How to Set Up Human Handoff Rules in a Shopify AI Agent

This guide covers the setup work. If you're still deciding which tickets an AI should handle at all, start with what AI support should and shouldn't answer on Shopify, listed at the end of this article. If you're working out how to split the work between your team and the AI, read Shopify AI vs human support, also listed at the end of this article.

Step 1: Write every rule as one plain sentence

Most handoff problems start with a vague instruction like "escalate tricky cases." An AI agent can't act on "tricky." A new team member can't either. Each rule needs a trigger, a limit and an owner.

Use this format: when a trigger happens, the AI does or doesn't do something, and a named person handles it.

Here's what that looks like for a skincare store:

• When a customer mentions a reaction to a product, the AI doesn't answer, and the founder handles it.
• When a refund request is above the amount we approve by hand, the AI drafts nothing, and the ops lead handles it.
• When a wholesale account writes in, the AI doesn't reply, and the account manager handles it.
• When an order edit comes in after the order has shipped, the AI explains the options, and anything outside policy goes to the ops lead.

Short sentences are easy to check. If two people on your team would read a rule differently, rewrite it.

Step 2: Pick your trigger types

Handoff triggers fall into a few groups. You don't need all of them on day one, but decide on each group before launch.

• Topic: what the ticket is about. Example rule to write: "Chargebacks and legal threats always go to the team."
• Value: how much money is involved. Example rule to write: "Refunds above your approval limit need a person."
• Customer: who is writing. Example rule to write: "Wholesale and repeat VIP customers go to a named person."
• Tone: how the customer feels. Example rule to write: "Clear anger or distress goes to the team."
• Certainty: how sure the AI is. Example rule to write: "If the answer isn't in your policies or order data, hand off."
• Explicit request: the customer asks for a person. Example rule to write: decide when this skips the AI entirely.

The explicit request case has its own nuances. Keep that rule simple for now and refine it once you've seen real tickets.

For topic rules, your existing decision trees are a good source. If you already have a process for damaged item claims or out of warranty goodwill requests (both guides are listed at the end of this article), mark the branches where a person has to decide. Those branches are your handoff rules.

Step 3: Set permissions action by action

Triggers decide when a person steps in. Permissions decide what the AI may do on its own before that point. Go through each action and pick one of three levels: the AI does it alone, the AI needs approval, or it always stays with your team.

• Answer policy and order status questions. AI does it alone: usually yes.
• Change a shipping address before fulfillment. AI does it alone: often yes.
• Send a discount code. AI does it alone within your limits. AI needs approval above your limits.
• Cancel an order. AI does it alone before fulfillment. Always your team after fulfillment, if you prefer.
• Issue a refund. AI does it alone below your limit. AI needs approval above your limit. Always your team for disputes and chargebacks.

The levels left open are deliberate. Your answers depend on your margins, your return policy and how much risk you're comfortable with. Start strict. Giving the AI more room after two weeks of clean results is easier than taking it back after a bad refund.

For order changes, our guide to handling order edit requests with AI, listed at the end of this article, covers which edits are safe to automate and which aren't.

Step 4: Decide what your team needs to see

A handoff only saves time if the person picking it up can act without starting over. What gets passed along depends on the tool you use, so treat this as a checklist to test, not something to assume.

On every escalated ticket, your team member should be able to find:

• The customer's message, in their own words.
• The order involved.
• Any reply the AI already sent, or any action it already took.
• Which rule caused the handoff.

During any trial, open a few escalated tickets and check each item yourself. If something is missing, add it to your team's routine, for example "check the order in Shopify before replying." The customer should never have to explain twice.

One more habit worth keeping: when a ticket moves to your team, don't promise the customer a response time. A time you can't keep on a busy Monday does more damage than no time at all. Tell them a team member will reply in the same thread.

Step 5: Test with real tickets before going live

Rules that look clear on paper break on real messages. Before launch, pull a sample of past tickets from your inbox and run them through your rules.

Pick tickets that test the edges:

• A refund request just under your approval limit, and one just over it.
• A polite message that mentions a sensitive topic in passing.
• An angry message about a simple question.
• A wholesale customer asking a routine question.
• A question your policies don't cover.

For each one, write down what you expected and what happened. Every mismatch means a rule needs rewording, a permission needs tightening, or your knowledge base is missing something.

Step 6: Review handoffs every week

Your handoff rules aren't finished at launch. Set aside time each week to read the tickets that went to your team.

Look for three patterns:

• The same question keeps getting handed off. The fix is usually a missing policy or product detail in your knowledge base, not a new rule.
• Tickets went to the team that the AI could have solved. Loosen that rule or raise that limit.
• The AI handled something a person should have seen. Tighten the rule right away and check similar tickets.

If you already run a weekly support routine, add this review to it. Our weekly workflow for Sidekick and a support agent, listed at the end of this article, shows one way to fit it in.

Where these rules live in Relay

Relay answers, refunds and resolves your Shopify tickets 24/7, following your rules. The rules you wrote in this guide are the ones to give it.

The homepage describes the four parts of how Relay works: Understand, Decide, Act, Escalate. Relay is an AI support agent that "knows when to bring in your team," and you decide where that line sits.

Setup takes three steps. Use them to apply the rules you wrote above:

• Teach Relay Your Business
• Set Your Rules
• Connect and Go Live

Run your test tickets from step 5 before the last one. Relay's homepage says You Stay In Control, with Permissions. Use your weekly review to decide which rules to tighten or loosen.

Setup mistakes to avoid

• Writing rules about feelings instead of signals. "Escalate upset customers" is vague. "Escalate when the customer threatens a review, a chargeback or legal action" can be checked.
• Starting with loose permissions. Open up step by step, based on what you've reviewed.
• Giving every handoff to "the team." If nobody owns a rule, those tickets sit unanswered.
• Never updating the knowledge base. Many repeat handoffs disappear once the missing answer is written down.
• Skipping the test run. Ten past tickets will tell you more than an hour of rereading your rules.

FAQ

How many handoff rules should I start with? Only as many as you need to cover your sensitive topics, your money limits and your key customers. Add rules based on what your weekly review shows, not on guesses.

Should refunds always go to a person? Not necessarily. Many stores let the AI handle refunds under a set amount that fit the policy, and send larger refunds or disputes for approval. Set the limit based on your margins and your return policy.

What if a ticket matches no rule and the AI isn't sure? Hand it off. A clear "when in doubt, escalate" rule is safer than letting the AI guess, and those tickets show you where your knowledge base has gaps.

How do I know my handoff rules are working? Read escalated tickets every week. Good rules send your team the tickets that need judgment, and very few that the AI could have solved with better information.

Want your handoff rules, permissions and approval limits in place before your next busy season?