Last month, I built an AI agent that handles my incoming email inquiries. It reads each email, classifies it (sales inquiry, support question, partnership request, spam), drafts an appropriate response, and either sends it automatically (for spam and common support questions) or sends it to me for review (for sales and partnerships).
The build took a weekend. Saturday for setup and testing. Sunday for refinement and edge cases. No coding required — I used n8n for the workflow and Claude for the AI processing.
An AI agent is different from a simple AI prompt. A prompt produces one output from one input. An agent makes decisions, takes actions, and handles multiple scenarios — more like a junior team member than a calculator. The agent evaluates the situation, chooses the appropriate response path, and executes. That decision-making capability is what makes agents powerful and what makes building them feel intimidating.
It should not feel intimidating. If you can describe a decision tree (“if this, then that; if that, then this”), you can build an AI agent. The tools are visual, the logic is sequential, and the test-and-refine cycle is fast.
What Your First Agent Should Do
Pick a task that meets three criteria:
1. High frequency. The task occurs daily or multiple times per day. Email triage, social media responses, data processing, content scheduling. The frequency matters because an agent that runs once a month does not justify the build time. An agent that runs fifty times a day recaptures significant hours.
2. Clear decision logic. The agent needs to make decisions, but the decisions follow patterns. “If sales inquiry, draft response and flag for review. If common support question, send standard response. If spam, archive.” The decisions are rule-based, not judgment-based. If the decision requires nuance that you cannot articulate as rules, the task is not ready for an agent.
3. Low stakes for errors. If the agent makes a wrong classification or drafts an imperfect response, the consequence is manageable. An email response that slightly misclassifies a sales inquiry as a partnership request delays the reply by a day — annoying but not catastrophic. An agent that miscalculates a financial report and sends it to investors — that is catastrophic. Do not build your first agent for high-stakes tasks.
The email triage agent is ideal for most founders because it meets all three criteria: emails arrive constantly, the classification follows patterns, and a misclassification is easily corrected.
Other good first agent candidates:
- Meeting note summarizer (reads transcripts, extracts action items, sends follow-up emails)
- Content scheduler (takes your pillar post, generates social media derivatives, schedules them)
- Lead qualifier (reads form submissions, scores leads based on criteria, routes high-quality leads to you)
- Expense categorizer (reads receipt images, categorizes expenses, updates your bookkeeping)
The Build Process (Saturday)
Morning: Set up the tools.
You need three things:
- n8n (self-hosted or cloud) as the workflow platform. n8n is open-source, visual, and designed for exactly this kind of automation. The cloud version starts at EUR 20/month. Self-hosting is free but requires a server.
- An AI provider account (Anthropic for Claude or OpenAI for GPT) for the intelligence layer. API access costs EUR 10-30/month for typical usage.
- Your email account connected to n8n via IMAP or Gmail integration. This is a standard n8n connection that takes five minutes to configure.
Setup steps: Install n8n (or sign up for the cloud version). Connect your email account. Connect your AI provider via API key. Test that n8n can read your incoming emails and that your AI provider responds to prompts. This setup takes one to two hours, including reading documentation for any unfamiliar steps.
Midday: Build the classification step.
Create a new workflow in n8n. The trigger: a new email arrives in your inbox.
The first processing node: send the email content to your AI provider with a classification prompt. The classification prompt is the brain of your agent. Structure it with XML tags — this reduces parsing ambiguity and helps the AI cleanly separate your instructions from the email content:
<context>
You are an email classifier for a SaaS company selling project management
tools to Austrian SMEs. You receive B2B inquiries in German and English.
Typical customers are teams of 5-50 people in manufacturing and services.
</context>
<task>
Classify the following email into one of these categories:
SALES, SUPPORT, PARTNERSHIP, SPAM, OTHER.
</task>
<email>
{{email_content}}
</email>
<output_format>
Respond with a JSON object: {"category": "CATEGORY", "confidence": "high/medium/low", "reasoning": "one sentence"}
</output_format>The specificity of the business description in the <context> tag matters. “A SaaS company selling project management tools to Austrian SMEs” produces better classification than “a software company.” The AI needs context to make good decisions, just like a human assistant would. The XML structure works because it gives the model clear boundaries between instruction, data, and expected output — no ambiguity about where the email ends and the instructions begin.
Test the classification with 20 real emails from your inbox. Pull emails from each category — five sales inquiries, five support questions, five partnership requests, five spam — and run them through the classifier. Note the accuracy. If accuracy is below 90%, add two or three examples of each category directly into the prompt. Examples activate pattern generalization — showing the model what a “tricky” sales inquiry looks like beats describing it abstractly.
Afternoon: Build the response step.
After classification, the workflow branches. Each branch handles a different category.
SALES branch: AI drafts a personalized response. Use a proper system prompt for this agent:
system="You are a sales response assistant for [your business].
Your task is to draft warm, professional replies to inbound sales inquiries.
You have access to the company's service descriptions and pricing tiers.
Keep responses under 150 words. Always acknowledge the prospect's specific need.
Always include a calendar link for a 30-minute call."Then pass the email content as the user message. The system prompt stays constant across all sales emails; only the email content changes. This consistency is why agents outperform ad hoc prompting — the same quality every time, regardless of your energy level. The draft goes to your inbox as a pending reply for you to review and send.
