How to Build an AI Automation Stack for Under $100/Month (The Exact Tools I Use)

Most people blow $300-500/month on AI tools they barely use. I run my entire automation infrastructure — content, lead gen, data ops, client delivery — for under $100/month. Here's the exact stack and how I think about building it.

Why Most AI Tool Stacks Cost Too Much

The typical mistake is subscribing to full-featured SaaS platforms when you only need 20% of what they offer. You're paying for dashboards, team collaboration features, and enterprise integrations that a solo operator or small team will never touch.

The workflow automation market hit $23.77 billion in 2025 and is growing at nearly 10% annually. That growth means there are more tools than ever competing for your money. The vendors want you on annual plans. They want you paying per seat. They want you locked in.

Here's the counterplay: buy capabilities, not platforms. Use free tiers aggressively. Pay for API calls instead of subscriptions wherever possible. And pick tools that compound — where learning one makes the others more powerful.

The Three-Layer Stack Framework

Every AI automation stack worth building has three layers. Miss one and the whole thing falls apart. Over-invest in one and you're wasting money.

Layer 1 — Orchestration: This is the workflow engine that connects everything. It's the central nervous system. n8n or Make lives here.

Layer 2 — Intelligence: This is your LLM layer. The AI brain that processes, generates, and decides. Claude API, OpenAI API, or both.

Layer 3 — Infrastructure: This is where your data lives and persists. Your database, your file storage, your CRM. Airtable, Supabase, or Google Sheets.

Think of it like a kitchen: orchestration is the recipe (the sequence of steps), intelligence is the chef (the decision-maker), and infrastructure is the pantry (where ingredients are stored and retrieved).

Layer 1: Orchestration — The Workflow Engine

This is where I spend the most time and the least money. Your orchestration layer is the backbone of every automation you build. Pick wrong here and you'll feel it everywhere.

My Pick: n8n (Self-Hosted)

Monthly cost: $5-10 (VPS hosting on Railway, Render, or a basic DigitalOcean droplet)

n8n is open source, self-hostable, and gives you unlimited workflows and executions for the cost of a small server. The Cloud version starts at €20/month, but self-hosting is where the economics get ridiculous.

With n8n you get 400+ native integrations, built-in AI nodes that connect directly to LLM APIs, a visual workflow builder that's genuinely powerful (not a toy), and the ability to run unlimited workflows without per-execution pricing.

The tradeoff: you need to be comfortable with basic server management. I'm talking 2 hours per month for updates and monitoring — not a full DevOps operation. If that still feels like too much, Railway and Render handle most of it for you.

The Alternative: Make.com

Monthly cost: $10.59/month (Core plan)

If you don't want to self-host anything, Make is the move. The free tier gives you 1,000 operations per month to test with, and the Core plan adds unlimited active scenarios with minute-level scheduling.

Make's visual builder is arguably more intuitive than n8n's for beginners. The drag-and-drop interface makes complex branching logic feel approachable. Where Make falls short is at scale — once you're running thousands of operations per month, the per-operation pricing adds up faster than a self-hosted n8n instance.

Why Not Zapier?

Zapier is the tool most people start with, and the tool most serious automators leave. The free tier is limited to 100 tasks per month with single-step Zaps. The Starter plan jumps to $19.99/month for 750 tasks. By the time you need multi-step workflows with conditional logic, you're looking at $49+/month for the Professional plan.

For the same money, you get dramatically more capability with n8n or Make. Zapier's advantage is simplicity — if you just need "when X happens, do Y" and nothing more, it works. But if you're reading this article, you want more than that.

Layer 2: Intelligence — The LLM Layer

This is where most people either overspend or underinvest. You don't need a $20/month ChatGPT Plus subscription to power automations. You need API access, where you pay per token — only for what you actually use.

My Pick: Claude API (Haiku + Sonnet)

Monthly cost: $5-30 (depending on volume)

Here's the key insight most people miss: you don't use one model for everything. You use the cheapest model that gets the job done for each specific task.

Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens. For classification, routing, summarization, and simple extraction tasks, Haiku is overkill-good at a fraction of the cost. Claude Sonnet 4.5 costs $3/$15 per million tokens and handles complex reasoning, content generation, and multi-step analysis.

For a typical small business running 10-15 automated workflows, you're processing maybe 2-5 million tokens per month total. At Haiku-heavy usage, that's $5-15/month. Mix in Sonnet for your heavy-lift workflows and you're at $15-30/month.

The Alternative: OpenAI API (GPT-4o Mini + GPT-4o)

Monthly cost: $5-25 (depending on volume)

GPT-4o Mini at $0.15 per million input tokens is absurdly cheap for lightweight tasks. GPT-4o at $2.50/$10 per million tokens handles the complex stuff. The economics are similar to Claude's tiered approach.

The real power move is using both providers. Route different workflow steps to different models based on the task. Use GPT-4o Mini for data extraction, Claude Sonnet for content generation, and Haiku for classification. n8n and Make both support multi-provider LLM routing natively.

Layer 3: Infrastructure — The Data Layer

Your automations need somewhere to store data, trigger from, and write results back to. This is the least glamorous layer and the one that breaks everything if you cheap out on it.

My Pick: Airtable (Free Tier) + Supabase (Free Tier)

Monthly cost: $0-20

Airtable's free tier gives you unlimited bases with 1,000 records per base, 100 automation runs per month, and a spreadsheet-like interface that non-technical team members can actually use. For most solopreneurs and small businesses, this is plenty for your CRM, content calendar, lead tracking, and project management.

