AI customer support agent: how we automate 80% of requests

Most support teams spend their days on the same few dozen questions: prices, delivery times, opening hours, order status, how-to steps. An AI customer support agent can take these off the team — but only if it's built right. A badly built bot invents answers, frustrates customers and creates more work than it saves.

This article shows how we build AI support agents that can resolve up to 80% of requests on their own, what the remaining 20% look like, how to build the knowledge base behind them, how to calculate the business case for your own team — and when you shouldn't use an AI agent at all. The headline numbers come from a live project.

What is an AI customer support agent?

An AI support agent is an assistant that answers customer questions in your channels — website, WhatsApp, Telegram, email — using your own knowledge. Unlike the rule-based chatbots of the past, it understands free-form questions, finds the relevant information in your documents and writes a natural answer. Unlike a generic AI chatbot, it's restricted to what your company actually knows and hands over to a human when needed.

How does an AI agent resolve up to 80% of requests?

The key is not the language model — it's everything around it. Every request goes through the same steps:

1. Understand the question — including typos, mixed languages and several questions in one message.
2. Search your knowledge base — documents, FAQs, price lists, policies, past conversations.
3. Draft an answer based only on what was found, in your brand's tone.
4. Check confidence — is the answer fully supported by the sources? Is the topic sensitive?
5. Answer or hand over — if in doubt, the agent passes the conversation to your team, with a summary.

This approach is called retrieval-augmented generation (RAG), and it's why we describe our agents as RAG chatbots rather than "AI that knows everything". It's what stops the agent from making things up: no source, no answer.

What does a good handoff look like?

The handoff is where customers decide whether they trust your AI support. Here is a short, illustrative example — not a transcript from a real customer — of a conversation that starts as a routine question and ends with a person:

Customer: Hi, how long does delivery to Austria take? And my last order arrived with a broken lid.

Agent: Delivery to Austria usually takes 3–5 business days. I'm sorry about the lid. Damaged items are handled by our service team so they can arrange a replacement or refund — I'm passing your conversation to them now. Could you share your order number and a photo of the damage, so they don't need to ask again?

Customer: Order 10482, photo attached.

Agent: Thank you. A colleague will take over here during business hours today. You won't need to repeat anything.

What makes this a good handoff: the agent answers the part it can (the delivery time from the knowledge base), recognises the part it shouldn't handle alone (a damage claim), collects what the team will need, and sets an honest expectation. The team receives the order number, the photo and a one-line summary. (The delivery time is a placeholder; in practice it comes from your shipping policy.)

How do you build the knowledge base for an AI support agent?

Before any AI is involved, we run a knowledge audit. In our experience, most answer-quality problems are knowledge problems, not AI problems — this phase decides more than the choice of model.

What to collect

Real past conversations are the single best source: they show how customers actually phrase questions, which topics come up most and where your written documents are silent. Ask the team for the "everyone knows this" answers and write them down.

How to structure price lists and tables

Price lists are where generic setups fail most often. A PDF table split into random text chunks loses the link between a product, its variant and its price — and the agent answers with the wrong row. We convert price lists into clean, one-item-per-entry records: product, variant, price, currency, whether tax is included, validity date and any conditions. If prices change often, the knowledge base should read them from the system where they're maintained rather than from a copy.

How to handle contradictions

Almost every company has them: the website says 14 days for returns, an old PDF says 30, the team does whatever the customer needs. Before launch we list every contradiction we find and ask one owner to decide which version is valid. Outdated documents are removed, not just "deprioritised". Each source gets an owner and a review date, so the next contradiction is caught by a person, not by a customer.

What else does it take to build one?

Retrieval, tone and boundaries

Documents are split into meaningful sections and indexed so the agent finds the right paragraph, not just the right document. We define how the agent speaks — formal or casual, short or detailed — and what it must never do: promise discounts, give legal advice, discuss competitors. These rules are tested, not just written down.

