AI in Customer Service: What an Agent Can Take On, and What It Can't
AI in customer service is about taking the repetitive, written enquiries and passing the rest on to a person straight away. Here is what an agent can take on, what it shouldn't, and how you get started.
Table of contents
In short: An AI agent (an AI system that can solve defined tasks) can in customer service sort enquiries, draft replies to practical questions, send status messages before the customer asks, and help with replies to reviews. Complaints, goodwill and anything that requires a judgement must go on to a person. Start with the repetitive, written enquiries, and measure how many are resolved correctly.
What is an AI agent in customer service?
Check whether your chosen solution can read enquiries, draft replies and look up order status with your data access, subscription and setup. The difference from an employee is that it should only take on what you have decided it may.
Three kinds of enquiries
A useful way to divide up customer service is by how much judgement the enquiry requires:
| Type | Examples | Who takes it |
|---|---|---|
| Practical questions | Opening hours, delivery, how to change an order | Can be tested: draft replies based on your own information |
| Lookups and changes | Order status, moving an appointment, new address | Can be tested: lookups and drafts, approved according to your rules |
| Judgements | Complaints, goodwill, dissatisfied customers, anything that requires discretion | A person, with the case gathered and summarised by the agent |
The line between the three is yours to set. Start cautiously, and only move one type over to the agent once you have seen it resolve it correctly.
What an agent can take on
Use the following as requirements for a pilot. For each function, check whether the solution supports it with your data access, subscription and setup.
Sorting and first reply
Check whether the agent can sort the inbox by type and urgency, draft replies to the practical questions and pass anything that requires a judgement on to the right employee with a short summary.
Status messages before the customer asks
When an order changes status in your system, for example ordered, dispatched or ready for collection, check whether the agent can send the corresponding message from a template you have approved. If the customer replies, the reply is treated like any other written enquiry.
Also agree how possible delays should be handled. Check whether an agent can keep an eye on orders that have sat too long in the same status, and either have a message ready for the customer or notify staff, as you choose. What counts as "too long" for an order is for you to decide.
Replies to reviews
If you have reviews on, for example, Google or Trustpilot, check whether AI can gather them in one place, draft replies in the company's tone based on your guidelines and distinguish between praise, specific criticism and the difficult reviews. The difficult ones are flagged to an employee with a calm draft. A person approves and posts every reply.
Summary and handover
When a case needs to go on to a colleague, check whether the agent can gather the history: what the customer has written, what has been answered, and what is still open. Check whether the summary gives the colleague enough context.
What an agent shouldn't do
- Decide complaints or goodwill. That is a judgement, and it belongs to a person.
- Promise anything nobody has approved. Times, prices and compensation go through an employee.
- Guess. If the answer is missing from your information, the agent should say so and pass the case on, not invent an answer.
- Hide that it is an AI. As a general rule, customers should know they are writing or talking to an AI.
The telephone
This guide is about the written workflows. If you are considering a solution where an AI answers phone calls, check how it discloses that it is an AI, how the call is handled and possibly recorded, and whether the call can be transferred to a person.
How to get started
- Find the repetitive enquiries. Go through a week's inbox, and count which questions recur.
- Gather the answers. The agent will be no better than the information it has: opening hours, delivery rules, returns policy and prices need to be up to date in one place.
- Run a pilot on one type of enquiry, where every reply is approved by a person.
- Measure the share of replies approved without corrections, how many cases were passed on, and what customers replied.
- Expand slowly, and only with the types the agent has shown it can handle.
Typical pitfalls
- Outdated information. If an outdated returns policy sits somewhere, the agent can answer wrongly. Keep the information up to date in one place.
- No route to a person. Make the handover clear in every channel, so the customer can get through to someone.
