Build Your Own Support Skills: Automate Any Support Scenario with Custom Support Skills

What are Custom Support Skills?

Rep AI comes with a set of pre-built Support Skills — Cancel Order, Order Status, Returns & Exchanges, Damaged Order, and more. They cover the most common requests your shoppers make, but not the ones unique to your business.

Custom Support Skills let you build your own. You describe a support scenario, write the steps you want the AI to follow, choose what the AI is allowed to do, and — optionally — connect your own APIs so the AI can look things up in your systems mid-conversation.

Think warranty activation, loyalty points disputes, gift card problems, subscription billing questions, product setup help. Anything your team handles with a repeatable playbook, your AI can now handle too, instead of escalating to a human.

Who can access Custom Support Skills?

Custom Support Skills require Expert Zone access on your plan.

• If your plan includes Expert Zone, you'll see an Add new skill card on the Support Skills page.

• If it doesn't, the card appears locked with a padlock. Account admins get a tooltip pointing to the plans page; other roles are told to ask their account admin to upgrade.

How Custom Support Skills work

When a shopper sends a message, your AI decides which skill should handle it. Custom skills join that decision automatically:

1. The AI matches intent using your skill's description. This is why the description matters — it's what the AI reads to decide "this message is a warranty claim."

2. Pre-built skills always win a tie. If a message could belong to both a pre-built skill and a custom one, the pre-built skill handles it. Custom skills fill the gaps, they never override the built-ins.

3. Your policy becomes the AI's playbook. The AI follows your written steps in order, collects the information you asked for, and resolves the issue the way you described.

4. Permissions decide what the AI can actually do. Your policy says when to refund; the permissions say whether the AI is allowed to. If your policy calls for an action you haven't permitted, the AI skips it and takes the next best path instead of failing.

5. Escalating to a human is always available. It can't be switched off, so a shopper never hits a dead end.

Before you start

A few limits to plan around:

Creating a custom support skill

Go to AI Support → Support Skills and click Add new skill. The card shows how many you've used, for example "3 of 10 custom skills."

Start from a template, or from scratch

You'll first choose a starting point. Six templates come ready to edit:

• Warranty Claim — warranty activation requests with proof validation

• Size Exchange — help shoppers swap for a different size

• Loyalty Program Issue — points balance, tier status, reward redemption

• Billing Dispute — subscription billing questions and payment disputes

• Product Setup Help — guide shoppers through assembly or configuration

• Gift Card Issue — balance, activation, and redemption problems

A template pre-fills the name, description, icon, information fields, permissions, and a starter policy — all of it editable. Templates don't pre-fill API endpoints; you add those yourself if you need them.

Prefer to build your own? Choose Start from scratch.

Step 1 — Name

• Skill name — what you'll see on the Support Skills page, and what appears in your analytics.

• Description — the single most important field. Write it as a description of the shopper's situation, not a description of the feature. The AI reads this to decide when to use the skill.

• Icon — pick one for the card.

Good description: "A customer's product became defective after use and they want to activate their warranty. They need to provide proof of purchase and describe the defect."

Weak description: "Warranty stuff."

Step 2 — Permissions and API endpoints

Action permissions control what the AI may do on this skill:

• Issue Refund — requires shopper confirmation before executing

• Cancel Order — cancel the shopper's order

• Modify Order — change details on the shopper's order

• Escalate to Human Agent — always on and locked, as a safety net

Leave a box unchecked and the AI simply can't take that action in this skill — even if your policy mentions it. Order actions use the same safety rails and shopper-confirmation steps as the pre-built skills.

API endpoints let the AI pull live data from your systems during the conversation. Click Add API endpoint and fill in:

• Endpoint name — how you'll refer to it in your policy, for example "Warranty Check"

• Description — helps the AI understand when calling it is appropriate

• URL — must be a publicly reachable http or https address

• Method — GET, POST, or PUT

• Authentication — None, Bearer token, or Basic auth

• Custom headers — optional, entered as JSON

• Test endpoint — sends a real request and reports the status code and response size, so you can confirm it works before you save

Your credentials are encrypted when stored and always shown masked afterwards (for example ****cde1). The AI can only call the endpoints you've configured here — it can't invent or guess a URL.

Requirements and limits for your endpoints:

• Public addresses only. Internal addresses, private IP ranges, and localhost are rejected for security.

• GET, POST, and PUT only.

• Each call times out after 10 seconds.

• Responses are read up to 50 KB.

• Information collected during the conversation is sent in the request body, which applies to POST and PUT. If your endpoint needs a value like an order number, use POST or PUT rather than GET.

Step 3 — Fields

Tell the AI what to collect from the shopper before it resolves anything. Add a row per field, name it, choose a type, and mark it required or optional:

• Text — free-form answers such as a defect description

• Email — the shopper's email address

• Order number — collected with an order lookup

• Photo — asks the shopper to upload an image

• Dropdown — you supply the choices, and the AI keeps the shopper to that list

Required fields gate the conversation: the AI won't move to a resolution until it has them. If a shopper volunteers something early or out of order, the AI notices and adapts rather than asking again.

Step 4 — Policy

This is where you write the playbook. Treat it like onboarding notes for a new support agent: numbered steps, clear conditions, explicit outcomes.

