AI systems 9 MIN READ
What Is an AI Operating System for Business? How I Built One
How I built an internal AI operating system for Nirox Labs: business context, working methods and reviewed output in one place. An internal project.
In this note
An internal Nirox Labs project, not a client case study.
Running a professional services business on your own means you are always the bottleneck.
Every proposal starts from scratch. Every new client gets whatever version of the pitch you happen to have in your head that day. If you’re busy, quality slips. If you’re ill, output stops. Everything lives in one person’s memory — which means it’s fragile, inconsistent, and impossible to scale.
I’ve been running Nirox Labs solo for a while now. Nirox Labs builds custom AI applications, knowledge assistants and agents. This is the system I use to run it. At some point I asked myself an obvious question: if I help other companies encode their operations into intelligent systems, why haven’t I done it for my own?
So I built one. What I ended up with is what I now call an AI operating system for business — not a chatbot, not a productivity app, but a structured system that runs underneath everything I do. I built it using Claude Code, Anthropic’s agentic AI tool, as the runtime.
This post explains what that actually means — and what it looks like in practice.
What an AI Operating System for Business Actually Is
The term sounds more abstract than it is. Forget the tech industry’s version — this isn’t a software platform from a big vendor.
Think about what an operating system does on a computer. It’s the layer that everything else runs on top of. It manages memory, coordinates processes, enforces rules about how things work. You don’t interact with it directly most of the time — you just notice when things work smoothly and reliably.
An AI operating system for business does the same thing, but for how you work.
It encodes the knowledge, judgment, and processes that usually live only in your head — and makes them available, consistently, every time you need them. It means you don’t have to remember how you price things, what tone you write in, how a proposal should be structured, or what questions to ask during a scoping call. The system knows.
This isn’t about replacing your judgment. It’s about making your best judgment available every time — not just when you’re sharp and unhurried.
The Three Layers
The system I built has three distinct layers. I’ll explain each in plain terms.
Layer 1: Context
This is the knowledge layer. It contains everything the business knows about itself.
How I position the business. Who the ideal clients are. What I charge and why. What I won’t do. How I write — the tone, the words I avoid, the structure I use for different types of output. What my delivery process looks like. What objections I typically face and how I handle them.
None of this is glamorous. But most businesses leave it entirely implicit — carried in the founder’s head or distributed across old emails and half-remembered conversations. Making it explicit and structured is the foundation of everything else.
In practice, this lives as a set of structured documents that the AI reads automatically before any task begins. When Claude Code starts work on a proposal, a blog post, or a scoping estimate, it already knows the business context — without me having to explain it every time.
Layer 2: Skills
If the context layer is knowledge, the skills layer is capability.
A skill is a defined, repeatable process for a specific type of work: writing a proposal from call notes, researching SEO keywords, estimating project scope from a brief. Each one exists as a structured document — inputs, steps, outputs, edge cases — written precisely enough that the AI executes it the same way every time.
The key difference from a prompt is specificity. A prompt is a rough instruction. A skill is a process. It defines what to check before starting, what order to do things in, what the output must contain, and what to do when something unexpected comes up.
Before I had this system, I was reinventing these processes every time. Now they’re consistent. And because they’re explicit documents, I can improve them — when a proposal wins, I can understand why and update the process.
Layer 3: Memory
This is the part most people miss.
Every proposal I write, every project I complete, every piece of content I produce — it all goes back into a Postgres database. The system can query this when working on new tasks. It knows what I’ve written about before, so I don’t repeat myself. It knows what a similar project looked like, so the next estimate is more accurate. It knows what has worked, so it can apply those patterns going forward.
This is where the compounding happens.
An AI operating system for business that doesn’t have memory is just a fancier way of doing things once. When Claude Code can reference six months of past proposals, project notes, and content — that’s when the system starts to feel less like a tool and more like institutional knowledge.
What Actually Runs It
The AI model at the centre of this is Claude, specifically running through Claude Code — Anthropic’s agentic CLI tool designed for multi-step, autonomous work.
What makes Claude Code suited to this is that it can read files, call external tools, query databases, write and execute code, and chain multiple steps together — all within a single session. It’s not a chatbot you talk to. It’s an agent you give a task to, and it works through the process using whatever it needs.
The reason this matters for the OS architecture is that Claude Code reads the context layer automatically at the start of each task, selects or follows the relevant skill, does the work, and stores the output — without me having to manage those steps manually. The OS works because the underlying tool is capable of acting on structured instructions, not just responding to prompts.
This isn’t a sponsored mention. Claude Code is simply what I evaluated and chose — if a better tool emerged tomorrow, the architecture would stay the same and the runtime would swap out.
