AI agents are specialist assistants that handle one job for your store, end to end, 24×7.
A shopper wants to cancel an order, track a parcel, return a product or file a warranty claim. Instead of a fixed chatbot flow, an AI agent understands what they type, decides what to do next, and uses tools (look up an order, cancel it, send an email, call your API…) to actually get the job done. When a real person is needed, it hands over to your team.
With the Agent Builder (you'll find it under AI Agents in the left menu of your Manifest dashboard, and you may hear it called Agent Studio), you can create your own AI agents without writing any code.
This article walks you through everything: the pre-built agents, every step and field in the builder, the full list of tools, customer verification, AI model selection, handover, testing and going live, plus worked examples and tips.
1. What are AI agents?
An AI agent takes over a conversation when a shopper needs something specific and handles it from start to finish. Each agent manages one task, like cancellations or returns, and defers to your human team only when necessary.
For every agent, you decide four things:
About: its name, what it does, and which channels it works on.
Trigger: when it should take over a conversation.
How it responds: its instructions, guardrails, tools, sub-agents and AI model.
Handover: where the conversation goes when a person needs to step in.
Agents stay in Draft until you click Go live, so you can build and test safely. Changes don't affect customers until you go live.
2. Pre-built AI agents
When you open AI Agents, you land on the Pre-built AI agents page. These are ready-made agents for the most common D2C jobs:
Pre-built agent | What it does |
Product Sales Concierge | Helps customers get information about a product and improves sales. |
Order Cancellation Agent | Helps process cancellation requests: looks up the order, checks if it's still cancellable, tags it in Shopify, and alerts the store team. |
Returns and Exchanges agent | Manages return and exchange requests end to end: looks up orders, understands the issue, and tries to save the sale before logging a return. |
If one of these fits your need, start from it. You can edit everything afterwards. If not, click + New AI agent in the top-right corner to build your own from scratch.
3. Create a new AI agent
Before you begin
Pick one job for the agent, for example Order cancellation. A focused agent works far better than a do-everything agent.
Connect your integrations first (Shopify, BigCommerce, ShipStation, Slack, Gorgias, Klaviyo, Omnisend, Segment…) from Integrations, so their tools appear and work in the builder.
Steps to create an agent
Go to AI Agents in the left menu of your Manifest dashboard.
Click + New AI agent.
Complete the four steps shown on the left of the builder: About this agent, When it steps in, How it responds and Handover rules.
Click Test agent at any point to try it out.
Click Save as draft (top right) to save your progress.
When you're happy, click Go live.
A new agent starts as Untitled agent with a Draft label next to its name. The note at the bottom of the builder is a helpful reminder: Changes don't affect customers until you go live.
4. Step 1: About this agent
Give your agent a name and describe what it handles.
Field | Required? | Limit | What to enter |
Agent name | Yes | 60 characters | What you and your team will see in the agents list, e.g. Warranty Assistant. |
What does it handle? (description) | No | 1,000 characters | One or two clear lines on what the agent does. If you use this agent as a sub-agent of another agent, this description is how the parent agent knows what it's for, so keep it clear. |
Example description:Helps customers who want to cancel an order: finds the order, cancels it if it hasn't shipped, tags it for the team and confirms with the customer.
5. Step 2: When it steps in (trigger)
Every agent needs a trigger: this tells Manifest when to route a conversation to your agent. Choose one of two options.
Option A: Manifest intent
Manifest's AI already understands what shoppers are asking about, with no keywords needed. Under If customer talks about, click + Intent and pick one or more intents. Your agent takes over whenever Manifest recognises that intent in a message, replacing Manifest's built-in reply for that topic. The changeover is automatic when you go live.
Intent | Example shopper message |
Product details | “Is this jacket waterproof?” |
Product customisation | “Can I get my name engraved on this?” |
Offers & rewards | “Any discount on this?” |
Checkout enquiry | “I can't complete my checkout.” |
Payment methods | “Do you accept cash on delivery?” |
Order tracking | “Where is my order?” |
Order cancellation | “I want to cancel my order.” |
Order enquiry | “I have a question about my order.” |
Product search | “Show me running shoes under $100.” |
One agent per intent. Each intent can belong to only one agent in your store, and drafts count too. If you pick an intent that another agent already uses, the builder tells you which agent has it, so you can change one of them.
