Can an AI agent answer your tier-1 support tickets?

Can an AI agent do tier-1 support?

Yes, for routine questions, with a person behind it. An AI support agent can answer the repeat questions that fill a tier-1 queue (where is my order, how do I reset my password, what is your return policy) from your own help content, around the clock, and hand everything else to your team. Organizations run this in production today. Gorgias reports that Orthofeet, an orthopedic footwear brand, automated 56% of its support tickets in under two months, with the AI agent handling order-tracking and return requests. Intercom's pricing page quotes Robin's Director of Technical Support reporting a 50% resolution rate with Intercom's Fin agent. Both figures are published by the vendors.

What the agent should not do on its own is give money back, bend policy or keep a customer who wants a person. Its replies go straight to customers, and a wrong one can cost you: a tribunal held Air Canada responsible for its chatbot's advice in one case, though legal responsibility depends on the facts and applicable law. So treat this as a supervised hire: the agent answers what your help content already answers, and a person owns refunds, exceptions and escalations.

This page covers answering customer questions in chat and email. For sorting and routing tickets see ticket triage, for calls see phone support, and for the role as a whole see customer service representative.

What the agent does and what stays with you

Agent

4of 9

• Answer how-to and policy questions from your help content
• Look up order status and tracking for a customer
• Hand off conversations the agent cannot or should not answer
• Record each conversation and its outcome on the ticket

Agent with review

1of 9

• Start a return within the written return policy

Human

4of 9

• Issue refunds, credits or goodwill gestures
• Decide exceptions to policy
• Handle complaints from angry or vulnerable customers
• Keep help content current when the agent gets something wrong

The split follows one rule: the agent can do anything a reviewer can check against your written content, and a person does anything that commits money or changes the rules. Answering from help articles and looking up order status are checkable, so they go to the agent. Returns sit in the middle: Gorgias describes Orthofeet's agent checking return eligibility against written guidance and sending the return portal link, which is fine for the standard case but needs a person for exceptions.

Handoff is a task the agent does, not a failure. Intercom documents that Fin escalates immediately when a customer clearly asks for a human, and offers escalation, rather than handing off at once, when a customer shows strong frustration or repeats themselves across three turns; Fin may first try to resolve a frustrated customer's issue. You can add rules on top that change this, for example escalating whenever a conversation contains an order total over a set value. Write that list before you switch the agent on.

A worked example: one ticket

This is an illustration, not a customer story. Assume, for this example, that you sell online, your return window is 30 days, and refunds over $200 need a lead's approval.

A customer writes in chat at 11 p.m.: "My shoes don't fit, can I send them back?" The agent opens with a line saying it is an AI agent, asks for the order number, and looks up the order: delivered 12 days ago, $140. That is inside the 30-day window, so it explains the return steps from your help article and sends the return portal link. The customer replies "thanks" and leaves. The conversation is tagged resolved.

A second customer asks the same question about an order delivered 45 days ago and adds, "this is ridiculous, I want my money back." The order is outside the window and the customer is frustrated, so the agent does not argue policy. It says it is passing the conversation to the team, tags it "return outside window, refund request," and hands off. In the morning, your support lead reads the thread and decides whether to make an exception. The agent never refused or approved the refund itself.

What you need to provide

The agent touches customer personal data (names, emails, addresses, order history) and needs access to your help desk and, if it looks up orders, your ecommerce or billing system. Give it the least access that gets the job done:

• Your help center articles and written policies. If a policy lives only in a senior agent's head, write it down first; the agent can only be as accurate as what it reads.
• Read access to the help desk and conversation history, so it can match your tone and see past tickets.
• Read-only access to order or account data. Do not give it permission to refund, cancel or edit accounts in the first months.
• A written list of topics that always go to a person, and the team inbox they go to.
• A disclosure line telling customers they are talking to an AI agent.

Disclosure rules vary by jurisdiction. In the EU, Article 50 of the AI Act requires providers to design AI systems that interact directly with people so that those people are informed they are interacting with an AI system, unless this is obvious from the context. Ask counsel which obligations apply to your role and jurisdiction. Separately from the law, Cursor said after its incident that AI responses in its email support are now clearly labeled.

Job brief template

Use this as the brief for a vendor, an operator or an internal team setting up the agent. Replace the bracketed values with your own.

The acceptance criteria are the part to keep. A weekly sample of 50 conversations read against your written policies tells you more than any resolution rate, because a resolution rate counts conversations that ended, not answers that were right.

What it costs

The human figure is the 2025 national median annual wage for customer service representatives (SOC 43-4051), $44,770, or $21.53 an hour, as published on O*NET OnLine from BLS data. It is wage only.

