Which customer support jobs can an AI agent take on today?

AI agents for customer service do real work today on the routine end of the queue, according to two vendors' product documentation. Intercom documents its Fin AI Agent answering customer questions using your available content and handing conversations to a person when it cannot help or the customer asks for one. Zendesk documents AI classification of incoming tickets by topic, sentiment and language. We recommend that people keep refunds and exceptions, angry or high-stakes customers, and the decision about what the agent may say. The practical question is which support job to hand over first and how you will check the agent's answers.

What AI agents do in customer support today

Two help desk vendors document how their agents work and, usefully, how they count a success.

• Intercom's help center documents outcome pricing for Fin AI Agent over chat and email, and describes Fin answering customers "using your available content." It also documents Procedures you configure that end in a handoff to a human or a workflow. Fin hands over when it cannot resolve the question, when a customer asks for human support, or when an escalation rule applies, and by default it escalates when a customer shows frustration.
• Zendesk's help center documents how it counts and bills AI agent resolutions, with separate rules for when email, messaging and voice conversations end. It also describes intelligent triage, a feature that classifies new tickets by topic, sentiment, language and entities such as product names. Each classification comes with a confidence field, and human agents can correct the values.

Read the vendors' definitions of "resolved" before you read any resolution rate. Intercom counts a resolution when, after Fin's last answer, the customer confirms it helped or leaves without asking for more help, which it calls an "assumed resolution." Its documentation states plainly that a frustrated customer who just leaves after Fin's answer is billed as an assumed resolution. Zendesk counts a conversation where the customer did not ask for more help, then has a large language model read the transcript to check the request was actually resolved. Conversations that pass are "verified resolutions" and are billed; those that fail are "contained resolutions" and are not. The two definitions differ, so two vendors' rates cannot be compared directly. We recommend not treating either as a stand-in for customer satisfaction; your own sample review is a better check.

The strongest independent evidence covers AI that helps human agents rather than replacing them. Brynjolfsson, Li and Raymond studied a generative AI assistant rolled out to 5,179 customer support agents at one company. Access to the tool raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. The paper was published in The Quarterly Journal of Economics in 2025, but its data come from one firm's chat support, and it studies suggestions to human agents, not an agent answering customers on its own.

Jobs you can hire an agent for

Each job has its own page with a task breakdown, a brief you can copy, risks and a trial plan.

• JobSupervised
  AI support agent for tier-1 customer support

  An AI support agent can answer routine chat and email questions from your help content and hand the rest to your team. A person still owns refunds, policy exceptions and every escalation.
• JobSupervised
  AI phone agent for customer service: scope, cost, checks

  An AI phone agent can answer routine inbound calls, such as FAQs, bookings and status questions, and route the rest. A person still owns escalations, account changes that move money and the call scripts it follows.
• JobSupervised
  AI ticket triage: what an agent sorts, routes and misses

  An AI agent can read each new ticket, tag its topic, language and urgency, and route it to the right queue. A person still owns the categories, the tickets it cannot classify and anything flagged as a safety, legal or security issue.
• JobSupervised
  AI agent for knowledge base upkeep: scope, checks, cost

  An AI agent can find questions your help center does not answer, draft new articles and propose edits to stale, duplicate or conflicting ones. A person still decides what is true and approves every change before it goes live.

Ready

0of 4

Supervised

4of 4

• AI support agent for tier-1 customer support
• AI phone agent for customer service: scope, cost, checks
• AI ticket triage: what an agent sorts, routes and misses
• AI agent for knowledge base upkeep: scope, checks, cost

Assist only

0of 4

Not yet

0of 4

They are not equally ready, and none of them is a job to run without a person watching. The table shows what you would hand over in each job and what we recommend a person keeps. Each job page cites its own evidence.

Job	What you hand to the agent	What a person keeps
Tier-1 support	Answering routine questions over chat and email using your content, and handing off the rest	Refunds, exceptions, complaints and the review of a sample of answers
Ticket triage and routing	Classifying new tickets by topic, sentiment and language so your rules can route and prioritize them	Correcting wrong labels and setting the routing rules
Knowledge base maintenance	Drafting updates to help articles	Approving what gets published
Inbound phone support	Answering calls and handling simple requests by voice	Disclosure obligations, and every call it transfers

We recommend triage as a first job when the agent's labels are used only for internal routing, because then the customer does not see them. If your setup sends automatic replies based on a label, that no longer holds. A wrong label can send a ticket to the wrong queue, so have agents correct labels when they spot them and review a sample of labels each week. When a tier-1 agent is configured to send answers without human approval, a wrong answer reaches the customer before anyone on your team sees it. Phone adds its own legal questions, which the phone support page covers.