Option B: Custom intent
If your use case isn't in the list above, select Custom intent and describe the trigger in plain English in Describe the trigger (up to 1,000 characters). For example:
Customer wants to file a warranty claim
Customer is asking about bulk or wholesale orders
Customer wants to exchange a product for a different size
Manifest's AI compares each shopper message with your description and hands the conversation to your agent when it matches.
Writing a good custom intent
Describe the shopper's goal, not keywords.“Customer wants to return an item” works better than “return, refund, send back”.
Keep it to one topic. If you have two different jobs, build two agents.
Make it distinct from your other agents' custom intents, so Manifest isn't torn between two agents.
Other ways to start an agent
Besides triggers, you can point a quick action, conversation starter, widget trigger, nudge button or custom handover directly at a live agent. When a shopper clicks it, that agent starts straight away.
6. Step 3: How it responds
This is where you tell the agent what to do and give it what it needs to do it.
Setting | Required? | Limit | What it's for |
Instructions | Yes | 10,000 characters | The agent's playbook: its goal, the steps to follow, decisions and tone. |
Guardrails | No | 2,000 characters | Hard “never do” rules the agent must not break. |
Tools | — | — | The actions the agent can take (look up an order, send an email, call your API…). Add them with the + button or by typing “@” in your instructions. |
Sub-agents | — | — | Other agents this agent can hand part of the conversation to. |
Require customer verification | — | Toggle | Customers must verify their identity before this agent responds. Needed for order tools. |
Advanced settings → Select AI model | — | Dropdown | Which Manifest AI model the agent runs on. |
Instructions
Write your instructions the way you'd brief a new support teammate, in plain English. The agent follows them closely, so the more specific you are, the better it performs. A good set of instructions covers:
The goal: “You help customers file warranty claims.”
The steps, in order: “1. Ask for the order number. 2. Look the order up with @Track an order. 3. Ask for a photo of the issue. 4. Email the details with @Warranty email.”
Decisions: “If the order is older than 12 months, explain that it's out of warranty and offer a discount code instead.”
Tone: “Be friendly and brief. Use the customer's first name.”
Use “@” inside your instructions to point at the exact tool or sub-agent to use at each step (see section 7).
Guardrails
Guardrails are the agent's hard “never do” rules. They're kept separate from the instructions and given to the agent as non-negotiable limits. Examples:
Never cancel an order without confirming the order number with the customer.
Never promise a refund amount.
Only approve returns within 30 days of delivery.
Don't discuss competitor products.
Never discuss competitor pricing. Always stay within the cancellation policy. Do not process refunds above ₹10,000 without human approval.
You can use “@” in guardrails too, for example to reference a variable or a tool by name.
On top of your guardrails, every agent follows a few built-in safety rules. For example, it never hands out support contact details itself: when a customer needs a person, it uses the handover you configure (see Step 4).
7. Using “@” to reference tools, sub-agents and variables
Type @ in the Instructions or Guardrails box to open the picker. You can insert three kinds of references, which appear as chips in your text:
You pick | What it does | Example |
A tool (browse by source, e.g. Manifest, Shopify, Slack) | Tells the agent exactly which tool to use at that step. Picking a tool here also attaches it to the agent. | “Look the order up with @Track an order.” |
A sub-agent | Tells the agent when to delegate part of the job to one of its sub-agents. | “For size questions, ask @Size guide expert.” |
A variable | Inserts a live value about the customer or your store, filled in fresh on every message. | “Greet the customer as @Customer Name.” |
Available variables
Group | Variable | Value |
Customer | Customer Name | The customer's name, from the chat if they shared it, otherwise from their store account. |
Customer | Customer Email | The customer's email, from the chat if they shared it, otherwise from their store account. |
Auth | Verified Email | The email the customer verified with an OTP (empty until they verify). |
Auth | Verified Phone Number | The phone number the customer verified with an OTP (empty until they verify). |
Store | Store Name | Your store's name. |
Store | Store URL | Your store's website. |
Store | Store Email | Your support email (the one set for human handover). |
Variables also work inside the Send an email tool (all fields) and the Call your API tool (URL, headers and instructions). If a variable has no value yet (for example Verified Email before the customer verifies), it's left blank.
