The AI automation stack I actually use every day.

Every few weeks someone posts a "50 AI tools you need in 2026" thread. I've never used 45 of the tools on any of those lists. The truth is that a working automation stack is small, boring, and mostly made of things that aren't marketed as "AI tools" at all.

This is the actual stack running across my stores and workflows right now. No affiliate links, no tools I installed once and forgot. If it's on this list, it's doing real work in production this week.

The orchestration layer: n8n

n8n is the spine of everything. It's an open-source workflow automation tool — think Zapier or Make, but self-hostable and far more flexible. I run it on a small VPS, which means no per-task pricing and full control over the data.

Almost every automation I have touches n8n at some point: order routing, inventory sync, CS classification, report generation, scheduled data pulls. The reason I use n8n over Make (which I also use, for lighter workflows) is the code nodes — when a workflow needs custom logic, I can drop into JavaScript without leaving the tool.

If you're starting from zero, this is the single highest-leverage tool to learn. It's the difference between automating one thing and being able to automate anything.

The intelligence layer: Claude API

Where I need actual language understanding — classification, summarisation, drafting, extraction — I use the Claude API. Not the chat interface. The API, called directly from n8n workflows.

The most common use: classifying incoming customer service messages by intent and urgency. A short prompt, a structured output, and the workflow can route the message correctly without a human reading it first. It costs fractions of a cent per message and it's more reliable than the keyword-matching approach I used before.

The trick with LLMs in production isn't using them for everything. It's using them for the one step in a workflow that actually requires understanding language — and using deterministic code for everything else.

I also use it for drafting: first-pass product descriptions, first-pass email replies, first-pass blog outlines. Always a first pass, never the final version. The model gets me to 70% and I do the last 30% myself.

The data layer: Supabase

Supabase is a hosted Postgres database with a good API layer on top. It's where all my operational data lives: order records, stock levels, CS ticket queues, error logs, configuration tables.

The reason this matters: without a central database, every workflow is an island. Data lives in whatever tool created it and nowhere else. With Supabase as the single source of truth, my n8n workflows can read and write shared state, my dashboards can query real data, and I have one place to look when something goes wrong.

It also has row-level security and a real-time subscription feature, which I use for the live operations dashboard that shows me the state of the whole portfolio at a glance.

The secondary automation: Make

For lighter workflows that don't need custom code, I use Make (formerly Integromat). It has a cleaner visual interface than n8n for simple linear automations, and its library of pre-built integrations is larger. I use it for things like social media scheduling, simple form-to-spreadsheet flows, and notifications that don't warrant a full n8n workflow.

Using both isn't redundancy — it's picking the right tool for the complexity level. Make for simple, n8n for anything with branching logic or custom transformations.

The voice layer: ElevenLabs

For content work — specifically turning written content into audio for short-form video and voiceovers — I use ElevenLabs. The voice quality is good enough that it's not distracting, which is the bar. I integrate it via API into content generation workflows.

This isn't core to store operations, but it's core to the content side of the business, and it's a good example of an AI tool that does exactly one thing well and doesn't try to be a platform.

The notification layer: Telegram

This is the least "AI" tool on the list and one of the most important. Every automation that needs to tell me something sends a message to a Telegram bot. Stock alerts, error notifications, daily summaries, exception flags — all of it.

The value: I don't have to check dashboards. The system pushes information to me when I need it, and stays quiet when everything is fine. My phone is the interface to the whole operation. If Telegram is quiet, the stores are running fine.

What's NOT in the stack

Worth saying explicitly, because it's as informative as what is:

No "AI agent" platforms. The autonomous agent tools that promise to run your whole business are, in my experience, not reliable enough for production yet. I use LLMs for specific bounded tasks, not open-ended autonomy.

No AI website builders. For anything that matters, hand-built or block-based beats AI-generated. The output isn't maintainable.

No dedicated "AI CRM" or "AI analytics" products. These are usually a thin AI layer on top of a database. I'd rather own the database (Supabase) and add the intelligence layer (Claude API) exactly where I need it.

The principle behind the stack

The thing that ties this together: use deterministic tools for deterministic work, and use AI only for the parts that genuinely require understanding. Most of a workflow is moving data from A to B based on rules — that's n8n, that's code, that's reliable. The one step that requires reading a message and understanding intent — that's the LLM.

The mistake I see people make is reaching for AI for the whole workflow. It's slower, more expensive, and less reliable than using AI for just the step that needs it. A small stack, applied precisely, beats a big stack applied everywhere.