What This Looks Like Day-to-Day
Concrete outcomes, because abstract benefits are useless:
Proposals. What used to take me 3 hours now takes under 30 minutes — and the output is more consistent, better structured, and better aligned to what actually wins work. The system draws on past proposals, applies the right pricing logic, and follows the same structure every time. I review and refine, but I’m not starting from a blank page.
Content. Blog posts, LinkedIn posts, and other content used to involve a lot of false starts. Now there’s a process: here’s the topic, here’s the keyword, here’s what I’ve already written so I don’t repeat myself, here’s the structure. The system does the heavy lifting. I supply the judgement and voice.
Pricing. No client ever gets a different number because I was tired when I answered their email. The pricing logic is in the system. It knows what I charge for what, and why. It applies consistently.
Institutional memory. When I work on a project, the context goes into the database. If I come back to a similar problem six months later, the system remembers. That’s the kind of thing that used to be locked in the memory of a team. Now it’s available even as a solo operator.
What this means if you’re considering a build
Here’s the honest reason I’m writing this post: it’s the clearest example I have of the approach I’d bring to a project, and it’s all from running my own business, not a client engagement.
What I learned applies to any project like it. Encode what is already sound, because AI makes decisions consistent, not better. Start with the context layer, because most of the value is in making explicit what only lives in people’s heads. Expect maintenance, because a system that reflects reality has to be kept current. And test it on real work before anyone relies on it.
The specifics of your business, processes and knowledge would be entirely different. The questions I’d ask first are the same.
The Honest Limitations
This wouldn’t be a useful post if I only told you the upside.
It took real time to build. This isn’t a template you can grab and install. Building it properly required thinking hard about how the business actually works — what the real processes are, not the idealised versions. That part can’t be rushed.
It requires maintenance. When how I work changes, the system needs updating. That’s not a flaw — it’s just what it means to have something that actually reflects reality. But it’s ongoing work, not a one-time project.
It only works because the underlying judgment was already sound. AI encodes decisions. It doesn’t invent them. If your pricing is inconsistent, your positioning is unclear, or your delivery process is chaotic — encoding those things into a system will just make the chaos more consistent. You need to do the business thinking first.
The payoff is real. But it’s not magic, and it’s not instant.
What to Do If You’re Thinking About This for Your Business
An AI operating system for business doesn’t have to be built all at once. Most businesses would benefit enormously from just codifying the context layer — making explicit what currently lives only in people’s heads.
That alone changes how consistent your output is, how quickly new work can be done, and how much you’re relying on any single person’s memory.
If you’re spending significant time on work that should be repeatable — proposals, reports, client communications, scoping, research — and you’re doing it differently every time, that’s the problem an AI OS is designed to solve.
I help businesses figure out where to start and what’s worth building. If you want a conversation about what this could look like for your operations, book a free call →. You can also read the free RAG guide.
No pitch. Just a conversation about whether this is the right problem to solve right now.
Liam Silk is the founder of Nirox Labs, an AI consultancy based in Hull, UK, designing and building custom AI applications, knowledge assistants and agents for businesses.
Practical notes from Nirox Labs.
Applied AI engineering, from Hull to wherever you work.
A LITTLE MORE CLARITY
Questions on this topic.
What is an AI operating system for business?
An AI operating system for business is a structured system that encodes how your business thinks, speaks, prices, and delivers — and makes that available consistently through AI. It's not a chatbot or a productivity tool. It's the layer that everything else runs on top of, so your output stays consistent regardless of how busy you are.
What is Claude Code and why did you use it?
Claude Code is Anthropic's agentic AI tool — designed for multi-step, autonomous work rather than simple question-and-answer. It can read files, call external tools, query databases, and chain actions together. That makes it suited to running a structured OS: it reads the relevant context, follows the relevant skill, does the work, and stores the output.
Do I need to be technical to build something like this?
Not necessarily — but you do need to be willing to think clearly about how your business actually works. The hardest part isn't the technology. It's making your processes, pricing logic, and judgement explicit enough to be encoded. If you can describe the rules, the technical side can usually be built.
How long does it take to build an AI operating system?
It depends on the complexity of what you're encoding. A basic context and skills layer comes first. A fuller system with memory and compounding outputs takes longer, and it keeps improving as more work flows through it. The foundation is worth getting right rather than rushing. For a client project, the timeline is agreed as part of the costed scope.
Can this work for my business?
Yes. The architecture — context, skills, memory — applies to any business that has repeatable processes and knowledge that currently lives in someone's head. It suits any knowledge-heavy work where consistency matters: operations teams, consultancies, and any business where output quality depends on one or two people's judgement.