Integrations

Answering "where is my order?" requires access to the order system; rebooking an appointment requires the calendar. We connect these through APIs or automation tools — our comparison of n8n, Make and Zapier covers the options.

Testing on real questions

We run the agent against hundreds of real, anonymised customer questions and grade the answers before launch. Only when accuracy and tone hold up does it go live.

Which requests can AI handle — and which should go to a human?

The 20% that go to people are exactly the conversations where people make a difference. That's the real benefit: your team stops answering "what are your prices?" and focuses on the cases that keep customers loyal.

How do you calculate the business case for automating customer support?

You don't need a vendor's promise to estimate the value — you need four numbers from your own support data:

Hours saved per month = requests per month × share automatable × minutes per request ÷ 60

Value per month = hours saved × fully loaded hourly cost of a support agent

An illustrative example with hypothetical numbers — not a client result:

Three things keep the estimate honest. First, don't plug in 80% just because our project reached it — tag a sample of a few hundred real requests and count how many could be answered from documented knowledge. Second, subtract the running costs: the model and hosting, plus the team's time for reviewing answers and updating the knowledge base. Third, count the benefits that don't show up in hours: answers outside business hours, no queue at peak times, and a team that spends its time on the harder 20%. For a quick hours-based estimate, use our ROI calculator with the hours you expect to save and your hourly rate.

How do you measure success?

We track four numbers from day one:

• Resolution rate — share of conversations closed without a human. In our live project: 80%.
• Customer satisfaction — rating at the end of the conversation. In our project: 70% positive.
• Handoff reasons — why the agent passed a case on. This is the to-do list for improving the knowledge base.
• Answer accuracy — sampled answers checked by the team every week.

Resolution rate alone can be misleading — a bot that never hands over has a "100% resolution rate" and angry customers. That's why we always look at it together with satisfaction and handoff reasons.

Can one AI agent work across channels and languages?

Yes. One knowledge base serves website chat, WhatsApp, Telegram, Instagram, email and voice. What changes per channel is the format: short messages without tables in WhatsApp, a complete structured reply by email, no links to click in voice.

Multilingual support works the same way. Your knowledge base can stay in one language; the agent answers in the customer's language. Two details matter: prices, legal terms and product names should be checked in every language you serve, and handoff summaries should be written in the language your team works in, even when the customer wrote in another one.

When should you not use an AI support agent?

An AI agent isn't the right answer for every team. In our experience, it's usually not worth it when:

• Volume is low. If the team answers a handful of requests a day, a well-maintained FAQ page may do the job.
• Almost every request is individual. Custom projects, negotiations and complex B2B cases rarely repeat enough to automate.
• The knowledge isn't written down and nobody will own it. Without a maintained knowledge base, accuracy drops quickly.
• The answers are regulated advice. Legal, medical or financial recommendations belong with qualified people.
• Nobody can take the handoffs. An agent that hands over to an empty inbox is worse than no agent.

What are the most common mistakes?

• Letting the model answer from general knowledge. It will sound confident and be wrong about your prices. Restrict it to your sources.
• No way out. Customers who can't reach a human leave. Make handoff easy and visible.
• Hiding that it's AI. Since August 2026, the EU AI Act requires chatbots to disclose that users are talking to an AI — see our Article 50 guide (not legal advice). Customers don't mind; they mind slow or wrong answers.
• Launch and forget. Products, prices and policies change. Without updates, accuracy drops within weeks.
• Measuring only cost savings. If satisfaction drops, savings disappear in churn.

How long does it take and what do you need?

A first agent for one channel is typically live within a few weeks. From your side we need access to your documentation and past conversations, one person who knows the answers and can review test results, and access to the channels and systems the agent should work with. If you handle sensitive customer data, the agent can also run on a local model in your own infrastructure. More about our approach on the AI agents and chatbots page.

Want to know what share of your requests an AI agent could resolve? Send us a sample of anonymised questions and book a free call — we'll give you an honest estimate.