Above the editor you'll see a reminder of everything available to you — the actions you permitted, the endpoints you configured, and the fields you're collecting. Refer to them by name in your steps ("Call the Warranty Check endpoint with the order number").

A worked example:

Short on time? Click Generate Policy and Rep AI drafts a policy for you based on your skill name, description, permissions, endpoints, and fields. Edit it from there — it's a starting point, not a finished policy.

Step 5 — Review

You'll see everything in one summary, with an edit link back to any step. Click Finish and the skill is saved and goes live on your storefront right away. Editing an existing skill shows Save Changes instead.

If you'd rather not go live yet, create the skill and then toggle it off from the Support Skills page.

Test it before your shoppers see it

Click Preview Conversation from the Policy step at any point. Rep AI saves your work in progress and opens your storefront in a new tab with the chat widget already running your draft skill, having sent "I need help with [your skill name]" for you. Continue the conversation as a shopper would and watch how your policy plays out.

Two things worth knowing:

• Preview runs against your draft version, so you can iterate before anything reaches live shoppers.

• If your store is password-protected, the new tab lands on your store's password page first. Enter the password and the preview continues.

Managing your skills

Every custom skill gets a card in the custom section of the Support Skills page, marked with a CUSTOM badge so you can tell it apart from the pre-built skills.

• Toggle — turn the skill on or off. Toggling off stops the AI from using it and keeps your configuration intact.

• Edit / Activate — reopens the wizard with everything pre-filled. Active skills show Edit; inactive ones show Activate.

• Delete — the trash icon, with a confirmation prompt. Deletion is permanent, and conversations already in progress aren't interrupted.

Changes take effect for new conversations. Conversations already running continue on the version they started with.

Measuring how your skills perform

Custom skills appear in the Support skill performance section alongside the pre-built ones, showing the conversations each one handled.

The performance data comes from a daily rollup, which has two practical consequences:

• Conversations from today won't appear until tomorrow. A skill you just tested will show 0 until the next daily aggregation runs.

• Your date range has to include the day the conversations happened. If you tested on the 19th and your filter covers the 16th to the 18th, you'll see zeros.

A skill legitimately showing 0 either had no matching conversations in the window you selected, or is toggled off.

What the AI will and won't do

It adapts your wording. Your policy controls the steps, the order, the conditions, and the outcome. The exact phrasing is rewritten to match your store's tone and the shopper's language, so write your policy as instructions and intent rather than as a word-for-word script.

It verifies identity before sharing order details. If a shopper asks about an order, the AI requires a second piece of matching information — an email or account — before revealing items, addresses, or status. This applies no matter what your policy says.

It degrades gracefully. If an API call times out or errors, if a required action isn't permitted, or if the conversation goes somewhere your policy doesn't cover, the AI acknowledges it and hands off to a human, passing along everything it already collected.

Tips for a policy that works well

• Invest in the description. Routing quality depends on it more than anything else in the wizard.

• Number your steps and name your conditions. "If the API confirms eligibility… otherwise…" beats a paragraph of prose.

• Say what to do when things go wrong. Give the AI an explicit fallback for out-of-warranty, not-found, and shopper-insists cases.

• Only collect what you'll act on. Every required field is one more thing the shopper must supply before getting help.

• Don't duplicate a pre-built skill. Pre-built skills win on ambiguity, so a custom "Where's my order" skill will rarely be reached.

• Preview, then refine. The fastest way to a good policy is watching a real conversation run against a draft.

Troubleshooting

The Add new skill card is locked.
Your plan doesn't include Expert Zone. If you're an account admin, the tooltip takes you to your plans page. Otherwise, ask your account admin to upgrade.

I can't create another skill.
You've reached the limit of 10. Delete one you no longer use, and the card becomes available again.

My skill isn't handling the conversations I expected.
Check, in order: the skill is toggled on; your description matches how shoppers actually phrase the request; and no pre-built skill covers the same ground — pre-built skills take priority when a message could go either way.

The AI isn't saying what I wrote.
That's expected. Your steps and decisions are followed; the wording is adapted to your store's tone and the shopper's language. Write intent rather than exact scripts.

My API endpoint isn't being called.
Reopen the endpoint and click Test endpoint to confirm it still responds. Then check that your policy names the endpoint exactly as you named it in Step 2, and that the endpoint is publicly reachable — internal addresses and private IPs are blocked.

My endpoint is called but doesn't get the shopper's information.
Collected information travels in the request body, which means POST or PUT. Switch the method if your endpoint currently uses GET.

My skill shows 0 conversations even though I tested it.
Performance figures come from a daily rollup. Wait for the next day, then make sure your date filter includes the day you ran the test.

A shopper reached a human instead of getting resolved.
Escalation is the built-in safety net. It fires when a required action isn't permitted, an API call fails, or the conversation moves outside your policy. The escalation carries a summary of what was already collected, so your team picks up where the AI stopped.

Related articles

• Streamline Customer Service with AI-Powered Support Skills

• Connect Your Own APIs to Create Custom AI Experiences

• Expert Zone: Advanced Configuration for Power Users

• Measuring Your AI Support Performance: Understanding the Support Analytics Dashboard