How we automate operations with local AI and n8n

Two-week sprints don't leave room for admin bloat. If a task takes more than a few minutes and repeats, we automate it. Our operations practice exists for this reason, but we practice what we preach internally. So we built a lightweight automation layer using self-hosted AI agents and n8n workflows. It runs on a single server in our office. No cloud costs, no third-party access, and full control over our data.

Why local AI instead of cloud APIs

We tried cloud-based automation (Zapier, Make, OpenAI APIs). They work, but they add latency, cost, and privacy risk. For client-facing data like project specs or financial figures, we don't want them touching external models. So we run Ollama with open-weight models (Mistral, Llama 3) and LocalAI for embedding search. Inference is slower than GPT-4, but for structured tasks, classify an email, extract a deadline, draft a summary, it's fast enough. And it costs nothing beyond the hardware.

Our core automations

Each workflow is an n8n pipeline triggered by a webhook, schedule, or Slack command. We run them on a Docker stack with PostgreSQL as the n8n database. Here's what we automate today:

• Client onboarding: When a new signed agreement hits our CRM, a webhook triggers n8n. It creates project folders (Google Drive), sends a welcome email, sets up a Slack channel with the client and our team, adds calendar blocks for kickoff and weekly checkpoints, and writes a summary to our internal Notion database. Total time: 90 seconds. Previously: 45 minutes of manual clicking.
• Weekly client reports: Every Thursday, n8n pulls hours logged, tasks completed (from Linear), and budget usage (from our custom dashboard). It feeds the data to an Ollama model that drafts a plain-language progress note. We review and send. No one spends an hour writing status updates.
• Internal standup summaries: Each morning, a Slack bot asks everyone for their focus and blockers. Responses go into a Notion database. At the end of the week, n8n aggregates them and Ollama produces a one-paragraph summary for our weekly retrospective. We catch drifting priorities early.
• Invoice reminders: If a payment is 3 days late, n8n sends a gentle in-app notification. If still unpaid at day 10, it flags the project lead, no awkward emails from our side. The AI drafts a polite follow-up, but a human always reviews before sending.

What we don't automate

We don't let AI write client-facing deliverables. It can draft an internal summary, but not a brand guide or a UI spec. We don't automate decisions that affect trust: pricing exceptions, scope changes, or hiring. Every automation has a human review step. n8n's approval gates make that easy. If a workflow fails twice, it pauses and alerts us.

The real win

Automation isn't about replacing people. It's about removing friction so we can focus on work that matters, design, engineering, strategy. Our operation runs with a team half the size of a typical studio our revenue. That's because we offload the predictable to machines. The result: we ship in two weeks, not two months. And we never ask a client to wait while we update a spreadsheet.