How I Built an AI Operating System for My Business

After the first 30 days of running my OpenClaw setup, it was already saving me €3.072,- per year in cancelled or reduced subscriptions on tools and services. This does not include time savings on my own side. That being said, the time save is probably the most important part here.
This is not a magical story of "How AI runs my business for me". There are a lot of pain points. I had to iterate a lot on the setup, and I still have to review what is going on. But in the end, the setup truly handles many things for me now which I struggled with in the past.
The first part of this blog series I want to focus on how I came to OpenClaw and how I reached the insights I have right now. Also I want to note that OpenClaw was the first AI Assistant. Currently there are already other options which can help in the same way and the market is rapidly evolving. Pick one based on your needs and see what it can do for you.
The Problem With Automations
A few months ago in January I was setting up some automations for my business. I was using Make.com and AntiGravity to help me work out the automations in JSON to prevent most of the manual work in the visual editor. There were many things I wanted to set up: manual admin work, dashboards and reports, product data, customer workflows etc.
The problem was not that Make.com was hard to understand. It's actually very easy to use. However, I was spending hours building automations that still had flaws in them.
Rate limits, Edge cases, Weird failures, small logic gaps. Things which worked one time, but not another time. I was definitely no automation expert. I overcame the issues, but felt I spend 4 hours automating a task that saved me 15m a day, and then with the added maintenance I started feeling I was wasting my time on automating a bunch of small tasks which in total were not the operational automation layer at all yet which I had envisioned.
Enter OpenClaw
I stumbled upon OpenClaw in February on Reddit. At that time it was still called ClawdBot, soon to be renamed to MoltBot and then again renamed to OpenClaw.
The idea was so simple that it caught my attention: install an AI agent on a machine, be able to speak to it from everywhere and it is always online and therefor can run things on a schedule (cron jobs). For a quick demo I installed it right away on my MacBook. Within 30 minutes I had set up OpenClaw with my GitHub Copilot subscription, hooked it up to Telegram and I was happily chatting away with it from my phone in bed, while my laptop remained on for the connection in another room.
This blew my mind. I was able to control an AI agent from my phone. This means I can run my business much more easily from my phone now. I could now speak to an agent which is always online. Generating reports, triggering workflows, run tasks on a schedule, query APIs, write code. All of this without sitting with my laptop. I already do a lot of my work out of the house. I spent time on the subway, in waiting rooms, coffee shops, taxi's. I usually carry around my laptop in order to get something done in these periods, and I got genuinely excited about the possibilities
Moving OpenClaw to a VPS
I kept the macbook setup for only two days. There are two obvious issues with installing OpenClaw on a personal machine.
- It has access to your personal data. This is always a risk. It could get leaked or deleted or corrupted in some other ways.
- The laptop needed to stay on for the agent to work. For an always-on agent I need something that is actually always on.
A lot of people were buying a Mac Mini for this use case, but I felt it didn't need a Mac Mini. That's quite an expense for a simple AI assistant. Perhaps, if your setup is refined enough and makes enough money for you the cost could be justified.
Since I host my websites with Hostinger, and they offer cheap VPS's for $4,99 a month I decided to get one and run OpenClaw on that. I set up basic security, set up a git repo for the openclaw configuration and I was good to go. In a later part of this series I will dive into the exact setup I used for all of this. For this post the important takeaway is that I moved OpenClaw away from my personal computer into a dedicated VPS environment where it had full access and could run 24/7.
The Basic Architecture
The simple version of the setup looks like this:
Telegram -> OpenClaw -> VPS -> Hanno / Forge / Muse -> tools and repos
| Layer | Role in the setup |
|---|---|
| Telegram | The interface I used from my phone to send instructions and review what came back. |
| OpenClaw | The assistant runtime that receives the message, keeps context and turns requests into work. |
| VPS | The always-on environment where OpenClaw runs without depending on my laptop. |
| Hanno / Forge / Muse | The agent team: orchestration, development and design. |
| Tools and repos | The places where the work actually happens: GitHub repos, project files, APIs, dashboards and business tools. |
Building my Agent Team
Once I had my setup up and running I wanted to have more than one agent. I saw people online building large agent teams. I didn't want to go all out like that. I figured I would create an agent for each job I would usually do myself or hire someone for.
I started with three dedicated agents.
Hanno -- The Orchestrator
Hanno is the main agent. He is always the one I speak to, in different chats. I named this agent after my late father. My father was always the entrepreneur, the thinker, the one who has the vision and delegates tasks to others to get things done. Naming this agent after him felt like a small tribute to his legacy.