When you outgrow the free tier, the Team plan is $20/user/month — but push that decision as long as possible. 1,000 records per base handles a lot more than you'd think when your automations are cleaning and archiving data regularly.

For anything that needs a real database — storing API responses, logging workflow executions, building custom apps on top of your data — Supabase's free tier is exceptional. You get 500 MB of Postgres database storage, unlimited API requests, and 50,000 monthly active users for auth. The only gotcha is that free-tier projects pause after 7 days of inactivity, so keep a heartbeat workflow pinging it.

The Alternative: Google Sheets + Google Drive

Monthly cost: $0

I know. Google Sheets as a database sounds like a hack. And it is. But for early-stage automations with low volume, it works surprisingly well. Both n8n and Make have native Google Sheets integrations. You can read, write, and update rows as part of any workflow.

The ceiling is low — once you're past a few hundred rows or need concurrent writes, Sheets buckles. But if you're testing a new automation before committing to infrastructure, Sheets lets you validate the logic with zero cost and zero setup.

The Full Stack: What It Actually Costs

Here's the exact breakdown, from minimum viable to comfortable production:

Even at the "comfortable" level, you're under $100/month with significant headroom. The budget option at $13/month is genuinely production-capable — I've seen businesses run five-figure monthly revenue on stacks that cost less than a Netflix subscription.

Five Workflows You Can Build on Day One

Knowing the tools is useless without knowing what to build. Here are five workflows that pay for the entire stack within the first month.

1. Inbound Lead Enrichment and Routing

Trigger: New form submission or email inquiry lands in Airtable. Process: n8n grabs the lead, hits a free enrichment API (Hunter.io for email verification, Clearbit free tier for company data), then uses Claude Haiku to classify the lead as hot, warm, or cold based on the enriched data. Output: Lead gets tagged in Airtable, and hot leads trigger an instant Slack notification or email alert.

Time saved: 15-30 minutes per lead, or 5-10 hours per month for a business getting 20+ inbound leads.

2. Content Repurposing Pipeline

Trigger: New YouTube video transcript dropped into a Google Doc or Airtable record. Process: Claude Sonnet summarizes the transcript, extracts key quotes, generates a blog post outline, writes 3 social media posts (LinkedIn, Twitter, Instagram caption), and creates an email newsletter draft. Output: Five pieces of content from one video, all stored in Airtable and ready for review.

Time saved: 3-4 hours per video. If you publish weekly, that's 12-16 hours per month.

3. Client Report Generation

Trigger: Scheduled weekly or monthly (cron trigger in n8n). Process: Pull metrics from your data sources (Google Analytics API, Stripe API, CRM records), feed them to Claude Sonnet with a templated prompt, generate a formatted summary. Output: A polished client report emailed automatically or dropped into a shared folder.

Time saved: 1-2 hours per client per report cycle.

4. Email Triage and Auto-Response

Trigger: New email arrives in your inbox (Gmail or Outlook integration). Process: Claude Haiku classifies the email (support request, sales inquiry, newsletter, spam). Based on classification, the workflow either drafts a response for your review, routes it to the right Airtable board, or archives it automatically. Output: Your inbox is pre-sorted every morning with draft responses waiting for a one-click send.

Time saved: 30-60 minutes per day, or 10-20 hours per month.

5. Competitor Price and Feature Monitoring

Trigger: Scheduled daily or weekly. Process: n8n scrapes competitor pricing pages (or uses their APIs where available), compares against your stored baseline in Airtable, and uses Claude to summarize what changed. Output: A weekly digest of competitor moves, delivered to Slack or email.

Time saved: 2-3 hours per week of manual research.

The Mistakes That Blow Your Budget

I've helped dozens of people build their automation stacks. These are the patterns that consistently push costs past $100/month for no good reason.

Mistake 1: Paying for ChatGPT Plus and Claude Pro simultaneously. If you're using AI for automations, you need API access, not chat subscriptions. The APIs are almost always cheaper for automation workloads. Keep one chat subscription for interactive use if you want it, but don't count it as part of your automation stack cost.

Mistake 2: Starting with Zapier and never leaving. Zapier's per-task pricing makes it the most expensive option at scale. People stick with it because switching feels hard. It isn't — export your Zap logic, rebuild in Make or n8n in an afternoon, and your monthly bill drops immediately.

Mistake 3: Overprovisioning your database. You don't need Airtable's $20/user/month Team plan on day one. You don't need Supabase Pro. Start on free tiers, hit the ceiling, then upgrade only the specific constraint that's blocking you.

Mistake 4: Using GPT-4 or Claude Opus for everything. Route your tasks intelligently. Classification, extraction, and simple generation should hit the cheapest model available. Reserve the expensive models for complex reasoning and high-stakes content generation.

How to Get Started This Week

Don't try to build all five workflows at once. Here's the sequence I recommend:

Day 1-2: Set up n8n (self-hosted on Railway or Render) or sign up for Make's free tier. Build one simple workflow — something like "new Airtable row triggers a Slack message." Get comfortable with the interface.

Day 3-4: Add the LLM layer. Get API keys for Claude and/or OpenAI. Build a workflow that takes text input, sends it to an LLM, and writes the output somewhere. The content repurposing pipeline is a great first real project.

Day 5-7: Connect your data layer. Set up Airtable as your central hub. Build one workflow that reads from and writes back to Airtable with an LLM step in the middle. The lead enrichment workflow is ideal for this.

Week 2 onward: Add workflows one at a time. Each new automation should solve a specific pain point you feel weekly. Don't automate theoretical problems — automate the thing that ate 3 hours of your Tuesday last week.