Fyncall wrote this page and sells an AI agent for Shopify stores. The security guidance we cite links to OWASP's own page, checked 19 November 2026.
What is an AI agent in customer service?
An AI agent is an AI that handles a customer conversation on its own: it understands the question, finds the answer in your data, and can carry out allowed actions. It works inside rules you set and passes the conversation to a person when it should.
Three parts make an agent more than a text generator:
- Knowledge: your policies, FAQs and product data, so answers come from your business rather than general guesswork.
- Tools: connections to systems such as your store, so it can look up an order instead of asking a person to.
- Rules and handoff: limits on what it may say and do, and conditions for bringing in your team.
How is an AI agent different from a chatbot or a copilot?
A scripted chatbot follows paths someone built in advance, an AI agent decides its own next step within limits, and a copilot helps a human agent but does not reply to the customer itself.
| Aspect | Scripted chatbot | AI agent | AI copilot |
|---|---|---|---|
| Who it talks to | The customer | The customer | Your support team |
| How it decides | Buttons, keywords and decision trees | Reads free text and chooses an answer or a tool | Suggests drafts and lookups when a person asks |
| Handles unexpected questions | Usually falls back to a menu or a person | Attempts an answer from its knowledge, or hands over | A person decides what to send |
| Main risk | Customers get stuck in loops | Wrong answers or wrong actions without review | Low, because a human sends every reply |
Many helpdesks offer both an agent and a copilot. In Fyncall, the AI Copilot sits beside a conversation once a person has taken over. It can draft a reply, look up orders and summarise the conversation, but the team member chooses what to send. Our blog post on AI inbox versus AI agent covers the trade-off in more detail.
What do tools and actions mean for an AI agent?
Tools are the specific functions an agent is allowed to call, such as "look up order" or "update shipping address". Read-only tools fetch information; action tools change something, so they need tighter control.
For a Shopify store, typical tools fall into two groups:
- Look-ups: order status and tracking, stock levels, product details, discount codes you have set up.
- Actions: updating a shipping address, cancelling an order that has not shipped, opening a ticket.
Fyncall's agents have tools in both groups, listed on the AI support agent page. Each tool can be switched off for each agent in its guardrails, the agent asks the customer to confirm before changing an order, and it checks refund eligibility but does not issue refunds itself.
What are the risks of an AI agent?
The main risks are confident wrong answers, actions taken on the wrong order or without consent, and customers steering the AI away from its instructions.
That last risk has a name. OWASP lists prompt injection first in its 2025 Top 10 for LLM applications, describing it as user prompts that alter the model's behaviour or output in unintended ways. Among its suggested mitigations are restricting the model's access to the minimum it needs and requiring human approval for privileged operations. Both apply directly to support agents.
What guardrails should an AI agent have?
A support agent should have limited tools, written rules, a way to review its replies before they go live, and clear triggers for handing over to a person.
- Least access: switch on only the tools it needs. Leave cancellations off until you have watched it handle look-ups.
- Written rules: topics to avoid, words never to use, and rules such as "never offer more than 10% off". Fyncall's guardrails cover Never Discuss, Never Say, Must Follow, tone, reply length and tool access.
- Grounded answers: answer from your documents. Fyncall's agent only uses knowledge base documents that have finished indexing.
- Review before automation: start with drafts a person approves. Fyncall offers Auto-Reply, Draft & Review or Off for each channel.
- Testing: run real past questions through a test sandbox and read which tools it used.
- Handoff triggers: a request for a person, signs of frustration, repeated failed attempts and sensitive topics. Fyncall's escalation settings cover each of these. See human handoff.
Common questions
Is an AI agent the same as a chatbot?
Not quite. Both talk to customers, but a scripted chatbot follows fixed paths, while an AI agent interprets free text and can use tools. Some products called chatbots now include AI agents, so check what a tool actually does.
Can an AI agent replace a support team?
It can take routine questions such as order status, but complaints, exceptions and upset customers still need people. Plan for handoff from the start.
Does Fyncall's AI agent work on every channel?
Fyncall gives each connected channel its own agent: WhatsApp, Instagram, email and web chat. Each can have its own personality, knowledge, guardrails and reply mode.
How much does Fyncall's AI agent cost?
The Free plan includes 25 AI messages a month, and Pro is $34 a month with 500. Human replies are not counted. See pricing for details.