Hanno plans and delegates. I tell it what I want and it thinks through the issue and context and either handles the task himself if it is simple or delegates to the other agents as subtasks when necessary.
I didn't want to deal with a chat with a coding agent, a chat with a design agent etc. I wanted to set up chat sessions by subject and simply say what I want and it would all be handled for me. Creating a single agent to speak with me which is "trained" to challenge me and think through issues sounded like the perfect solution to that.
Forge -- The developer
Forge is the developer agent. I stole the name from every post on the internet about coding agents. He specializes in coding, architecture, git, tests and technical implementation.
I gave Forge access to all of my agent skills I already was using with my other AI tools in order to write code the way I want it to be written. It was immediately very capable at writing code in my style, properly following code conventions I adhere to, and using code quality tools such as ESLint, Prettier and Fallow to check its work and make sure I would approve of the code.
Muse -- The designer
Muse is the designer agent.
She is configured around clean, minimal, premium design work. I gave her access to Firecrawl to inspect websites and extract their branding and design system. I also connected her to Google Stitch so she could quickly generate design concepts or full designs herself.
A lot of my work sits between a client not knowing what they want, but wanting to maintain their current branding. I was now able to quickly scrape existing branding information from a client's website and spin up new designs according to their wishes. This allows for much faster communication and handling of new website discussions.
The First Real Test
While I was busy setting up OpenClaw on the VPS, I was speaking to a potential client. He did not have the budget available for what I usually charge for what he wanted. So, I told him upfront I usually charge roughly triple what he was willing to pay, but I would like to try to give him a good result using basically only AI. This would give me much less hours of work, and he would still get a quality site for the price he wanted. He was perfectly ok with this experiment.
I started a new chat with my OpenClaw agent, explained what the client wanted, passed along the website url and wishes for the new site, which pages he wanted, which functionalities he wanted and it started thinking. It came back with the answer that Muse was generating designs based on the provided context, and that new improved SEO texts were written and passed to Muse for use in the designs.
I came back 30 minutes later to my laptop with a link to the Google Stitch project. I shared this with my client. He had minor feedback but loved the direction. I passed the feedback in, and in a total of two hours I had designs for both mobile and desktop which the client was very happy with.
It was time for the part I was expecting to go wrong. Turning a custom webdesign into a WordPress theme. I set up an empty WordPress site on the VPS and told OpenClaw to take the Google Stitch designs and make a very clear plan on how to turn this into a WordPress theme. I wanted it not to try to make the pages, but to think about each component on the page and make that into a reusable set so we have an actual Design System in place before we start stitching things together. It made a bunch of documentation which looked reasonable and I set it to work.
I truly expected a lot of hallucination here, but what came out was pretty solid. I made corrections here and there on the components, strengthening the agent skills further according to my wishes. But by the end of the day I had all the templates for all the pages ready. It had generated custom post types for the different types of content, and we had a beautiful responsive site ready from scratch in less than a day.
I was very happy with the result, both visually and the code quality. The client was very happy as well. He had some feedback on some of the content, which he was able to fix in the WordPress editor himself. I had to do some of the work myself in the editor as well, setting up plugin license codes, adding WPForms and ACF plugins etc. but it was 90% automated.
This made me really interested in turning it towards my own business instead of client work.
Automating my E-commerce Business
I sat down with my morning coffee the next day and made a short plan of where I believed OpenClaw could help me the most in terms of automating tasks, improving productivity, saving costs and generating more revenue/profit.
The first list looked roughly like this:
- Build a productivity dashboard that replaces manual Notion/Todoist interaction and helps my ADHD brain work more efficiently.
- Send me a daily morning briefing with reports on issues in the business, the most important todos (derived from Notion, E-mail and WooCommerce).
- Automate my accounting (upload incoming invoices to my accounting software, connect webshop order invoices to payments, report on issues, etc).
- Analyse my campaigns on Google Ads and give through insights on potential improvements and issues.
- Replace some third-party tools I was paying for that were not a perfect fit for my business, but still costed money.
After 30 Days
Just one month after setting up OpenClaw it was saving me €3.072,- per year.
The breakdown of this number looks like this.
| Area | Annualized saving |
|---|---|
| Reduced accounting/admin support | EUR 2,160/year |
| Cancelled Make.com subscription | EUR 192/year |
| Moved hosting from Sevalla to Dokploy/self-hosted | EUR 300/year |
| Replaced Todoist with a custom ToDo app | EUR 60/year |
| Replaced SaaS invoice generator | EUR 180/year |
| Reduced Mailchimp costs | EUR 180/year |
| Total | EUR 3,072/year |
The most important aspect is not even reflected in this list. OpenClaw helped me build a web application to be my "Business Hub" as I like to call it.
