Contents9
- 01What can AI do with an incoming ticket or email?
- 02Tickets or emails: what changes?
- 03Which replies can go out without review?
- 04Why keep a human in the loop?
- 05How much does AI sorting cost, and how do you keep it low?
- 06What we run ourselves
- 07What to set up before going live
- 08Where to start
- 09Frequently asked questions
Automating customer support does not mean replacing the team that answers. The surest gain comes before the reply: reading every incoming ticket or email, understanding what it asks, finding the customer and their order, looking up the right information in your documents. AI does this work in a few seconds and prepares a draft; a member of the team reviews it, corrects it if needed, and sends it. This guide in our Automation and bots section explains how to split the roles, and what each step costs.
In short
AI reads, sorts and prepares. A rule sends what is simple. A person approves what commits the company.
Tickets or emails, the principle is the same: an email just needs one extra step to become a clean ticket.
Not every step needs AI. An acknowledgement of receipt or an order status is handled by a rule, with no paid call.
With the right model at each step, sorting a few thousand messages a month costs less than a dollar. The cost that matters remains review time.
You start with everything under review, and only open automatic sending to the categories that have earned it, backed by figures.
What can AI do with an incoming ticket or email?
A customer message goes through five steps. AI handles two, rules handle two others, and the last goes to your team as soon as the reply commits to something.
Who does what, from message received to reply sent
- Receipt, by a rule
The message arrives, an acknowledgement goes out, a ticket opens.
- Reading, by AI
Subject, urgency, language, tone, order number, several requests in one message.
- Context, by a rule
The customer, their orders and past exchanges are found in the CRM or the online store.
- Draft, by AI
A reply written from your documents, with the source of each piece of information.
- Sending, by your team or a rule
Reviewed and sent, or sent on its own if the category allows it.
Reading is the step where AI pays off the most. A customer writes “the box arrived open, the instructions are missing and I would like to know if I can change the size”: three requests in one sentence, none in the vocabulary of your categories. Keywords classify this kind of message badly; a language model breaks it down, files it and flags the irritated tone. This is exactly the case described in our article on automation without AI: AI is justified when the input is free text, and not elsewhere.
The draft comes next, and it is only as good as its sources. A model that answers from memory invents plausible return conditions; a model that first searches your terms and conditions, your FAQ and your procedures quotes what you actually wrote. That is the principle of an AI assistant that answers from your documents. When the information is missing, a good draft says so, instead of filling the gap.
So the support agent who opens the ticket finds the request summarised, the customer identified, their order displayed and a reply ready to review. They keep their time for what no tool does in their place: judging whether the reply suits this customer, today.
Tickets or emails: what changes?
The principle does not change, the starting point does.
With a ticketing tool (Zendesk, Freshdesk, Zammad, or a CRM such as Twenty that handles requests), the message already arrives filed in a queue, with fields. AI fills in those fields (category, priority, language) and drops its draft in as an internal note, which the agent turns into a reply with one click. The team keeps its tool and its habits: the automation plugs into it, it does not replace it.
With a shared mailbox (support@, contact@), you first have to build the ticket. An email has no structure: a twenty-line signature, the history quoted underneath, attachments, sometimes three different customers in the same thread. The workflow cleans up the message, isolates the new request, links the email to the right customer, then AI does its reading job. The draft can be dropped straight into the mailbox as a prepared reply, so that the team reviews it without changing tools.
With a messaging app (WhatsApp, Telegram, website chat), the customer expects a live reply, and reviewing before every message becomes difficult. Automatic sending is reserved there for simple questions, with a quick handover to a support agent: this is what a WhatsApp or Telegram bot connected to your CRM does, sorting and preparing without deciding in your place.
Which replies can go out without review?
The question is not “AI or human”, but “what does a mistake cost in this category?”. Three levels are enough.
- Automatic sending, without AI. The acknowledgement of receipt, the status of an order read from the online store, a parcel tracking link, a reminder of opening hours. The information comes from your systems, the reply is always the same: a rule delivers it better than a model, with nothing to pay on each message.
- AI draft, reviewed by the team. Most requests: a question about a product, an order change, a usage problem. The agent reviews in a few seconds what would have taken several minutes to look up and write.
- A person, from start to finish. Anything that commits the company towards the customer: a refund, a goodwill gesture, an exception to the contract, a dispute, a complaint that could end up with a lawyer or on social media. AI can still summarise the file, but the reply and the decision stay human. Our article on automated decisions explains why European law pushes in the same direction.
