Most support teams answer the same twenty questions all day. Where is my order? Can I change the delivery address? Can you send the invoice again? In between sit the messages that really need a person: a broken product, a double charge, a customer who has had enough.
AI can take a large share of the first kind off your team. It can also make the second kind worse if you let it answer everything. This guide covers what AI does well in support today, what to keep human, the order to roll it out in, and how to tell whether it’s working.
Six jobs AI can do in a support team
They’re listed roughly from the lowest risk to the highest. The first two never talk to a customer on their own, which is why they’re the safest place to start.
1. Sorting and routing incoming requests
Every email, chat message or form gets read the moment it arrives. The AI tags the topic (delivery, returns, billing, technical fault), the language and the urgency, and pulls out details such as the order number. Plain rules then send it to the right queue.
This is a classic AI workflow: one AI step inside an otherwise fixed process. Our older post on workflows, AI workflows and agents walks through this exact example. For an e-shop selling into Slovakia, Czechia and Hungary, simply detecting the language and routing the message to the right person saves a surprising amount of forwarding.
2. Suggested replies for your team
The AI drafts an answer from your help articles, your policies and the customer’s order data. A person reads it, fixes what’s wrong and sends it. The customer gets a reply a person has checked, and your colleague skips most of the typing.
This is where the best evidence is (see the numbers below). It helps new team members most, because good drafts carry the way your most experienced people answer.
3. Self-service answers from your knowledge base
A chat on your website or in your customer portal answers questions from your own help articles and links the article it used. This works through retrieval: the system first finds the relevant passages, then the model writes an answer from them. We explain the mechanics in RAG explained.
Two conditions. The articles must be correct and current, and the assistant must be allowed to say “I don’t know” and pass the customer on. An assistant that confidently quotes last year’s return policy is worse than no assistant at all.
4. Fully automating simple requests
Some requests have one right answer, and it sits in one of your systems: order status, a copy of an invoice, a new delivery slot, opening hours over a holiday. Here the AI can do the whole job, because it looks the answer up through a tool (a defined connection to your order system or ERP) instead of guessing.
- You: Hi, I need a copy of the invoice for my coffee machine order.
- Tool call: orders.find(email: customer, product: coffee machine)
- Result: order 20931 · delivered 3 June · invoice INV-2026-1184
- AI: Here it is: invoice INV-2026-1184 for order 20931. I’ve also sent it to your email. Anything else?
- You: Yes. The machine has been leaking since the second week and I want my money back.
- Complaints and refunds go to a person
- Peter: Hi, I’m Peter from the service team. I can see your order and what you wrote. Could you send a photo of where it leaks?
The important part is what the assistant is not allowed to do. It can read orders and resend documents. It can’t approve a refund, promise a replacement or argue about a warranty.
5. Summaries for hand-offs
When a conversation moves to a person, or from first-line support to a specialist, the AI writes a short summary: what the customer wants, what has been tried, what was promised and by when. Nobody has to scroll through 30 messages, and the customer doesn’t have to explain everything again.
6. Quality monitoring
Most teams review a small sample of conversations by hand. AI can read all of them and flag the ones worth a look: an unhappy customer, a promise that wasn’t kept, a reply that contradicts your policy, the same product fault mentioned ten times this week. A person still decides what to do about each flag. The product fault pattern alone can be worth more than the time saved on replies.
Which requests to automate first
Export a few months of tickets and sort them by type. Two things decide where each type belongs: how many of them you get, and how much judgement one takes.
Volume against complexity
- 1 Where is my order?
- 2 Change of delivery address
- 3 Invoice copy
- 4 How do I use the product?
- 5 Technical fault
- 6 Wholesale price request
- 7 Complaint about damage
- 8 Refund dispute
Before you automate a high-volume request, ask why customers need to ask at all. If a third of your tickets are “where is my order”, a tracking page, an SMS when the parcel ships or a client portal with live order status may remove those tickets entirely. That’s cheaper than answering them faster, with or without AI.
The bottom-left corner is a trap of its own. A request that comes in twice a month doesn’t justify building, testing and maintaining an automation, however easy it looks.
What the numbers say so far
There isn’t much solid public data on AI in support yet, and most of it comes from vendors. Two sources are worth knowing.
Two data points
- 15% more issues resolved per hour, on average, when support agents got an AI assistant that suggested replies
- 2/3 of Klarna’s customer service chats handled by its AI assistant in the first month
The study found the biggest gains among less experienced agents. The most experienced ones got slightly faster, and their quality dipped a little. So suggested replies are a training tool as much as a speed tool.
