By The D2C Agent Desk team · Updated October 2026

Gorgias AI Agent setup tips: a calm, draft-first rollout for Shopify brands

Most Gorgias AI Agent problems are not model problems. They are knowledge, scope and handover problems. Here is the order we set things up in, and why.

Gorgias AI Agent can answer and resolve a meaningful share of routine Shopify support tickets, but only when it is fed clean knowledge, given a narrow scope at first, and told clearly when to step aside. Brands that switch it on across every channel on day one usually spend the next month apologising for confident wrong answers. Brands that roll it out topic by topic tend to keep it.

This guide assumes you already run Gorgias on Shopify and want to turn on (or reset) AI Agent. Menu names below follow Gorgias's own help docs at the time of writing; Gorgias ships changes often, so check their documentation if a screen looks different.

1. Start with your real ticket mix, not the template list

Before touching settings, export or filter the last 30 days of tickets and tag the top reasons customers write in. For most D2C stores the list looks something like this:

• Where is my order (WISMO) and delivery delays
• Returns, exchanges and refund status
• Order changes: address edits, cancellations, adding an item
• Product questions: sizing, ingredients, materials, compatibility
• Subscription changes (skip, pause, cancel) if you use Recharge or similar
• Discount codes that did not apply

Rank them by volume and by risk. WISMO is high volume and low risk, which makes it the classic first topic. Ingredient or allergy questions may be lower volume but higher risk, so they should go to a human until your written knowledge on them is airtight.

2. Fix the knowledge before you blame the AI

AI Agent answers from the knowledge you give it: help center articles, website content, documents and guidance. If your shipping page says "3–5 days" and your returns FAQ says "ships within a week", the agent will eventually quote one of them to the wrong person.

Do a one-hour cleanup pass:

• Make one canonical page each for shipping, returns, exchanges and warranty, and delete or redirect duplicates.
• Write numbers explicitly: cut-off times, processing days, the return window, who pays return shipping, and which countries you ship to.
• Add product-level facts (materials, care, sizing) to product pages or a document, not just to agents' heads.
• Remove promotional language from policy pages. "Lightning fast shipping" is not a fact the agent can use.

3. Use guidance for single-topic rules

Gorgias recommends that each piece of guidance covers one topic so the agent can pick the most relevant one. Good guidance reads like a short internal SOP:

• When it applies: "Customer asks whether a product is safe for sensitive skin."
• What to say: the facts, in your words, with the link to the relevant page.
• What not to say: "Never give medical advice; suggest a patch test and point to the full ingredient list."
• When to stop: "If the customer mentions a reaction, hand over."

Before publishing, make sure the guidance is toggled to be in use by AI Agent and test it with a message it should catch.

4. Add skills and actions only where the procedure is boring and safe

Skills are step-by-step procedures, and actions let AI Agent do things in connected apps, such as cancelling a Shopify order, updating a shipping address or sending a returns portal link. Gorgias notes the agent can only run an action that is included in a skill or guidance, and that you do not need actions at all to go live.

Our rule of thumb: add an action only when (a) the policy is completely unambiguous, (b) the action is reversible or low-cost if wrong, and (c) you have tested it in the Playground against edge cases. "Edit shipping address before fulfillment" usually passes. "Issue refund" usually waits until you have weeks of clean data.

5. Design handover before go-live

Gorgias's docs describe three reasons AI Agent hands over: built-in conversation signals (anger or frustration, an explicit request for a human, sensitive topics such as legal threats), your configured handover topics, and cases where it cannot generate a reliable answer. You can configure the second group under AI Agent > Settings > Handover and exclusion.

Handover topics we commonly start with for Shopify brands:

• Chargebacks, disputes and "I'll contact my bank"
• Damaged, unsafe or allergy-related product reports
• Wholesale, press, influencer and partnership requests
• Orders above a value you choose, or customers you tag as VIP
• Anything involving a refund outside your written policy

Write each topic the way a customer would say it ("my package arrived broken", "I want to talk to a manager"), because that is how the agent recognises it. Gorgias also suggests keeping handover instructions minimal at the start; we agree, with the exception of anything safety- or money-related, which should be on the list from day one.

For email and SMS there is a toggle to tell customers when they are being handed to a person. We usually turn it on: a short "I'm passing this to our team" sets expectations and avoids a second "hello??" email.

6. Test in the Playground like a mystery shopper

The Playground lets you run test conversations by channel and audience (for example a new customer versus one with order history). Gorgias states that test conversations are not billed, do not send real messages and do not affect reporting, so use them generously.

Build a test set of 25–40 real (anonymised) customer messages, including awkward ones: typos, two questions in one email, a polite message hiding a refund demand, a customer who gives the wrong order number. For each answer, mark it pass, edit or fail, and use Show reasoning to see which knowledge it relied on. Fix knowledge first, guidance second, and only then consider new skills.

7. Go live narrow, on one channel

A sensible first live scope is WISMO plus one or two FAQ topics on email or chat only. Keep everything else with your team. Then expand weekly:

1. Week 1: one channel, two or three topics. Review every AI-handled ticket daily.
2. Week 2: add topics that passed testing. Review a sample (say, 20 tickets a day).
3. Week 3–4: consider a second channel and your first low-risk action.

8. Measure more than automation rate

Automation rate on its own can be gamed by customers giving up. Track these alongside it every week:

• Handover reasons, top five, with example tickets. These are your knowledge to-do list.
• Reopen rate: AI-closed tickets the customer replied to again within a few days.
• CSAT on AI-handled tickets versus human-handled ones.
• Wrong-answer log: every factual error, the source it came from and the fix.

A weekly one-page digest with those four numbers is usually enough for a founder to decide whether to widen the scope. We cover the format in our weekly KPI digest guide.

Common mistakes we see

• Turning on every channel at once, so problems are hard to trace.
• Duplicated or outdated policy pages feeding contradictory answers.
• A persona/tone so cheerful it reads badly when a parcel is lost.
• No owner. Someone on the team has to read the handover reasons every week.

If you already have AI Agent running and the numbers have stalled, our Gorgias AI Agent audit checklist walks through how to find what is holding it back.

— The D2C Agent Desk team