You run a service business. You get 100-200 emails a day. 80% of them are the same questions. How much does it cost, when's the next available slot, are you open on Saturday. And someone has to answer them. Every day. From scratch.
Sound familiar? Because that's the reality for most service businesses. I've implemented email automation for over a dozen clients in the past year. Let me tell you what works, what doesn't, and how much it costs.
Case study: cleaning company, Krakow, 200 emails per day
The client came to us at kaminski.link with a simple problem. Three people were handling emails. Eight hours a day. Answering questions about pricing, availability, and service scope. Same answers, copy-paste, day after day.
We deployed an AI assistant based on GPT-4 with the company's knowledge base. Simple scenario:
- Email arrives in the inbox
- AI analyzes content and classifies it into one of 12 categories
- If the category is known - generates a response based on the knowledge base
- If uncertain - escalates to a human with a suggested response
- Human approves or edits before sending
Result after 3 months? One person instead of three. 40 emails per day requiring human intervention instead of 200. Response time dropped from 4 hours to 11 minutes.
How to set it up technically
You don't need a team of developers. Seriously. The whole thing can be built on n8n or Make in a few days. If you're not sure which tool to pick, I compared them in my Zapier vs Make article.
The tech stack I use for clients:
- n8n as the workflow engine (self-hosted, zero license costs)
- OpenAI API for classification and response generation
- Knowledge base in Notion or Google Docs - easy for the client to update. A CRM that collects conversation history helps here too
- IMAP/SMTP for receiving and sending emails
The whole flow works like this: n8n checks the inbox every minute. New email? It goes to the OpenAI API with a system prompt and knowledge base. AI returns a category, suggested response, and confidence level (0-100). Confidence above 85? Email goes to the auto-send queue. Below? It goes to a human with a draft response.
One important thing here. I never send emails fully automatically from day one. For the first 2 weeks, every response goes through a human. Only when accuracy hits 95%+ do I enable auto-send for the highest-confidence categories.
Costs - no fluff
Let me break it down, because I know that's what you're looking for:
- Implementation (n8n setup, prompt engineering, knowledge base): $1,000-1,700 one-time
- OpenAI API cost: $50-120/month at 200 emails per day
- n8n hosting: $12-25/month (VPS)
- Knowledge base maintenance and updates: $120/month (optional, by our team)
Total monthly cost: $180-265. Salary savings: $1,700-2,200/month. ROI? First month. Sometimes even the first week.
But heads up. These numbers are for a company with 200 emails per day. If you get 20 emails a day, automation won't pay for itself. The break-even point is roughly 50 emails per day.
Where AI falls flat
Because it does. And we need to talk about it.
Complaints. AI can't read an angry customer's emotions. When considering automating customer support, it's also worth checking whether a website chatbot could handle some of these issues. It writes a technically correct response, but the tone is too smooth, too corporate. An upset customer gets a textbook customer service reply and gets even more upset.
Non-standard requests. Customer asks about a service that's not in the knowledge base. Instead of saying "I don't know," AI makes things up. Hallucinations are a real risk and you need to take them seriously. That's why the 85% confidence threshold and human-in-the-loop aren't optional. They're essential.
Multi-thread context. Customer writes their third email in a thread, referencing agreements from the first one. AI without conversation history responds generically. Solution? Feed the entire thread as context. More expensive (more tokens), but necessary.
How to get started - step by step
Don't start with tools. Do an email audit first:
- For one week, tag every email with a category (pricing, scheduling, complaint, other)
- Count how many emails fall in each category
- Check which categories have repetitive responses
- Start automation with the most repetitive category
Then write your knowledge base. Answers to the 20 most common questions. Simple, specific, in your company's voice. It'll take you half a day. And it's the most important work in the entire project.
Only then choose your tools. n8n if you have someone technical. Make if you don't. Zapier if budget isn't a concern. And if you don't want to deal with it yourself - reach out to us at kaminski.link. We do this every day.
Email automation isn't about replacing people with machines. It's about freeing people from work that machines do better and faster. Let your employees do what requires empathy, creativity, and common sense. The rest? Let AI handle it.