On disclosure, the EU AI Act requires providers to design AI systems meant to interact directly with people so that those people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person given the context (Article 50(1)). Article 50(5) requires the information at the latest at the first interaction. Rules elsewhere differ. Ask counsel which obligations apply to your role and jurisdiction before an agent talks to your customers.

Roles they affect

The role most affected is the customer service representative. Its page goes task by task through what an agent can take and what stays with the person.

• Role
  Can AI replace customer service representatives?

  Support products from Intercom, Zendesk and Gorgias document AI agents that answer questions, cancel or change orders and process refunds within rules you set; Zendesk and Intercom document handing unresolved requests to a person. Follow-ups, payments, coverage decisions and disputed goods have no cited agent evidence and stay with a rep.

The Bureau of Labor Statistics puts the median wage for customer service representatives at $21.53 an hour ($44,770 a year) in May 2025, with about 2.7 million jobs in 2025. BLS projects employment in the occupation to decline 5% from 2025 to 2035, and attributes lower demand partly to self-service systems and automation of tasks. It also projects about 289,500 openings a year over the decade, all from replacing workers who leave the occupation.

For a sense of the agent side, Intercom lists $0.99 per outcome for Fin over chat and email when used with Intercom, charged at most once per conversation; an outcome is a resolution (as defined above) or a configured procedure handoff. That price excludes Intercom's own seat plan, and Fin Voice is "contact sales." These are prices with different units, not a savings estimate. The AI agent vs employee cost guide shows how to compare them for your own volume.

Where to start

We recommend starting where mistakes are cheap and visible: ticket triage, then tier-1 answers on a narrow set of topics.

An example, with assumptions labeled. Assume, for this example, that a customer writes in by email asking how to change the delivery address on an order that has not shipped. The agent classifies the ticket as a delivery question with neutral sentiment, finds your help article on address changes and replies with the steps. The customer does not answer. A person on your team reviews a daily sample of such conversations. On this one they check that the article was the right one and that the order really had not shipped. If the customer had written "this is the third time I'm asking," the escalation rule would have sent it to a person instead.

Before you hand over any support job:

1. Fix your help content first. Intercom describes Fin answering from your available content. In our view, gaps and outdated articles in that content become wrong answers.
2. Write the escalation rules. Decide which topics always go to a person (refunds, cancellations, legal threats, safety), and make sure any customer who asks for a human gets one.
3. Define "resolved" yourself. Do not rely on the vendor's count. Review a sample of conversations the agent closed, each week, and track how many you would have answered differently.
4. Start narrow. Turn the agent on for a few topics or one channel and widen it only when the sample review holds up.
5. Know what data it touches. Depending on the job and the systems you connect, a support agent may see customer names, contact details, and order or account data. Treat these as examples to check. Check what the provider stores, where, and who can see transcripts.

If you would rather describe the job and have someone help you scope it, use the form below. The guide to hiring an AI agent covers scoping, trials and handover in more detail.

Sources

1. 1
   Fin AI Agent outcomes

   IntercomPrimaryAccessed Nov 14, 2026

2. 2
   About automated resolution tiers

   ZendeskPrimaryAccessed Nov 14, 2026

3. 3
   Automatically classifying tickets with intelligent triage

   ZendeskPrimaryAccessed Nov 14, 2026

4. 4
   Generative AI at Work (NBER Working Paper 31161; published in The Quarterly Journal of Economics, vol. 140(2), 2025)

   National Bureau of Economic ResearchPrimaryAccessed Nov 14, 2026

5. 5
   Customer Service Representatives: Occupational Outlook Handbook

   U.S. Bureau of Labor StatisticsPrimaryAccessed Nov 14, 2026

6. 6
   Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal L, 12.7.2024

   Publications Office of the European UnionPrimaryAccessed Nov 14, 2026

7. 7
   AI Act, Article 50: Transparency obligations for providers and deployers of certain AI systems (consolidated text)

   European Commission, AI Act Service DeskPrimaryAccessed Nov 14, 2026