Klarna is the other half of the story. In 2025 its chief executive told Bloomberg that cost had been too dominant a factor and the result was lower quality, and the company started hiring people for customer service again. Automating most chats is possible. Doing it mainly to cut headcount tends to show in the customer experience.
What to keep human
Keep a person in charge of:
- Complaints and angry customers. A model can be polite. It can’t take responsibility.
- Refunds, credits and goodwill gestures above a limit you set. Below the limit, a plain rule may be enough.
- Vulnerable customers. Bereavement, illness, financial difficulty. Detect the signals and route them to a person straight away.
- Legal threats and data protection requests, and anything else that may end up with a lawyer.
- Key B2B accounts, where the relationship matters more than speed.
- Anyone who asks for a person. Make that easy, not a test of patience.
Who checks what, and when the line can move as trust grows, is covered in human in the loop.
How the pieces fit together
One request, start to finish
- Message arrives Email, chat, form or phone
- AI sorts it Topic, language, urgency, order number
- Rules pick the path Simple and known: answer. Anything else: a person
- AI answers or drafts From help articles and order data
- A person handles the rest With a summary of the conversation
- Logged and measured Outcome, satisfaction, reopened or not
The same flow works for email, chat and the phone. On the phone, an AI voice agent does the listening and talking; the rules and the hand-off stay the same. A clinic or a property manager that takes most requests by phone may want to start there. Our page on AI voice agents shows how a call moves through these steps.
Buy or build?
Many helpdesk tools, such as Zendesk, Freshdesk, Intercom or Gorgias, now include AI features: suggested replies, summaries, a chatbot over your help centre. If your requests are standard and your data lives in the helpdesk, try those first. They’re quick to switch on and easy to switch off.
A custom build makes sense when the answers depend on your own systems: a custom order system, an ERP, a booking calendar, a portal your clients log into. Or when you want one set of rules across email, chat and the phone. In practice it’s often a mix: keep the helpdesk your team knows and add the connections and rules it lacks.
Customer data and privacy
Support conversations are full of personal data: names, addresses, order histories, sometimes details about health or money. Before any of it goes to an AI service, check three things: where the data is processed, whether the provider may use it to train its models, and how long conversations are kept. Business and API terms usually exclude training on your data, but read them. Most large providers can process data in the EU, directly or through the big cloud platforms.
Tell customers when they’re talking to an AI assistant. Since August 2026 the EU AI Act requires it for systems that talk to people directly, unless it’s obvious from the context. And keep card numbers and passwords out of chats altogether: send a secure link instead.
How to measure it
Measure before you switch anything on. Without a baseline you’ll end up arguing about impressions.
- First response time and time to resolution, per channel and request type
- Automated resolution rate, counted honestly. A conversation counts as resolved only if the customer doesn’t come back about the same thing within, say, seven days.
- Reopen and escalation rates
- Customer satisfaction (CSAT), the short “how did we do?” survey, split between automated and human conversations
- Handle time for your team on the tickets they still touch
- Cost per resolved request, including the AI running costs and the time people spend checking AI output
Be careful with “deflection”, a favourite vendor metric. A customer who gave up on the chatbot and left counts as deflected in some dashboards. That isn’t a resolved request. It may be a lost customer.
A sensible rollout order
An example rollout over about 12 weeks
- 1
Weeks 1–2
Read your ticketsCount request types, set the baseline, fix or delete outdated help articles.
- 2
Weeks 3–5
Sorting and suggested repliesThe AI tags and routes every request and drafts answers. People still send everything.
- 3
Weeks 6–9
Automate one simple requestOrder status, for example, with access to the order system and an easy way to a person.
- 4
Weeks 10–12
Summaries, monitoring, decisionAdd hand-off summaries and quality checks. Compare with the baseline, then widen or stop.
The order matters. Sorting and suggested replies teach you where the AI is wrong while a person still checks every answer. By the time you automate anything, you know which request types it handles reliably and which it doesn’t.
What to do next
- Export three months of tickets and count your ten most common request types.
- Write down the baseline: response time, resolution time, satisfaction, cost per request.
- Read your help articles. Fix or delete anything outdated before an AI starts quoting it.
- Start with sorting and suggested replies. Automate one simple request only once those work.
- Decide, in writing, which cases always go to a person.
If you’d like a second opinion on which of your requests to automate first, send us a short description of your support setup. We’ll tell you where we’d start and what we’d leave alone.