- Starting too broad. Start with one type of question, and expand once it has been tested.
- No follow-up on quality. Regularly review a sample of the replies, including those sent without corrections.
What to ask the supplier
| Question | Why it matters |
|---|---|
| Can it read your own information and systems? | Clarify what basis the answers are built on |
| Can we control what it may answer on its own? | You should be able to move the boundary gradually |
| How is a case passed on to a person? | Agree how the customer gets through to an employee |
| Where is customer data processed, and by whom? | If the supplier processes personal data on your behalf, a data processing agreement is required |
| What does it cost at your volume of enquiries? | The price may depend on users, usage and channels |
Customer data and rules
Customer enquiries contain personal data. If they are processed by an AI solution, there must be a legal basis, customers must be informed about the processing, and the data must be limited to what is necessary for the purpose. If the supplier processes the data on your behalf, you need a data processing agreement. Under Article 50 of the EU AI Act, providers of AI systems that interact directly with people must, as a general rule, ensure that the person is informed that it is an AI.
We cover how AI can be used in the rest of the sales and customer contact in AI in sales.
If you want help finding the enquiries where an AI agent can take work off your customer service team's plate, get in touch.
Sources
- The Danish Data Protection Agency, the basic principles: https://www.datatilsynet.dk/regler-og-vejledning/grundlaeggende-begreber/hvad-er-dine-forpligtelser/de-grundlaeggende-principper (accessed 12 October 2026)
- The Danish Data Protection Agency, You are also responsible when you share (data processors): https://www.datatilsynet.dk/regler-og-vejledning/gdpr-univers-for-smaa-virksomheder/trin-7-du-er-ogsaa-ansvarlig-naar-du-deler (accessed 12 October 2026)
- The Danish Agency for Digital Government, Transparency obligations for certain AI systems: https://digst.dk/tilsyn/ai-forordningen/reglerne-i-ai-forordningen/gennemsigtighedsforpligtelser-for-visse-ai-systemer/ (accessed 12 October 2026)
Frequently asked questions
Can AI replace customer service?
The guide's recommendation is to let AI take the repetitive, practical enquiries and let people keep complaints, goodwill and anything that requires a judgement. The agent passes those cases on with a summary.
Which enquiries can an AI agent resolve?
It depends on the solution. Try practical questions, lookups such as order status and changes according to your rules, and decide for yourself which types the agent may take on its own.
Do customers need to know they are talking to an AI?
Yes, as a general rule. Article 50 of the EU AI Act requires the provider to ensure that the person is informed that it is an AI, unless it is obvious.
Can AI reply to reviews?
Check whether the solution can draft replies in the company's tone and flag the difficult reviews to an employee. A person approves and posts every reply.
What happens when the agent doesn't know the answer?
It should say so and pass the case on to an employee with a summary, not invent an answer. Set that up from the start, and test it in the pilot.
What should you start with?
Go through a week's enquiries, find the ones that recur, and run a pilot on one type where every reply is approved. Measure how many were approved without corrections.
Tags: Customer Service, Efficiency, AI, Guide, Scaling
Contents
How many of your enquiries are the same ten questions?
We look at what can genuinely be answered automatically for you, and what cannot. An honest answer, even if it is not much.
Book a reviewHow many of your enquiries are the same ten questions?
We look at what can genuinely be answered automatically for you, and what cannot. An honest answer, even if it is not much.
How we built it · Email & prioritisation
The inbox, sorted
An inbox where AI sorts the emails, writes the reason next to each email and can draft a reply based on the customer’s case.
- Email in
- AI sorts
- You reply
Related articles

Latency in Sales Calls: Why 500ms of Delay Costs You Orders
Sales calls with low AI latency are completed far more often. What delay costs in sales, and what the technical standard is in 2026.

Voice-to-CRM: how automatic call data transforms RevOps in 2026
Voice-to-CRM: how sales conversations become filled CRM fields without manual entry. The method, the pitfalls, GDPR and what a Danish business can have built.

Sales budget with AI: how to set one that holds
How to set a sales budget and forecast that holds: the three classic errors, a simple method, pipeline checks, and where AI can help in forecasting.