AI customer support agent: automate tickets without losing quality

How an AI support agent differs from a chatbot

A scripted chatbot follows pre-written branches, while an AI agent understands free-form questions, searches your knowledge base and can take actions in connected systems. That means you can delegate more to it, and that its mistakes cost more.

A working agent has four parts:

• a knowledge base of policies, terms, guides and common answers;
• search over it, so answers rest on a document and not on model memory (see our RAG explainer);
• tools to check an order, book a slot or open a ticket through your systems' APIs;
• routing rules that decide when a conversation goes to a person, and to whom.

In August 2026 Gartner advised using generative AI for intake and intent recognition, attempting resolution only when confidence is high and keeping a clear route to a human. That is a sound frame for a first launch.

Which tickets should you automate first?

Start with frequent questions that have a verifiable answer in your knowledge base, and leave money, complaints and exceptions to people. It helps to sort tickets by the cost of an error, not by topic.

To pick a first scenario, export recent tickets and label them against this table. A strong candidate is a topic that comes up often, gets resolved the same way each time and is already documented. If the policy isn't written down, write it first: an agent won't fix a knowledge gap, it will only expose it faster.

How should handoff to a human work?

A good handoff is visible to the customer, happens immediately and keeps the context, so the human continues the conversation and the customer doesn't repeat themselves. This is where automation earns or loses trust.

In Gartner's survey of 3,566 B2B and B2C customers (February–March 2026), 50% said generative AI makes interactions easier, yet 87% said an option to reach a human is essential. Customers unwilling to use AI most often said they would change their mind if they could switch to a person when needed. Gartner's advice is direct: don't make generative AI mandatory, and don't force customers through several failed attempts.

Define the handoff triggers in advance:

• the customer asks for a person, with no pushback or repeat offers;
• confidence falls below the threshold, or the knowledge base has no answer;
• the agent misunderstands the question again;
• the customer is frustrated or raises a complaint;
• the topic is on your sensitive list;
• the action exceeds the limits you gave the agent.

The human agent should receive a short summary, the conversation history, customer data, the sources found and what the AI already tried. Outside staffed hours, the AI should state a realistic response time and create a ticket, not promise what nobody can deliver.

Which metrics actually show quality?

The main metric is not the share of conversations without a human but the share of issues resolved with no repeat contact. Containment can climb simply because customers gave up.

In its analysis of voice assistants, ML6 warns that containment rewards the wrong behavior, since a bot that leads a caller into a dead end still counts as a success. A more honest measure is contained resolution: the AI handled the contact alone, the customer's intent was resolved, and no repeat contact followed within a defined window, commonly seven days.

How many tickets can realistically close without a human? There is no universal number. Gartner forecasts that by 2029 agentic AI will resolve 80% of common issues and cut operational costs by 30%, but that is a forecast limited to common issues. Klarna reported that in February 2024 its assistant handled 2.3 million conversations in its first month, about two-thirds of chats, with average resolution under 2 minutes. W2W's pilot target for its support skill is up to 70% of requests without an operator; the site labels this a target based on industry data, not a result of its implementations. Your number depends on your share of routine questions, the completeness of your knowledge base and your integrations, so a pilot has to establish it.

Where AI agents fall short

An agent can be confidently wrong, your company answers for what it says, and people can try to manipulate the model. These risks don't rule out automation, but they set its limits.

• Liability for answers. In February 2024 the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada. A chatbot misdescribed the airline's bereavement fare terms, and Air Canada argued the bot was responsible for its own actions. The tribunal disagreed: the bot was part of the website, and the airline answered for its content. It awarded about CAD 650 plus interest and fees. This is a Canadian tribunal decision, not a rule for other countries, but the lesson is clear: check answers about terms and prices with particular care.
• Over-weighting cost. In May 2025 Klarna's CEO said the company had gone too far and that a focus on cost had lowered quality. The assistant still handles about two-thirds of inquiries, but Klarna now wants customers always to have the option of a person and is recruiting support staff again.
• Manipulation. OWASP ranks prompt injection first among its 2025 LLM application risks: text from a customer or an external document can make a model behave in unintended ways. Defenses include least privilege, keeping sensitive functions in application code rather than under model control, and human approval of high-risk actions.
• Stale knowledge. If terms change and the document doesn't, the agent will be confidently wrong. You need a knowledge base owner and regular runs against a test set.
• Languages. If customers write in Russian, Uzbek, English or a mix, test quality in each language separately. A good result in one tells you little about another.

How Company Brain handles this

The Customer support skill in Company Brain answers from your knowledge base with a source link and passes anything it doesn't know to an employee with the full conversation context. It works in Telegram, WhatsApp, on your website and by email; the other scenarios are on the Company Brain skills page.

What the site states:

• when confidence falls below the threshold, it says "I don't know" and calls in a person;
• access rights apply before an answer is generated, and every request and source is logged;
• quality, tone, boundaries and escalation scenarios are tuned during the pilot;
• acceptance uses 50 real questions and a threshold of 85% acceptable answers, meaning correct in substance, citing a source and containing nothing outside the asker's access rights; below that, W2W keeps improving at its own expense;
• a dashboard covers requests, topics, answer quality and knowledge gaps;
• timeline: prototype in 2 weeks, pilot of 2–4 weeks, production launch in 6–10 weeks;
• pricing: implementation from $5,000, running costs from $500 per month in the cloud.

Where to start

Pick one ticket topic, collect real questions for it and agree on the metric before launch. You test the idea at small scale and get numbers from your own data.

1. Export your tickets and label them by type and cost of error.
2. Choose one scenario where the answers are already documented.
3. Clean up the knowledge base and assign it an owner.
4. Write down the handoff triggers and what the agent must never do.
5. Build a test set from real questions and set a quality threshold.
6. Roll out to part of your traffic and compare resolved tickets, repeat contacts and customer ratings.

Get a preliminary budget and timeline from the implementation calculator, or walk through your ticket data with us: book a meeting.