The agent figures are published per-resolution prices from two vendors, and each sits on top of a help desk subscription. Gorgias charges $0.90 per automated interaction on an annual contract, $1.00 on a monthly contract and $1.50 for overage above the plan's included allowance; its helpdesk plans start at $10 a month for 50 tickets, and each automated interaction also counts as a helpdesk ticket. Intercom charges $0.99 per Fin outcome, at most once per conversation, plus seats from $19 per seat a month on its Essential plan billed annually; used with another help desk, Fin needs no seats but has a monthly minimum, which Intercom's pricing page gives as "e.g. 50 outcomes."

Read the billing definition before you compare. Intercom counts a resolution when the customer confirms the answer helped or leaves without asking for more help, and it bills that second case too. A customer who gave up is billed the same as one who was helped. Your reviewer's time and the work of writing down policies sit outside both numbers; measure them in your trial.

Risks and human checkpoints

Wrong answers are the risk that justifies the checkpoints. In February 2024 a British Columbia tribunal ordered Air Canada to pay a passenger $812 (Canadian) after its website chatbot gave incorrect advice about bereavement fares; according to CBC News, the tribunal member rejected the airline's argument that the chatbot was responsible for its own actions and found Air Canada did not take reasonable care to ensure its chatbot was accurate. That was one tribunal's finding in one dispute; legal responsibility depends on the facts and applicable law, so ask counsel how it applies to you. CBC reported that it appeared to be the first such case in the CanLII database, which is why it is cited here despite its age. In April 2025, Cursor's AI support bot told users that being logged out when using several machines was expected behavior, a policy that did not exist; Cursor's co-founder apologized publicly, and according to The Register, his early investigation suggested the logouts appeared to result from a session race condition.

Our recommendation: answer both cases with process, not tooling. Keep written policies the agent can read, a list of topics it must hand off, and a person reading a sample every week.

How to run a trial

Run the agent on part of your queue for one month before you widen it.

1. Pick a slice: 3–5 high-volume topics (order status, returns inside the window, password resets) on one channel, usually chat.
2. Configure handoff first. Write the always-hand-off list and test it with your own messages before any customer sees the agent.
3. Turn on the AI disclosure line and read-only access only.
4. Each week, read 50 agent-closed conversations against your written policies and count wrong answers against the brief's criteria. Read every conversation with a negative rating.
5. Compare satisfaction and repeat contacts for agent-closed and human-closed conversations on the same topics.
6. Track handoff reasons. A topic that keeps getting handed off usually needs a help article, not a different agent.
7. Watch the bill. Set a usage limit and compare billed resolutions with the conversations you judged correct.

1. 1

   Pick a slice

   3–5 high-volume topics on one channel, usually chat

2. 2

   Configure handoff first

   Write the always-hand-off list and test it before any customer sees the agent

3. 3

   Turn on the AI disclosure line

   Read-only access only

4. 4

   Each week, read 50 agent-closed conversations

   Against your written policies; read every conversation with a negative rating

5. 5

   Compare satisfaction and repeat contacts

   Agent-closed against human-closed conversations on the same topics

6. 6

   Track handoff reasons

   A topic that keeps getting handed off usually needs a help article

7. 7

   Watch the bill

   Set a usage limit and compare billed resolutions with the ones you judged correct

If the samples meet the criteria for two weeks in a row, add the next topic. For more on structuring the test, see how to run a paid trial with an AI agent, and for writing the brief, how to write a job description for an AI agent. Related support jobs: knowledge base maintenance and the customer support overview.

Sources

1. 1
   43-4051.00 - Customer Service Representatives

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

2. 2
   Article 50: Transparency obligations for providers and deployers of certain AI systems

   European Commission, AI Act Service DeskPrimaryAccessed Nov 14, 2026

3. 3
   Regulation (EU) 2024/1689 (Artificial Intelligence Act)

   EUR-Lex, Publications Office of the European UnionPrimaryAccessed Nov 14, 2026

4. 4
   Manage Fin AI Agent's escalation guidance and rules

   IntercomPrimaryAccessed Nov 14, 2026

5. 5
   Fin AI Agent outcomes

   IntercomPrimaryAccessed Nov 14, 2026

6. 6
   Intercom pricing

   IntercomVendorAccessed Nov 14, 2026

7. 7
   Gorgias pricing

   GorgiasVendorAccessed Nov 14, 2026

8. 8
   Orthofeet automating 56% of tickets in 2 months with Gorgias

   GorgiasVendorAccessed Nov 14, 2026

9. 9
   Air Canada found liable for chatbot's bad advice on plane tickets

   CBC NewsPressAccessed Nov 14, 2026

10. 10
    Cursor AI's own support bot hallucinated its usage policy

    The RegisterPressAccessed Nov 14, 2026