The freelance and consulting world is currently obsessed with "doing more with less." But doing more with less usually just means you are doing more. When I transitioned from designing for corporate giants like Marvel and Disney to running my own remote consulting practice, I realised that subscribing to 15 different SaaS tools didn't free up my time; it just turned me into an administrator of my own subscriptions.
That changed when I stopped renting software and started orchestrating my own local AI. I built my first persistent local AI agent—let's call him "Hermes"—in under 60 minutes. He doesn't cost a monthly subscription, he doesn't hold my data hostage on a cloud server, and most importantly, he handles the grunt work while I step back onto the balcony to direct the strategy.
Here is exactly how I built a local AI employee and why you need to do the same.
The Problem with Cloud AI and "Walled Gardens"
Most creatives are currently using AI backward. They treat ChatGPT or Claude like a super-powered search engine, pasting in prompts and copying out text.
The problem is twofold:
- No Memory: You have to explain who you are and what you do every single time.
- Data Hostage: Your client data is being fed into a corporate machine.
To build a real AI employee, you need persistence and data sovereignty. You need a model that runs locally on your machine, reads your files (with permission), and retains context between sessions without sending your intellectual property to a server farm.
My 60-Minute Architecture
You don't need a computer science degree to set this up. The open-source community has done the heavy lifting. Here is the stack I deployed in an hour:
1. The Engine: Ollama
Ollama is a free, open-source tool that lets you run large language models locally. I downloaded it, opened my terminal, and pulled a lightweight open-source model (like Llama 3 or Mistral). Because it runs on my Mac's local hardware, my client NDAs remain perfectly intact.
2. The Brain: A Local Vector Database
To give Hermes memory, I hooked him up to a local vector database (using a tool like AnythingLLM or a custom Python script). I dumped all my brand guidelines, previous proposals, and email templates into a folder. The vector database indexes this folder. Now, when I ask Hermes to draft a proposal, he doesn't guess my tone—he perfectly mimics the proposal I wrote last week.
3. The Hands: API Webhooks
An employee needs to act. I follow a strict "No API, No Buy" rule for software. I connected Hermes to n8n (an open-source alternative to Zapier) so he can access my email and Notion databases. If a client emails an invoice, Hermes sees it, extracts the data, and updates my Notion tracker.
The "Human on the Balcony" Shift
When you run a persistent agent like this, your role changes. You stop being the person writing the emails or formatting the PDFs. You become the director. You stand on the balcony, observe the workflow, and make strategic decisions.
In my charity work with Become Inspired, where we deliver creative education programs, handling the logistics of booking 120 schools used to take weeks of admin. Now, Hermes processes the intake forms locally, drafts the schedules, and flags conflicts. I just review and approve.
Your First Steps
Don't overcomplicate it. Download Ollama today. Run a model locally. Start moving your daily "slop" tasks to a system that you actually own. The future belongs to the orchestrators.
FAQ
Q: Do I need to know how to code to build a local AI agent?
A: No. Tools like Ollama and AnythingLLM have graphical interfaces that make running local AI as simple as installing a normal desktop application. You only need code if you want to build custom API integrations.
Q: Is a local AI agent safe for client data?
A: Yes, local AI is the safest option. Because the model runs entirely on your own hardware, no data is ever sent to a third-party server, ensuring compliance with strict NDAs.
Q: Can my local AI agent send emails for me?
A: Yes, by connecting your local agent to a local automation tool like n8n, you can grant it API access to draft or send emails based on specific triggers. I recommend keeping a "human in the loop" to approve emails before they send.