💡 Tip: Referencing tools with “@” is the single most effective way to make an agent reliable. “Use @Track an order” beats “check the order status”.
8. Tools
Tools are the actions your agent can take. Add them in How it responds, either from the Tools panel (click +) or by typing “@” inside your instructions. The Tools panel groups them by source, such as Shopify, Manifest, Integrations, REST API, Send email and Google Sheets. An agent can only use the tools you attach.
Two columns in the tables below matter:
Needs verification: the tool reads or changes a customer's private data (usually orders). On your website chat, the customer must verify their email or phone before the agent can use it, so turn on Require customer verification for the agent (see section 10).
Needs connection: the tool works through another app, so that app must be connected to your store first. The builder shows a Connect button next to tools that aren't connected yet.
8.1 Product tools
Use these to answer product questions, recommend products and compare them.
Tool | What it does | Needs verification | Needs connection |
Search products (fast) | Searches your catalog and returns prices, variants and stock, so the agent can choose what to recommend. | No | No |
Search products (thorough) | Searches your catalog more thoroughly, following your search rules and requests like best sellers or new arrivals. Slower than the fast search. | No | No |
Get details for one product | Gets everything about one product: description, variants, price, stock, discounts, materials or care, and its FAQs. | No | No |
Get details for multiple products | Gets the same details for several products at once, so the agent can compare them. | No | No |
Ask about a product (older) | Answers any question about a product using your catalog and trained content. Prefer Get details for one product for new agents. | No | No |
Get customisation options (older) | Answers whether and how a product can be customised or personalised, using your trained content. | No | No |
Find offers on a product (older) | Finds discounts, coupons and offers available on a product. | No | No |
When to use:Search products (fast) for quick recommendations; Search products (thorough) when shoppers ask for best sellers, new arrivals or when your search rules matter; Get details for multiple products for side-by-side comparisons.
8.2 Shopify tools (Shopify stores)
Use these to look up, cancel, tag or annotate orders, manage customer accounts and give discount codes.
Tool | What it does | Needs verification | Needs connection |
Get customer's orders | Shows the customer their orders to pick from, then reads the one they choose. | Yes | Shopify |
Track an order | Shows the status and tracking of one of the customer's orders, using its order number. | Yes | Shopify |
Cancel an order | Cancels one of the customer's Shopify orders, using its order number. | Yes | Shopify |
Tag an order | Adds a tag to one of the customer's orders, so your team can sort or find it. | Yes | Shopify |
Add a note to an order | Adds a note to the customer's order for your team, like a delivery request or address change. The customer doesn't see it. | Yes | Shopify |
Look up the customer's account | Checks whether the customer has a store account, and shows the details and number of orders on it. | Yes | Shopify |
Create a customer account | Creates a store account using the customer's name, email and phone number. | No | Shopify |
Give a discount code | Creates a working discount code from one of your existing Shopify discounts, when your instructions say the customer qualifies. | No | Shopify |
Search products | Searches your Shopify store's products by keywords and filters. | No | Shopify |
Get products by ID | Gets specific products or variants from your Shopify store, up to 10 at a time. | No | Shopify |
Get product details | Gets full details for one Shopify product, including its options like size or colour. | No | Shopify |
Search policies and FAQs | Answers questions using your Shopify store's policies and FAQs. | No | Shopify |
Tip: For Give a discount code, spell out in your instructions exactly when a customer qualifies, e.g. “Only offer @Give a discount code if the customer was about to cancel because of price.”
8.3 BigCommerce tools (BigCommerce stores)
Tool | What it does | Needs verification | Needs connection |
Get customer's orders | Lists the customer's recent BigCommerce orders and how long ago each was placed. | Yes | BigCommerce |
Find an order by number | Finds one of the customer's BigCommerce orders by its order number, with its items, status and tracking. | Yes | BigCommerce |
Cancel an order | Cancels the customer's BigCommerce order if it hasn't shipped yet. This can't be undone. | Yes | BigCommerce |
Refund an order | Refunds the full order amount to the way the customer paid. Partial refunds aren't possible. This can't be undone. | Yes | BigCommerce |
⚠️ Cancel an order and Refund an order can't be undone. Add guardrails such as “Never refund without confirming the order number and getting the customer's explicit yes.”