Watch out
An AI draft must never be able to trigger, on its own, an action that commits the company
Orchestrators provide for this: n8n, for example, lets you require human approval before an AI agent uses a specific tool. Its documentation lists sending external communications among sensitive actions. The approval request can arrive by email, on Slack, Teams, Telegram or WhatsApp; a refusal cancels the action, and the AI is informed.
Moving from one level to another is decided on figures, not on gut feeling. For the first few weeks, everything goes through review, and you measure, category by category, the share of drafts sent without changes. A category where the team hardly corrects anything any more can move to automatic sending, with a sample check. A category where it rewrites one draft in two stays under review, or does not deserve AI at all.
Why keep a human in the loop?
First, because customers ask for it. Here is what the 5,728 customers surveyed in December 2023 answered for a Gartner survey published in July 2024:
say they would prefer that companies did not use AI in customer service
would consider switching to a competitor if the company used it
Their first fear: that it becomes harder to reach a person. An AI that works behind the scenes, on replies that your team signs, does not raise that problem.
Second, because “customer service without agents” does not keep its promises. Gartner predicted in June 2025 that half of the organisations that intended to cut their support teams sharply thanks to AI would abandon those plans by 2027, according to The Register’s report. Of the 163 leaders surveyed, 95% planned to keep human agents alongside AI. Brian Weber, the firm’s VP analyst, explains there that generative AI projects “are not going as planned due to both the results and the unexpected costs”, and that a contact centre with no agents is “not yet technically feasible, nor is it operationally desirable”. The case of Klarna, which went back to hiring human agents after betting on its assistant, is told in our article on AI agents in production.
Third, because AI makes the team better instead of making it redundant. A study published in the Quarterly Journal of Economics, on real data from 5,172 customer service agents, measures an average 15% increase in issues resolved per hour thanks to an AI assistant. The least experienced improve the most; the most seasoned gain little. In other words, the tool passes part of the best agents’ know-how on to newcomers, and the best remain indispensable: their replies feed the documents, and their judgement settles the cases that AI cannot classify.
What your support team keeps that is valuable
Knowledge of the customers
A sense of what can be accepted or refused
The ability to defuse anger
What automation takes off its hands
Reading
Sorting
Searching
Typing
Your team thus answers faster, to more requests, without spending its evenings on it.
How much does AI sorting cost, and how do you keep it low?
The cost of AI-assisted support depends less on volume than on architecture choices. Three levers keep it low.
No AI where a rule is enough. The acknowledgement of receipt, the order status and the link to the customer are done without a model: every call avoided is a cost saved on every message, for good.
The right model for each step. Classifying a message does not call for the most powerful model in the catalogue. Writing a nuanced draft on a delicate case, perhaps. In our own routing table, sorting and intent detection go to the smallest model available, with a fallback to a model that runs on our own server; drafting goes to an intermediate model; sensitive data stays on a self-hosted model, which we run in production. Combining several models means paying for each task at its fair price instead of paying full price for everything.
A worked example, with our assumptions and Anthropic’s price list consulted on 9 October 2026. A team receives 3,000 messages a month.
| Step | Model | Assumption per message | Cost per month |
|---|---|---|---|
| Sorting the 3,000 messages | Claude Haiku 5.5 ($0.10 / $0.50 per million tokens) | 1,500 tokens read, 100 written | about $0.60 |
| The same sorting, on the biggest model | Claude Opus 5.5 ($4 / $20) | same | about $24 |
| Drafts for 2,000 messages | Claude Sonnet 5.5 ($2 / $10) | 5,000 tokens read (message, history, document excerpts), 500 written | about $30 |
The same sorting costs forty times more on the big model, with no guarantee of better classification. And even when well routed, the AI bill stays small next to the review time and the integration into your tools, which tokens do not show. These figures get corrected after a week of real measurements; the method for forecasting them is in our guide to the cost of AI in business, and our AI at the right cost offer sets up this routing for you.
What we run ourselves
We do not run a high-volume customer service operation, and we do not claim otherwise. But the rule we apply to our own automations is the one in this guide.
From the field
AI prepares, a person approves whatever touches money
Our expense-entry agent has run on Telegram in production since July 2026: it starts from the photo of the till receipt, extracts the information, then waits for human validation before writing to the accounting system.