It's basically a bunch of dashboards where both me and OpenClaw can control everything about my business using the connectors to WooCommerce, HubSpot, Google Analytics, Google Ads, Google Workspace, Gmail, my payment service provider, and my accounting software. I will list some of the features:
- See different marketing dashboards specific for my business through data aggregated from my tooling.
- See user accounts with all their data from WooCommerce/Mailchimp/HubSpot. Here I can quickly delete spam, see which clients are active, fill in missing data.
- A customer dashboard where I see my top clients, clients who are at risk of leaving (with an option to add them to a win-back campaign).
- Automatically generate invoices in my accounting software for each WooCommerce order. This replaced a paid third-party app.
- Stock management. I order stock in batches by container. It keeps track of the sale velocity of items and how long it takes for stock to arrive. It checks every day and once I am at risk enough and how enough stock I need to fill a container, it prepares a draft email to my supplier for me with all the needed information which I can send with a single button press.
- Marketplace connection. It takes orders on marketplaces such as Amazon and automatically connects them to our usual logistics and accounting flows. This was a lot of manual work for us in the past.
Besides the direct business side it also made a custom "Todo app" for me based on my Notion. I track all my goals, projects and tasks in Notion, but I am terrible at keeping the data clean and in-sync. OpenClaw sees the scheduled data, due date etc of the tasks and makes sure that all the issues are reported in the app. This has been a tremendous help for me in keeping myself organised around administrative tasks, running my own business, doing freelance work and personal things which used to fight for my attention.
What This is Not
I want to be careful with the framing here. OpenClaw is not a magic bullet. It does not autonomously run my business for me. Most things I did with it could have been done before with other tools. However, it has made it so easy to set up systems where before it was more difficult, more time consuming and you needed to be more technically proficient to set it up that I fully understand why OpenClaw blew up so much.
What I would do differently if I could restart
I figured out a lot of stuff on the go here. I should have planned my OpenClaw setup better. I started out on Telegram, but moved to Discord later. I blindly updated OpenClaw a few times which broke my setup costing me a bunch of time to get everything working again. I could have put more thought into that.
Most importantly, I would have focussed on setups where OpenClaw can make the most impact. I spend a lot of time on the setup where it was cool to automate something, but the flashiest automations were in the end the least useful and the simple automations were the most valuable. I would have prioritised by time and money saved or potentially by money earned from the automations.
What Comes Next
This post is the first part of a series on building an AI back-office.
The next posts will go deeper into the practical setup such as:
- How I installed OpenClaw on a VPS and how I handled the security
- How I set up Discord and OpenClaw to keep my projects, discussions and tasks organized
- How I structured my full multi-agent team
- Which workflows I actually use and how they work in-depth
- What I changed after now running OpenClaw for 3+ months
No theory-only hyping up tools, no pretending OpenClaw is anywhere near perfect.
I just want to share my journey and setups which have helped me so far, explain what worked for me, what broke, what didn't work and what I am changing as I use it.
Frequently Asked Questions
What is an AI back-office?
An AI back-office is an operational layer of AI agents, workflows and internal tools that helps with recurring business work such as reporting, admin checks, customer data, invoices, stock alerts and task triage. It should support decisions, not run the business without review.
How do you build an AI back-office for a small business?
Start with one always-on assistant connected to a chat interface, then give it access to a limited set of tools, repos and workflows. The best first version solves repeated tasks that already cost time or money every week.
What business tasks should you automate first with AI?
Good first candidates are daily reports, campaign checks, invoice handling, customer data cleanup, stock monitoring, support triage and replacing paid tools that do not quite fit your workflow. Prioritise tasks by time saved, cost reduced or revenue created.
Should an AI assistant run on a VPS or a laptop?
A laptop is fine for a quick prototype, but a VPS is usually better for an AI back-office because it can run 24/7 and keeps business automation separate from personal files. That makes scheduled workflows and remote access more reliable.
Do you need multiple AI agents for business automation?
No. A single well-configured assistant can already be useful. Multiple agents only become valuable when the roles are clear, such as one agent for orchestration, one for development and one for design or content work.
I'm turning this setup into an AI Back-Office Planning Worksheet. Subscribe below or reply to me on socials if you want it when it is ready.