8.4 Shipping and returns tools
Tool | What it does | Needs verification | Needs connection |
Request a return (AfterShip) | Starts a return in AfterShip for a whole order that has shipped or been delivered. Your team reviews it and sends the return label. Returning only some items isn't supported. | Yes | No, available to every store |
Get tracking details (ShipStation) | Finds the tracking number, carrier and ship date for an order in ShipStation. | Yes | ShipStation |
Cancel shipping labels (ShipStation) | Cancels every shipping label on an order. Labels the carrier has already scanned can't be cancelled. This can't be undone. | Yes | ShipStation |
Add a note for the warehouse (ShipStation) | Adds a note to the order's shipments in ShipStation. Only your team sees it. | Yes | ShipStation |
Cancel an order (ShipStation) | Cancels the order in ShipStation so it isn't packed or shipped. It doesn't cancel shipping labels already bought. | Yes | ShipStation |
8.5 Helpdesk and marketing tools
Use these to open tickets for your team, or to save customers and their actions in your marketing tools.
Tool | What it does | Needs verification | Needs connection |
Create a Gorgias ticket | Opens a Gorgias ticket with a summary of the customer's issue so your team can follow up. Needs the customer's email. | No | Gorgias |
Save customer to Klaviyo | Adds the customer to Klaviyo, or updates them if they're already there, after they share their email and agree to hear from you. | No | Klaviyo |
Record an action in Klaviyo | Records something the customer did, like asking for a restock alert, so you can follow up from Klaviyo. | No | Klaviyo |
Save customer to Omnisend | Adds the customer to Omnisend, or updates them if they're already there, after they share their email and agree to hear from you. | No | Omnisend |
Record an action in Omnisend | Records something the customer did, like asking for a restock alert, so you can follow up from Omnisend. | No | Omnisend |
Save customer to Segment | Adds the customer to Segment, or updates them if they're already there, after they share their email and agree to hear from you. | No | Segment |
Record an action in Segment | Records something the customer did, like asking for a restock alert, so you can follow up from Segment. | No | Segment |
8.6 Custom actions (tools you configure)
These tools are templates: you fill in a short form and the result becomes a tool for this agent. You can create as many as you need, e.g. one email to your warranty team and another to your wholesale team.
Tool | What it does | What you configure | Needs verification | Needs connection |
Send an email | Emails the details collected in the chat to the address you choose. | Action name, Send to, CC, Reply to, Subject, Message instructions (how the agent should write the email), and whether to include the full chat history. Each field can be a fixed value or a prompt such as “the customer's email”. | No | No |
Call your API | Sends a request to an API you set up and uses the reply in the chat. | Name, request URL, HTTP method (GET/POST/PUT/PATCH/DELETE), headers, the request fields (with a plain-English description of what goes in each), and instructions for how the agent should build the request. | No | No |
Send a Slack message | Posts a message with details from the chat to your Slack channel. | Slack channel ID, and instructions for what the message should contain. | No | Slack |
Look up a Google Sheet | Reads rows from a Google Sheet you connect, to answer the customer's question. | Sheet URL, what each column holds, and when/how to use the results. The sheet must be shared publicly (view). | No | No |
Add a row to a Google Sheet | Saves details from the chat as a new row in a Google Sheet you connect. | Sheet URL, the column headers and what goes in each, and when to add a row. The sheet must be public and give edit access to | No | No |
When to use:Send an email or Send a Slack message to alert your team (e.g. a new warranty claim); Call your API to read from or write to your own systems (e.g. a loyalty points balance); Look up a Google Sheet for simple reference data (e.g. pin codes you deliver to); Add a row to a Google Sheet to log requests (e.g. wholesale enquiries).
8.7 Built in: added to every agent automatically
You don't need to add these. Every agent has them:
Hand over to a person: used when the customer asks for a human or the issue can't be resolved by the agent (see Step 4).
