AI in customer service: what it handles, what it costs and how to start

AI works well in customer service for recurring questions with clear answers, for sorting incoming requests and for giving staff ready-made draft replies. It works poorly for complaints, exceptions and anything that requires a decision. For a small or mid-sized business the best start is usually small: a limited solution for the most common questions, with a clear route to a human and simple follow-up on what it actually resolves.

What AI handles well

Today's language models understand questions written in everyday language, typos and half sentences included. They are best suited where the answer is documented and the same for everyone:

• Common questions. Delivery times, return policies, opening hours, payment options and size guides. In most businesses the same questions come back again and again.
• Order status and bookings. If the AI assistant is connected to your order or booking system, it can tell customers where a parcel is or when an appointment is booked. Without that connection it can only refer them onwards.
• Sorting and summarising. AI can read incoming emails, judge what the request is about and how urgent it is, and route it to the right person with a short summary.
• Draft replies for staff. Staff get a suggestion based on your routines and previous replies and adjust it before sending. This is often the safest place to start.
• Evenings, weekends and multiple languages. Simple questions can be answered outside office hours and in more languages than you have staff for.

What should stay with people

An AI model can answer confidently and still be wrong. It might promise a return right you do not offer or invent a delivery time. Your company is responsible for what the chat says, just as it is for what an employee says. So let people handle:

• complaints and warranty claims, especially when the customer is upset
• exceptions, compensation, discounts and other decisions that cost money
• matters involving sensitive information, such as health or personal finances
• key customer relationships, where the customer expects to talk to someone they know
• anything where a wrong answer could have legal or financial consequences

Be clear that the customer is talking to an AI, and make it easy to reach a person. It is good customer service, and the EU AI Act also requires that people are informed when they are interacting with an AI system.

What it costs

The cost depends more on how closely the solution is connected to your systems than on the AI technology itself. Roughly, there are three levels:

1. A ready-made AI feature in your existing support tool. Many help desk and chat tools offer AI as an add-on, charged per user per month or per resolved conversation. Lowest starting cost, but limited room for customisation and hard to connect to your own systems.
2. A custom AI assistant with your own knowledge base. The assistant is built on your terms, product data and routines and placed on your website or in your support flow. It is a well-defined development project, which we normally deliver at a fixed price after a short pre-study.
3. AI connected to orders, CRM or booking systems. Here the assistant can look things up and in some cases act on the customer's behalf. Greatest benefit, but also the largest project, since every connection has to be built, tested and kept running.

On top of that come running costs: the fee to the AI provider, charged by usage, and time to keep the knowledge base up to date. The latter is often underestimated. If your terms change and nobody updates the material, the AI keeps answering according to the old ones.

GDPR and what happens to customer data

When a customer writes to an AI assistant, the text is sent to an AI service for processing. That service is often run by a provider outside the EU. Make sure you have the following under control:

• Data processing agreement. The AI provider processes personal data on your behalf, which requires a data processing agreement.
• Transfers outside the EU. If data is sent to, for example, the US, there must be a valid legal basis for the transfer. Providers offering processing and storage within the EU make this simpler.
• Training on your data. Check that the provider does not use your conversations to train its models. Business agreements usually exclude this, but it should be in writing.
• Retention. Decide how long chat logs are kept and where.
• Data minimisation. Do not send more than necessary. Personal identity numbers, card details and the like should never need to pass through the AI service.
• Informing customers. Describe the processing in your privacy policy.

Start small

1. List the 10–20 most common questions. Go through recent months of support requests and see what keeps coming back.
2. Write down the answers. An up-to-date knowledge base is the foundation. It often turns out the answers were never written down, and that alone has value.
3. Start internally. Let the AI draft replies for your staff for a few weeks. You will see how it answers without customers being affected.
4. Run a customer-facing pilot in one channel, such as the chat on your website, with a clear handover to a person.
5. Expand only when the results hold up.

How to measure whether it works

Measure before you start so you have something to compare against. Then follow up on:

• Resolved requests. How many are closed without a person having to step in? Check that the customer does not come back about the same thing – otherwise the issue was not resolved, just moved.
• Incorrect answers. Read a sample of conversations every week during the pilot. This is the most important quality measure.
• Response time and customer satisfaction. Compare with how things were before.
• Staff time. How much time is freed up, and what is it used for?

Our recommendation

Choose a ready-made AI feature in your support tool if your questions are simple and the tool you already use offers it. Choose a custom solution if the answers depend on your own data, if you want control over where customer data ends up, or if the AI needs to connect to order or booking systems. Whichever route you take: start narrow, measure and let people handle the difficult cases. We build custom AI solutions and connections to existing systems, and you can read more about how we work under AI integrations.