What customers really expect#
For most requests, customers don't expect a brilliant answer: they expect a fast, accurate one. Where's my order? Are you open on Saturday? How do I change my appointment? Can I pay in installments? These questions often make up a large share of the messages a small business receives, and the answers already exist somewhere: on your website, in your software, in a colleague's head.
Slow responses rarely come from difficulty. They come from a queue: messages arrive while you're busy, in the evening, on weekends, and pile up.
An AI customer service agent works on that queue. It doesn't replace your relationship with your customers; it handles what's simple so you have time to handle well what isn't.
Three levels of autonomy#
The question isn't “do I need an agent?” but “how far should I let it go?” There are three levels, and you can move from one to the next gradually.
| Level | The agent does | You do | Best for |
|---|---|---|---|
| 1. Prepare | Sorts the request and suggests a reply as a draft or internal note | Review, adjust, send | Getting started, sensitive topics |
| 2. Answer simple questions | Answers common questions on its own, passes on the rest | Handle the requests passed to you | Opening hours, order tracking, practical information |
| 3. Resolve | Answers and performs simple actions (change a delivery address, resend an invoice) | Approve the rules, supervise | Tightly defined processes |
Most businesses do well to start at level 1 for a few weeks, then move to level 2 for the question categories where drafts consistently go out without edits.
Four setups depending on your channel#
You use a ticketing tool#
The Ticket replies template is triggered by every new ticket in Zendesk: it categorizes it (delivery, billing, technical), looks for the answer in your knowledge base and past tickets, and adds a reply as an internal note, ready to send. Your team stays in control of sending. The same approach works with Freshdesk, Gorgias, Intercom or Help Scout.
You run an online shop#
The Online shop customer service template answers common questions (order tracking, sizes, lead times, address changes) every hour from 9 a.m. to 7 p.m., checking the real status of the order and parcel. It passes on anything that requires a decision: a refund outside your policy, a defective product, an unhappy customer.
Your customers message you on WhatsApp#
The WhatsApp replies template answers common questions on WhatsApp Business (hours, prices, availability, address) from the information you give it, and forwards other requests to you with a summary. It's often the preferred channel for customers of local shops, tradespeople and practitioners.
Your complaints drag on#
The Complaint tracking template checks response times in Freshdesk every business day, nudges the team on overdue cases and sends a progress update to any customer who has been waiting more than 48 hours. It doesn't resolve the complaint for you: it makes sure it isn't forgotten.
The knowledge base: the foundation for everything else#
A customer service agent is only as good as the information it's given. Without it, it can only answer vaguely, or worse, make things up. Before launching the agent, gather:
- Your practical information: opening hours (including exceptions), address, directions, contact methods.
- Your policies: delivery, returns, refunds, warranties, payment methods.
- Your prices, when they're public.
- Your standard answers to the ten or fifteen most frequent questions.
- Your hand-off rules: what the agent must never handle on its own.
These documents are added as the agent's context files (PDF, Word, Excel, text), or written directly in the conversation. They're indexed, and the agent consults them for every request. When you change a policy, update the document: the agent takes it into account from the next run.
To keep this knowledge base alive, the Always-current FAQ template analyzes the questions received each month, spots the recurring ones, drafts answers and updates your FAQ in Notion, flagging the ones that need approval.
Rules that protect your customer relationships#
Here are the rules we recommend building into every customer service agent:
- Never make things up. If the information isn't in the knowledge base or your tools, the agent says so and passes the request on.
- Never promise a lead time, goodwill gesture or refund that isn't in your policies.
- Escalate immediately messages from unhappy customers, threats of legal action, health or safety questions, and any request it doesn't understand.
- Introduce itself honestly: if the agent replies on its own, it doesn't pretend to be a named team member. A sign-off like “The shop team” is enough.
- Respect data: ask only for the information needed (an order number, not a card number).
A sample brief to copy#
For my online tea and accessories shop (Shopify), reply to customer emails received at contact@. Questions you can handle on your own: order tracking (check Shopify and the carrier), delivery times, brewing advice based on my product pages, address changes if the order hasn't shipped yet. For everything else (refunds, damaged products, complaints, trade inquiries), don't reply: forward the request to me on Slack with a two-line summary and a suggested reply. Warm tone, signed “The Tea House team.” My terms of sale and FAQ are attached.
Scenario: a physiotherapy practice#
A practice with three physiotherapists receives messages by email and WhatsApp: appointment requests, questions about what to bring, cancellations, requests for invoices for insurance claims.
The practice sets up an agent with strict rules:
- It answers practical questions on its own (hours, address, accessibility, what to bring to the first session), based on a document approved by the practitioners.
- It points to the online booking link for any appointment request.
- It forwards cancellations and invoice requests to the front desk, with a summary.
- It never answers a medical question, however simple: it passes it on and tells the patient a practitioner will get back to them.
The front desk now only handles the requests that need a person. Patients get practical answers the same evening, including outside opening hours.
Setting it up in five steps#
- Take stock of your requests. Go through the last few weeks of customer messages and roughly sort them: order tracking, practical information, quotes, complaints, other. You'll see where the agent will be most useful.
- Write or gather the answers to the most frequent questions. It's often the longest step, and the most rewarding: this document will serve your team too.
- Describe the agent in the chat: channel, categories it handles alone, categories it passes on, tone, sign-off. The architect asks about anything missing and shows you sample replies to dummy requests.
- Start at the “prepare” level: the agent drafts, you send. Note the corrections you make most often and tell the agent, which remembers them.
- Expand category by category. When order-tracking replies no longer need edits, allow direct sending for that category only, and keep reviewing the rest.
Every run appears in the agent's log, in your dashboard: at any time you can check what it replied, to whom and on what basis.
Measure what matters#
After a few weeks, look at:
- The share of requests handled without intervention: it tells you whether the agent is useful.
- The share of drafts sent without edits: it tells you whether you can move to the next level of autonomy.
- Requests passed on: their reasons show which information to add to the knowledge base.
- Customer feedback: the Satisfaction survey template sends a short survey after each service and analyzes the answers every month.
Pricing and first steps#
An agent that drafts replies from a knowledge base is usually in the “Essential” or “Advanced” tier, depending on the number of tools and the volume; the exact price is calculated and quoted before launch. See the pricing page. Templates for customer support are grouped in the library, and our article on automated email management complements this one for requests that land in your inbox.