SUPPORT branch: AI generates a response using your FAQ knowledge base — a document containing your standard answers to common questions. Create this document: list your top 20 FAQs with complete answers. Upload it to n8n or include it directly in the prompt context. Structure the handoff to the support agent with XML tags so the model can distinguish the FAQ data from the customer’s question:
<knowledge_base>{{your_faq_document}}</knowledge_base>
<customer_email>{{email_content}}</customer_email>
<task>Draft a reply using only information from the knowledge base.
If the question is not covered, acknowledge it and say a team member
will follow up within one business day.</task>For common support questions (covered by the FAQ), the response can be sent automatically. For unusual questions (not in the FAQ), the response goes to you for review.
PARTNERSHIP branch: AI drafts an acknowledgment: “Thank you for reaching out about a potential partnership. I would be interested in discussing this. Let me review the details and get back to you within two business days.” The draft goes to you for review before sending.
SPAM branch: Archive automatically. No response.
OTHER branch: Forward to you with the original email and the AI’s classification reasoning, so you can handle it manually.
Test each branch with the real emails from your morning test set. Verify that the drafted responses are appropriate, the tone matches your brand, and the routing logic works correctly.
The Refinement Process (Sunday)
Morning: Handle edge cases.
Run through 50 more emails — a broader sample that includes unusual cases. Identify cases where the agent misclassified or drafted poorly. Common edge cases:
- A customer email that is both a support question AND a sales inquiry (upsell opportunity). Solution: add a “SUPPORT-UPSELL” category that routes to you with a note about the upsell potential.
- An email in German when your FAQ is in English. Solution: add a translation step before the classification, or create a German FAQ document.
- An email from a known VIP customer that should always be handled personally. Solution: add a “VIP check” step that compares the sender against a list of important contacts and routes directly to you regardless of classification.
Adjust prompts. Add examples of tricky cases to the prompt context. Each refinement improves the agent’s accuracy. Expect to make 5-10 prompt adjustments during Sunday morning.
Afternoon: Add the human oversight layer.
For any action the agent takes autonomously (support responses, spam archiving), set up a daily digest: a summary email showing what the agent did, so you can spot-check for errors. The digest should include: number of emails processed, classification breakdown, and the full text of any auto-sent responses.
For actions requiring your approval (sales and partnership responses), set up a notification that lets you review and send with minimal friction. A Slack notification with a “Review” button, or an email with the draft response pre-loaded and a “Send” button.
The oversight layer is not about distrust. It is about responsible automation. Even a well-built agent makes mistakes, especially in the first weeks. The oversight catches errors before they reach customers and provides data for further refinement.
Evening: Go live.
Enable the workflow trigger. The agent starts processing emails in real time. The first email that arrives and gets correctly classified, drafted, and routed feels like watching a colleague start their first day. The agent is working for you.
After the Weekend
Week 1: Monitor daily. Read the daily digest. Correct classification errors by updating the prompt with counterexamples. If the agent misclassifies a support email as sales, add that email as an example in the prompt with the correct classification. One critical lesson: when adding corrections, tell the agent what to do, not what to avoid. “Emails mentioning existing subscriptions are SUPPORT, even if they ask about upgrades” works better than “DO NOT classify upgrade emails as SALES.” Negative instructions cause overtriggering — the agent becomes so focused on avoiding the mistake that it over-corrects in other areas.
Week 2-3: Monitor every other day. The agent improves as your prompts become more refined. Track the accuracy rate. By week two, you should be above 90% classification accuracy. Consider the reversibility of each action your agent takes. Drafting a response for your review is fully reversible — low blast radius. Auto-sending a response is not. Keep auto-send limited to truly routine categories until you have confidence built from data, not from hope.
Month 2: Monitor weekly. The agent handles 80-90% of incoming emails without your involvement. The digest confirms accuracy. You handle the 10-20% that require personal attention.
Ongoing: When new types of emails appear (a new product launch generates a new category of inquiries), add classification categories and response templates. The agent evolves with your business.
The Value
The email triage agent saves me approximately 45 minutes per day. That is 3.75 hours per week, 15 hours per month. At my hourly value, that is EUR 1,500+ per month of recaptured time.
The cost: EUR 20/month for n8n hosting + EUR 10-30/month for AI API usage = EUR 30-50/month. Net value: over EUR 1,400/month.
The non-financial value is equally significant. I no longer dread opening my inbox. The anxiety of a full inbox — will there be something urgent? something I forgot? something that needs immediate action? — is gone. The agent handled it. What remains when I open my inbox are the emails that actually need me: strategic decisions, important relationships, complex problems. The noise is filtered. The signal remains.
Your Second Agent
Once the email agent is running, the pattern is clear. Pick the next task. Define the classification logic. Build the workflow. Test. Refine. Deploy. Monitor.
Good second agent candidates:
- Customer service triage (similar to email but from your help desk or contact form)
- Content production assistant (generates weekly content derivatives from your pillar posts)
- Competitive intelligence monitor (weekly competitor analysis delivered to your inbox)
- Meeting preparation agent (before each meeting, compiles relevant information about the attendee from your CRM and public sources)
Each agent follows the same build pattern. Each weekend produces a permanent productivity gain. The compound effect of three or four agents running simultaneously transforms your operational capacity.
One weekend of building. Permanent time savings. That is the promise of AI agents for founders, and it is achievable with tools available today. No coding required. Just a clear task, a clear decision framework, and a willingness to ship something imperfect and improve it over time.
Build your first agent this weekend.