Remember customer details: saves a name or email the customer shares, so the agent (and the Customer Name / Customer Email variables) can use it.
Order and product pickers: when an agent has order tools, it can show the customer a list of their orders or products to tap on, instead of asking them to type an order number.
9. Sub-agents
For bigger jobs, you can split the work. An agent can have sub-agents, each with its own instructions and tools. Add them under Sub-agents in How it responds (click +).
The main agent decides when to call a sub-agent. Use “@” in your instructions to tell it when, e.g. “For size questions, ask @Size guide expert.”
The sub-agent's answer comes back to the main agent, which carries on the conversation.
The sub-agent's description (from Step 1) is how the main agent knows what it's for, so keep it clear.
Sub-agents don't have their own trigger, and they can't have sub-agents of their own.
When to use sub-agents
Most agents work best doing one job on their own. A sub-agent is useful when part of that job needs its own specialist, with its own instructions and tools, while the main agent stays in charge of the conversation.
Use a sub-agent when:
One step needs a specialist. For example, a returns agent that also needs to help the customer pick a different size or product for an exchange. The main agent handles the return and asks the sub-agent for the product part, then carries on.
A step needs different tools. Each sub-agent has its own tools, so the main agent can keep its order tools while a sub-agent uses, say, the product search tools. Each agent then only sees the tools it needs.
Your instructions are getting long or hard to follow. Moving a self-contained part of the job (like size advice) into a sub-agent keeps each set of instructions short and focused, which makes both agents more reliable.
You want to reuse a specialist. An agent you've built can be added as a sub-agent of another agent. Its description tells the main agent what it's for, so you only write that expertise once.
Don't use a sub-agent when:
The job is simple. If one agent with a few tools can handle it, keep it in one agent. One agent, one job.
The customer is switching to a completely different task. Build a separate agent with its own trigger instead. When the customer changes topic to something another live agent handles, the conversation moves to that agent automatically.
It needs to start on its own. Sub-agents don't have their own trigger. They only run when the main agent calls them.
It needs its own sub-agents. Sub-agents can't have sub-agents of their own, so keep the setup to one level.
You only need a single action, like sending an email or tagging an order. Add that tool to the main agent instead.
Examples from our Pre-Built Agents :
Product Sales Concierge + size guide expert: the main agent recommends and compares products. In its instructions: “For size questions, ask @Size guide expert.” The sub-agent answers the size question, and the main agent continues the sale.
Returns and exchanges agent + product finder: the main agent looks up the order and handles the return. When the customer wants a different product instead, it asks a Product finder sub-agent (with the product search tools) for options, then completes the exchange conversation.
💡 Tip: Give every sub-agent a clear description, and reference it with “@” at the exact step in the main agent's instructions where it should be called.
10. Customer verification
Order tools work with a customer's private data, so the agent must know who it's talking to before using them. Turn on Require customer verification for any agent that uses a tool marked Needs verification, and choose one method:
Email OTP: the customer enters their email and a one-time code sent to it.
Phone OTP: the customer enters their phone number and a one-time code sent to it.
How it behaves on each channel:
Channel | What happens |
Website chat (Manifest widget) | Before the agent starts, the customer sees a verification form. Once verified, they aren't asked again in the same chat for 24 hours. Customers already logged in to your store skip the form. |
No form. The customer is identified by their WhatsApp number. | |
No form. The customer is identified by the email address they wrote from. | |
Instagram can't verify a customer, so order tools won't share order details there. The agent offers to connect the customer to a person instead. |
If the customer closes the form or can't verify, the agent doesn't start and the customer is told it can't continue without verification.
11. Model selection
Under Advanced settings → Select AI model in How it responds, you can choose which AI model the agent runs on. The options are the same Manifest model tiers you see on the AI model card in Customise → Behaviour:
Option | Best for |
Manifest Pro | Balanced performance for reliable, cost-efficient customer support. |
Manifest Ultra | Ultimate precision for complex queries and personalised interactions. |
Credit usage follows the tier, as shown on that card. If you're unsure, leave it unselected and your agent will match the rest of your Manifest experience.
