription alone doesn't establish AI entitlement. Official setup requirements
We checked the official documentation on September 19, 2026. This guide focuses on launch scope and acceptance testing rather than account creation or inbox migration. It isn't a report of a deployment we have run. The pouch-store examples below are fictional planning examples.
Our recommended first goal is narrow: answer routine delivery-policy questions correctly and route exceptions to the team. That is a manageable evaluation target, not a promised automation rate or implementation timeline.
Separate reference knowledge, procedures, and real operations
Knowledge describes the business; Skills describe a process; Actions perform operations in connected apps. Explaining a cancellation policy is different from canceling an order. Keeping that distinction explicit helps you identify whether an error came from the source material, the procedure, or a permitted operation. AI Agent structure
| Component | Input | Acceptance question |
|---|---|---|
| Knowledge | Shipping regions and return conditions | Do the sources agree? |
| Skills | A procedure for delivery questions | Does it handle missing details appropriately? |
| Actions | Permitted order operations | Are allowed changes tightly defined? |
| Handover | Exceptions requiring a person | Can the team receive and handle them? |
Actions aren't required to launch. Consider proving the informational workflow before adding order changes. Human-facing response templates remain useful, but creating a Gorgias macro is not equivalent to authorizing AI operations.
Make the knowledge consistent before making it comprehensive
Check the sync status, not just the upload: This official Knowledge example shows a website sync still in progress. The banner warns that AI Agent may not have the latest content yet. Wait for completion, review what was imported, and then test your questions. Starting a sync is not the same as making the answers ready.
Start by resolving conflicting policies. An expired free-shipping promotion can undermine an otherwise correct delivery FAQ. Identify the current source for each topic and remove outdated instructions from the material AI Agent uses.
Supported knowledge includes help articles, website content, individual URLs, documents, and Guidance. A single-page URL doesn't automatically include its child pages or embedded media; uploaded-document images also aren't processed. Official knowledge guide
For our fictional store, split dispatch time from carrier transit time. A policy of dispatching within two business days doesn't promise delivery everywhere within two days. State the region, qualifying conditions, and exceptions alongside the rule. Replace that illustrative timing with your actual policy.
Check which sources are enabled after adding them. A document's presence doesn't prove every policy is available in usable text. Avoid including customer exports or unrelated internal personal information. Assign a policy owner so a wrong response can lead to a specific correction rather than an unstructured rewrite of the entire FAQ.
Build acceptance tests that include failure paths
Test ordinary questions, boundary conditions, and requests for a person. Playground conversations normally don't reach shoppers and aren't billable. However, enabling Actions can affect real customer or order data. Leave Actions disabled for initial conversational tests. Playground behavior
| Question | Expected behavior |
|---|---|
| When will you dispatch my order? | Use the current dispatch policy |
| Can you guarantee delivery tomorrow to a remote address? | Avoid an unsupported guarantee |
| I need an exception to your return conditions | Follow the defined escalation path |
| I want a person | Recognize the handoff request |
| Change this order for me | Don't make an operation outside the enabled scope |
Repeat tests using incomplete sentences, alternative wording, and two requests in one message. For each intended language, check that the conditions retain their meaning. Equal quality across all languages remains unverified for your store until tested.
Record the question, response, cited knowledge, expected outcome, and correction. Re-run affected tests after a change. If you later test real operations, plan their data and side effects separately with an authorized administrator; don't assume the Playground label makes them harmless.
Prove that handoffs reach someone
An escalation message is only the customer-facing half of a handoff. Your team needs a monitored destination and a realistic response expectation. Gorgias distinguishes configured handover topics from rules that prevent AI Agent from engaging with specified tickets at all. Handover and exclusion documentation
For returns, distinguish explaining the normal policy from approving an exception. Describing the exception precisely is more useful than sending every mention of returns to a person. This is an editorial scoping recommendation, not a required Gorgias configuration.
Check chat behavior both during and outside staffed hours. A customer should not receive an immediate-response promise when the team is unavailable. Email and chat have different handover experiences, so test the actual channel you will use.
Name the person responsible for the handover queue. In the final channel check, verify the customer message and the team's received ticket, including the information needed to continue the conversation. Our tags and rules guide covers routing mechanics separately.
Confirm the subscription and launch only the intended scope
Don't treat the Helpdesk entry price as an all-inclusive AI quote. Ask about AI allowance, overage, enabled channels, billing period, and the trial offered to your account. Use the same configuration when comparing quotes. Official pricing and our pricing guide explain the distinction.
Before enabling a channel, confirm the selected store, intended audience, knowledge, and staffed handover route. Assign someone to review early conversations and decide when the rollout should pause. A narrower start makes errors easier to trace and correct.
The current pricing page promotes AI Agent across plans while describing charges for resolved conversations. The setup guide still requires an active AI Agent subscription. Confirm the account's entitlement and quote: Starter's $10 monthly Helpdesk entry price is not an unlimited AI budget, and the applicable AI-inclusive minimum remains unverified. Official billing explanation
Questions to take into the demo
Can we start with a small FAQ?
At least one knowledge source is required. Whether the available content is enough for your intended workload is a separate question. Use representative test outcomes rather than document count to decide.
Is Playground completely isolated from real orders?
Not when Actions are enabled. Separate response testing from operation testing and keep Actions off in the initial phase.
What should we prepare first?
Bring three target conversation types, their current policies, mandatory human exceptions, and a few test questions. That gives the demo a concrete purpose: establish whether the available controls and price fit your team's intended scope.
What should we check about the trial and cancellation?
The creation guide describes 30 days and the Shopify listing shows seven. Check your signup route for duration, card requirements and charging after the trial. Trial signup is available through the official site. Paid plans renew automatically; stop renewal through Gorgias Billing and usage, or the customer success manager for managed contracts. Removing the Shopify app does not cancel the subscription. Account creation and cancellation
Can we deploy Japanese answers immediately?
Japanese AI accuracy for your policies remains untested. Evaluate dispatch versus delivery, return exceptions and human handoffs separately from widget localization. Full Japanese admin UI coverage, Japanese vendor support and JPY billing are unverified. Our Japanese support guide provides a test plan; our Gorgias review covers public ratings and reviewer concerns. Bring the requirements to a demo before deciding your launch scope.
See whether Gorgias fits your businessBook a demoA free trial is availableAsk about the trial length and included features during your demoRequest a Gorgias demo
