# AI customer service agent

> An AI customer service agent for email, web chat and WhatsApp: it resolves with your knowledge base, triages tickets and escalates.

URL: https://gradual.pro/en/use-cases/ai-customer-service

Updated: 2026-09-29

Answers email, chat and WhatsApp with your company’s information and passes the tricky cases to your team.

- One agent handles email, web chat and WhatsApp with the same answers.
- It answers from your documentation (RAG) and real customer data, not invented replies.
- It classifies and prioritises every ticket and assigns it to the right person when it can't resolve it.
- It works inside your current helpdesk: Zendesk, Freshdesk, HubSpot, Intercom or Gorgias.

## What does an AI customer service agent resolve?

It resolves the repetitive queries that take up most of your support team's day: where is my order, how do I change my plan, what does the warranty cover, how do I download my invoice. In a small business these are often half of all tickets or more.

To answer well, the agent combines two sources. The first is your knowledge base: manuals, policies, terms and the answers your team already uses. The second is customer data, which it looks up in your store, ERP or CRM: order status, tracking number, renewal date. With both, it replies with specifics rather than "please see our terms".

## How does the agent avoid making up answers?

We use a technique called RAG: before answering, the agent searches your documentation for relevant passages and replies only with what it finds. If it finds nothing reliable, it does not improvise: it creates the ticket, tells the customer a person will reply and leaves a draft ready. There is more detail in what [RAG](/en/glossary/rag) is.

We also limit actions: the agent can look up an order, but a refund or compensation always goes through human approval. During the pilot we review a sample of conversations every week and extend the knowledge base with the questions that went unanswered.

## Ticket classification and routing

Even when the agent does not resolve a case, it saves work. Every incoming ticket gets tagged by topic, urgency, language and sentiment, and is assigned to the queue or person that matches your rules.

The agent also summarises the conversation and attaches the customer's data, so whoever receives it starts working without asking the same questions again. Complaints with a very negative tone or customers flagged as key accounts can jump straight to a manager.

## Example: a sporting goods store with sale-season peaks

An online sporting goods store receives about 60 tickets a day, tripling during sales and at Christmas. Three people handle email and chat in Zendesk, and at peaks replies take two or three days.

The AI agent connects to Zendesk, Shopify and the carrier's tracking. When a customer writes "my order hasn't arrived", the agent identifies the order by email, checks tracking and replies with the real status and estimated date. If the parcel has been stuck for days, it opens a claim with the carrier and tells the customer. It prepares returns according to policy and leaves the refund pending approval. Unusual cases, such as a defective product with photos, reach the team already classified and summarised. More ideas for stores in [AI for ecommerce](/en/industries/ecommerce), and if your main channel is WhatsApp, see the [WhatsApp AI agent](/en/use-cases/whatsapp-ai-agent).

## Process

1. **We analyse your tickets**: We export a sample of tickets from recent months and group the reasons to see what can be resolved automatically.
2. **We prepare the knowledge base**: We organise manuals, policies and saved replies, and spot what is missing or out of date.
3. **We connect helpdesk and data**: We integrate Zendesk, Freshdesk, HubSpot or another tool, and the customer data sources: store, ERP or CRM.
4. **We define limits and escalation**: Which actions the agent can run, which need approval and how cases are assigned to the team.
5. **Pilot by channel**: We start with one channel, measure resolution and satisfaction, then extend to the others.

- of tickets resolved without human involvement: 40-70%
- first response in chat and WhatsApp: < 1 min
- of tickets classified and assigned on arrival: 100%

Typical ranges in support projects; they depend on how repetitive the queries are and on the quality of your knowledge base.

### Rule-based chatbot, AI agent and human team

|  | Rule-based chatbot | AI agent | Human team |

| --- | --- | --- | --- |

| Understands free-form questions | Poorly | Yes | Yes |

| Looks up orders or contracts | Limited | Yes | Yes |

| Handles volume peaks | Yes | Yes | With delays |

| Complaints and exceptions | No | Escalates with a summary | Yes |

| Cost per ticket | Very low | Low | High |

Tools: OpenAI, Anthropic Claude, Zendesk, Freshdesk, HubSpot Service Hub, Intercom, Gorgias, WhatsApp Business API, n8n, Supabase

## FAQ

### How much does an AI customer service agent cost?

It depends on the number of channels, integrations (helpdesk, store, ERP), ticket volume and languages, plus ongoing maintenance. We budget per phase with a fixed price agreed before we start. AI usage is paid directly to the provider.

### Does it work with my current helpdesk?

In most cases, yes. We work with Zendesk, Freshdesk, HubSpot, Intercom, Gorgias and any tool with an API. The agent writes inside the tickets, so your team keeps working in the same screen.

### Do customers get annoyed talking to an AI?

They get annoyed when the AI fails to resolve the issue and won't let them reach a person. If the agent replies with real data in seconds and offers a handover as soon as the customer asks, the experience usually beats waiting two days for an email.

### What is the difference from the bots built into Zendesk or Intercom?

Native assistants work well with the documentation already inside that tool. We build the agent when you need to query external systems (store, ERP, carrier), run actions or unify several channels. Sometimes the best option is to configure the native assistant well, and we tell you so.

### How long until it is running?

A pilot on one channel is usually in production in 4-6 weeks. The timing depends mostly on the state of your knowledge base: if answers are scattered across emails and people's heads, we spend the first week organising them.

### AI chatbot vs live chat: which is better for customer support?

They work best together. The AI agent answers instantly at any hour and resolves the routine share of tickets; live chat with a person covers complaints, exceptions and high-value customers. The agent hands the conversation over with a summary so nobody starts from zero.