Our n8n workflows run in production for the companies we work with, with alerts when an execution fails: a support automation that stops silently leaves customers without a reply, and monitoring matters as much as sorting. We explain this in how to monitor your n8n workflows.
What to set up before going live
- Customer data. A ticket contains names, addresses, sometimes sensitive information. Choose where they go: an AI provider under contract, with no data retention, or a model hosted on your side for what must not leave. The four possible levels are described in private AI in business: the four levels of control.
- Informing the customer. Since 2 August 2026, Article 50 of the EU AI Act requires that people who interact directly with an AI system be informed of it, unless this is clear from the context. A chatbot must therefore introduce itself as such. The details of the obligations, GDPR included, are in the AI Act and GDPR: what an SME that uses AI must do.
- The handover. The customer must be able to reach a person without searching, and the AI must know which requests to pass on without trying to answer them.
- A starting measurement. First response time, time spent per ticket and reopening rate, recorded before launch. Without them, no gain can be proved.
- A log. What the AI read, sorted and proposed, and what the agent changed. It is also the material used to improve the instructions.
Where to start
Getting started takes four steps:
- Record.
For two weeks, note what comes in to support: the volume, the categories, the time spent on each.
- Choose.
The three most frequent categories are almost always the right ones to start with.
- Plug in.
Sorting and drafts on these categories, all under review.
- Measure.
The rest is decided on the figures.
Our process automation with n8n offer sets up this circuit on your current tools, from the mailbox to the CRM, with human approval where it matters. If your replies need to rest on extensive documentation, an AI assistant that cites its sources from your documents prepares the drafts. And if support is only one project among others, the digital and AI maturity audit tells you in what order to tackle them and what each will bring.
Frequently asked questions
Can AI answer customers’ emails on its own?
For some of them, yes, but not straight away and not for everything. Factual replies drawn from your systems (order status, parcel tracking) go out on their own, often without AI. The others start as a reviewed draft. A category only moves to automatic sending when the team hardly corrects anything in it any more, and anything that commits the company, such as a refund or a dispute, stays approved by a person.
Do you need to change ticketing tools to add AI?
No, in most cases. Common ticketing tools and shared mailboxes connect to an orchestrator such as n8n, which reads the new messages, calls the model and drops the result into the existing tool: fields filled in, draft as an internal note. The team keeps its habits. Changing tools is only justified if the current one offers no way to read or write tickets from outside.
Do you have to tell the customer that AI prepared the reply?
For a chatbot that converses live, yes: since 2 August 2026, Article 50 of the EU AI Act requires the person to know that they are exchanging with an AI, unless it is obvious. For a draft prepared by AI, reviewed and sent by a support agent, it is your agent who replies and signs. Many companies still choose to state it in their service policy, for transparency.
Does customer data go to the AI provider?
It depends on the architecture chosen. Via an API, messages pass through the provider, and you need to check its retention policy and processing location, and list it as a processor in your processing register. For requests containing sensitive data, a model hosted on your own server keeps them from leaving. A routing setup can in fact combine the two depending on the content of the message.
What happens to the support team?
It spends less time reading, sorting and searching, and more time on the cases that call for judgement: an unhappy customer, an exception, a problem nobody had spotted. The available studies show a gain mostly among beginner agents, and Gartner observes that most customer service leaders keep their teams alongside AI. Their replies also become the knowledge base that feeds the drafts.
Sources
- Require human approval before an AI agent uses a specific tool (new tab)docs.n8n.io
- A Gartner survey published in July 2024 (new tab)gartner.com
- The Register’s report (new tab)theregister.com
- A study published in the Quarterly Journal of Economics (new tab)doi.org
- Anthropic’s price list (new tab)platform.claude.com
- Article 50 of the EU AI Act (new tab)ai-act-service-desk.ec.europa.eu
Jim · founder of Brain Plus AI
Also in automation and bots
Choose and price9 min
Automation without AI: when a rule is enough
Not every automation needs AI. A fixed rule costs less and always gives the same result: here is how to spot the steps where AI really pays off, and those where it only adds cost.
Understand9 min
Enterprise RAG: an AI that answers from your documents
RAG connects an AI model to a company's documents so that it answers by citing its sources. How it works, when it is useful, why it sometimes misses the point, and what to set up before opening it to the team.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.