AI for customer service pays off for small firms mostly behind the scenes: it sorts incoming requests, drafts replies, keeps the knowledge behind them at hand, and summarizes long threads. The reply itself, every promise, every goodwill gesture and every conflict stays with a person. That is the opposite of the picture many people have of "AI customer service": a bot that stalls customers until they give up. A chatbot on your website can still make sense, but only with a narrow job, a visible way to reach a person, and a notice that an AI is answering. Article 50 of the EU AI Act has required that notice from the provider since 2 August 2026.
In my business the pattern looks like this: Frieda separates my inbox into Now, Today, Can wait and Trash, prepares reply drafts and returns anything that needs my personal judgment. She has never sent anything on her own: every reply stays a draft until I approve it. Sonja writes business emails for me, usually handovers to a new client after a website redesign: she checks name, address, demo link and price against the source, brings the text to my language level and drops it into my inbox as a draft. In short: Frieda brings order to my inbox, Sonja writes what actually goes out. Inbox sorting in detail is in Sorting Your Inbox: The Pile System, the tool question in AI Email Triage: Agent or Tool?.
AI for customer service: prepare, don't fend off
In a small firm, customer service is rarely a department, and the bottleneck is rarely writing a single reply. It's everything before it: understanding what the customer wants, checking what was discussed last time, finding the right information. AI can take on that groundwork.
Fending off keeps the customer from reaching a person. Preparing helps the person answer faster and better. The test for any tool: who writes to the customer in the end, and who read it first? If the answer is "the machine, and nobody", it's fending off, however friendly the bot sounds.
The four channels and what AI prepares there
With email and reviews the AI stays invisible: the customer reads a reply a person approved. In chat and on the phone it talks to the customer directly, without approval before every sentence, so those two channels need tighter rules.
What AI prepares: four jobs
1. Sorting. Every request gets a label: what the customer wants (question, appointment, quote, complaint, invoice), how urgent it is, who is responsible. The AI rarely needs the full text for that. Frieda reads sender, subject and preview first, and the full text only for emails that actually need a reply.
2. Reply draft. The AI writes a reply that a person reads, edits and sends. A draft gets good with three ingredients: real examples of your voice, a source for facts such as prices and opening hours, and a rule for what's missing. Sonja works exactly like that for me: if a name, address, demo link or price is missing, she asks instead of guessing. How an AI picks up your voice is in Can AI Write Emails in My Voice?.
3. Knowledge base. Here a knowledge base isn't a software project but a maintained list: the frequent questions, your checked answer, the source and the date of the last check. How to build it from real conversations is in Build FAQs From Customer Conversations. The same list later feeds reply drafts, a chatbot and the FAQ on your website.
4. Summary. Long email threads, phone notes, the last quote: the AI sums up what the customer wants, what has been promised and what's open. That helps most when someone takes over an unfamiliar case.
What stays with a person
- Every promise. Prices, dates, delivery times, discounts. An AI may take them from a source, never invent new ones.
- Money and goodwill. Refunds, credit notes, rework at your own cost.
- Complaints and conflicts. A person decides what happens; the draft only puts that decision into words. The process for that is in Handling Complaints: A Process With AI Drafts.
- Decisions with legal effect. Article 22 GDPR gives people, in principle, "the right not to be subject to a decision based solely on automated processing" that has legal or similarly significant effects; paragraph 2 sets out exceptions. House rule: a person decides whether a customer gets a contract, a refund or a refusal.
- Sending. As long as a person approves, a mistake in a draft lands on your desk and not with the customer.
Chatbot on your website: pros and cons
Besides an AI receptionist, a chatbot is the only place where an AI talks to your customers without approval. What sets it apart from an agent is in AI Agent, Chatbot, or Assistant?.
Microsoft admits that generative AI responses "aren't guaranteed to be 100% factual". My rule of thumb, not a proven threshold: a chatbot pays off only once the knowledge base is in place and you have written down which questions it may answer.
An example: say a painting business gets the same questions every week about its service area, lead time and whether it also paints facades. A chatbot that answers exactly those questions from the maintained list and, for a quote, takes down address, photos and a callback number is narrow enough. A chatbot that estimates prices or promises dates is not.
Five rules I would give any chatbot:
- It answers only from the knowledge base; anything not in there it takes down as a request.
- It never gives prices, promises, diagnoses or legal information.
- Every conversation shows a visible way to a person.
- Someone reads the conversations: questions without an answer go into the knowledge base.
- The first sentence says that an AI is answering.
Disclosure duty: the notice in the first sentence
Under Article 50(1) of the EU AI Act, providers ensure that AI systems intended to interact directly with people are designed "in such a way that the natural persons concerned are informed that they are interacting with an AI system", unless that is obvious to a reasonably well-informed person. Under paragraph 5 this must happen "at the latest at the time of the first interaction", "in a clear and distinguishable manner". The duty has applied since 2 August 2026. Regulation (EU) 2026/1744, the Digital Omnibus on AI, rewrote only paragraph 7 of Article 50. For machine-readable marking under paragraph 2 it added a deadline of 2 December 2026 in Article 111(4), only for systems already on the market before 2 August 2026. Paragraph 1 and the start date are unchanged.
The duty is on the provider. Under Article 3(3) a provider is whoever develops an AI system or has it developed and "puts the AI system into service under its own name or trademark". If you only use a ready-made chatbot service, you are usually a deployer. If you have a chatbot built and put it on your website under your own name, you may be the provider yourself, so have that checked by an expert. The notice itself is easy: "You're chatting with the digital assistant of Example Painters, an AI." What else Article 50 covers is in AI Disclosure Rules Under the EU AI Act. This isn't legal advice, just the operator's perspective.
Privacy: three points
- Information. Under Article 13 GDPR you inform people "at the time when personal data are obtained". Chatbot, AI receptionist and the service providers behind them belong in your privacy notice.
- Processing agreement. Under Article 28 GDPR you use "only processors providing sufficient guarantees", with a contract.
- As little as needed. For Frieda, sensitive folders are off-limits: personal matters, job applications, health and legal disputes, the instructions say she doesn't touch them, not even to sort them. How I decide what an AI gets to see is in AI and Privacy: What the AI Gets to See.
This isn't legal advice either.
Frequently asked questions
What can AI take on in customer service?
The groundwork: sorting requests, drafting replies, using a knowledge base and summarizing long threads. Promises, goodwill, complaints and sending stay with a person.
Does a chatbot have to say it's an AI?
Yes, unless it's obvious. Since 2 August 2026, Article 50(1) of the EU AI Act requires it, at the latest at the first interaction. The duty is on the provider; if you have a system developed and put it into service under your own name, that can be you. This isn't legal advice.
Can AI answer customer emails automatically?
Technically yes, sensibly rarely: you are responsible for the content of a sent email, whoever wrote it. In my business Frieda sorts my inbox, and every reply from her stays a draft until I approve it. A support inbox works the same way.
What are examples of AI in customer service for small firms?
A pre-sorted inbox with reply drafts, summaries of phone notes before the callback, reply drafts for Google reviews, and an FAQ built from real customer questions. A chatbot or AI receptionist comes later, once the knowledge base is in place.
How to continue
Don't start with a tool, start with a week of notes: which channel, what about, who replied? Your checked answers to the most frequent questions become your knowledge base. Then begin where most requests arrive, with sorting and drafts; a chatbot or AI receptionist comes once that list holds up.
Frieda and Sonja are available as ready-made packages in my community AI Employee Framework, with course, rules and approval limits. There you adapt categories, tone and escalation paths to your business, and in the weekly live Q&A call we go through rules like these together.