12. Step 4: Handover rules
Handover is what happens when a customer needs a real person. The agent hands over when:
the customer asks for a human, person or support agent, or
the issue can't be resolved with the agent's instructions and tools.
The agent never makes up support contact details. It always uses the handover you configure here.
Where the conversation goes
Switching between the two options doesn't lose the journeys you've picked.
When to hand over
Use this field to describe the situations where this agent should bring in a person, for example “The parcel is marked lost or has been stuck for more than 7 days.”
💡 Also write the same rule into your Instructions (e.g. “If tracking shows the order is lost, hand over to a person”), so the agent acts on it during the conversation.
To learn more about your store's standard handover setup, see Agent Handover.
13. Test, save as draft and go live
Test agent: click Test agent at the bottom of the builder at any time. It opens the real chat widget, so what you see is exactly what shoppers will get.
Save as draft: click Save as draft in the top-right corner to save your work. Drafts don't affect customers.
Go live: when you're happy with your tests, click Go live. From then on, your agent takes over conversations that match its trigger, on the channels you switched on.
What to test before going live
The happy path: a customer who does exactly what you expect.
A confused customer: vague messages, typos, missing order numbers.
A topic change halfway through the conversation.
“Talk to a human”: check your handover works.
Verification: if the agent uses order tools, check the verification form appears and works.
14. What happens during a conversation
A message matches your trigger (Manifest intent or custom intent), or the customer taps a button pointed at your agent.
Verification runs, if it's enabled and needed on that channel.
The agent owns the conversation. Every following message goes to the same agent, which keeps working through its instructions and tools.
The agent steps aside when: the customer changes topic to something another live agent handles (the conversation moves to that agent); the customer changes topic to something Manifest handles (the conversation goes back to Manifest's normal replies); the customer needs a person (your handover rules run); or the customer types EXIT.
15. Worked examples
Example 1: Order cancellation agent (Shopify)
Step | What to set |
About | Name: Order Cancellation Agent. Description: Helps customers cancel an order: finds the order, cancels it if it hasn't shipped, tags it for the team and confirms with the customer. Channels: website chat and WhatsApp. |
Trigger | Manifest intent: Order cancellation. |
Tools | @Get customer's orders, @Track an order, @Cancel an order, @Tag an order, and a Send a Slack message action called Cancellation alert (or a Send an email action to your ops team). |
Verification | On (Email OTP or Phone OTP), because the order tools need it. |
Model | Not selected (uses your store's model). |
Handover | Default. When to hand over: The order has already shipped, or the customer is unhappy with the outcome. |
Instructions:
You help customers cancel orders. Be friendly, brief and use @Customer Name.
1. Use @Get customer's orders so the customer can pick the order they want to cancel.
2. Check its status with @Track an order.
3. If the order has already shipped, explain it can't be cancelled any more and hand over to a person.
4. If it hasn't shipped, ask the customer why they want to cancel, then ask them to confirm the order number.
5. After they confirm, cancel it with @Cancel an order.
6. Tag the order "AI-cancelled" with @Tag an order.
7. Alert the team with @Cancellation alert, including the order number and the reason.
8. Confirm the cancellation to the customer.
Guardrails:
Never cancel an order without confirming the order number with the customer.
Never promise a refund amount or a refund date.
Always stay within the cancellation policy.
Example 2: Warranty claim agent (custom intent)
Step | What to set |
About | Name: Warranty Assistant. Description: Collects warranty claims: checks the order, gathers the issue details and emails them to the warranty team. |
Trigger | Custom intent: Customer wants to file a warranty claim. |
Tools | @Track an order, @Give a discount code, and a Send an email action called Warranty email (Send to: your warranty team; Reply to: the customer's email; include the full chat history). |
Verification | On, because Track an order needs it. |
Handover | When to hand over: The customer disputes the warranty decision. |
Instructions:
You help customers file warranty claims. Be friendly and brief. Use the customer's first name.
1. Ask for the order number.
2. Look the order up with @Track an order.
3. If the order is older than 12 months, explain that it's out of warranty and offer a discount code with @Give a discount code instead.
4. Otherwise, ask for a photo and a short description of the issue.
5. Email the details with @Warranty email and tell the customer the team will be in touch at @Customer Email.
6. If the customer disputes the decision, hand over to a person.
Guardrails:
Never promise a replacement or refund. The warranty team decides.
Don't discuss competitor products.
16. Tips for writing great instructions
One agent, one job. A focused Order cancellation agent beats a do-everything agent.
Write numbered steps in the order the agent should follow them.
Reference tools with “@” at the exact step they're needed. “Use @Track an order” beats “check the order status”.
Spell out decisions with clear “If… then…” sentences, e.g. “If the order has shipped, hand over to a person.”
Put “never” rules in Guardrails, not buried in the instructions.
Set the tone in one line, e.g. “Be friendly and brief. Use the customer's first name.”
Use variables (@Customer Name, @Store Name…) to personalise replies.
Repeat handover rules in your instructions as well as in When to hand over.
Turn on verification whenever you add a tool marked Needs verification.
Connect integrations first (Shopify, Slack, Gorgias, Klaviyo…) so their tools appear and work.
Test before going live with Test agent: the happy path, a confused customer, a topic change and “talk to a human”.
Start from a pre-built agent if one fits. You can edit everything afterwards.
17. Troubleshooting
Problem | What to check |
I can't select an intent. | Each intent can belong to only one agent, drafts included. The builder tells you which agent already has it. Remove it from that agent, or use a custom intent instead. |
A tool shows a Connect button or doesn't work. | The tool needs an integration. Connect that app (e.g. Shopify, Slack, Gorgias) to your store first. |
The agent can't see order details. | Turn on Require customer verification. On Instagram, customers can't be verified, so the agent offers to connect them to a person instead. |
My agent never takes over. | Check that it's live (not a draft), that the channel is switched on in About, and that your custom intent describes the shopper's goal clearly and doesn't overlap with another agent's. |
The agent uses the wrong tool or skips a step. | Reference the exact tool with “@” at each step, and write the steps as a numbered list. |
A variable comes out blank. | It has no value yet, e.g. Verified Email before the customer verifies, or Customer Name before the customer has shared it. |
Look up a Google Sheet returns nothing. | Make sure the sheet is shared publicly (view). |
Add a row to a Google Sheet fails. | The sheet must be public and give edit access to |
The customer wants to leave the agent. | They can type EXIT, change topic, or ask for a person. |
FAQs
1) Do changes affect customers while I'm building?
No. Agents stay in Draft until you click Go live, and changes don't affect customers until you go live.
2) How do I see exactly what shoppers will see?
Click Test agent. It opens the real chat widget.
3) Can two agents use the same Manifest intent?
No. Each intent can belong to only one agent in your store, including drafts.
4) What's the difference between a Manifest intent and a custom intent?
A Manifest intent uses Manifest's built-in understanding of common topics (like order tracking), with no keywords needed. A custom intent lets you describe any other trigger in plain English, and Manifest's AI matches shopper messages to it.
5) What's the difference between instructions and guardrails?
Instructions are the playbook (goal, steps, decisions, tone). Guardrails are hard “never do” limits that the agent treats as non-negotiable.
6) Do I need to add a handover tool?
No. Hand over to a person, Remember customer details and order and product pickers are added to every agent automatically.
7) Which AI model should I choose?
If you're unsure, leave it unselected and the agent uses your store's model from Customise → Behaviour. Choose Manifest Ultra for complex, personalised conversations, or Manifest Pro for balanced, cost-efficient support. Credit usage follows the tier.
8) Will verified customers be asked to verify again?
On website chat, not again in the same chat for 24 hours. Customers already logged in to your store skip the form. On WhatsApp and email, there's no form at all.
9) Can a sub-agent have its own sub-agents or trigger?
No. Sub-agents are called by the main agent only.
10) Can the agent make a partial refund or return only some items?
No. Refund an order (BigCommerce) refunds the full order amount, and Request a return (AfterShip) works on whole orders only.
11) Can I start an agent from a button?
Yes. You can point a quick action, conversation starter, widget trigger, nudge button or custom handover at a live agent.
For further assistance or to raise feature requests related to AI Agents, please contact [email protected].