What an AIOS is and why a business needs one
Sergei Pak · designs AI operating systems · русская версия
Say "AI operating system" and people picture a chatbot. A chatbot sits there until you ask it something. An operating system has already sorted your inbox by six in the morning, recalculated the cash position after the overnight statement, and left you the two or three questions that actually need you. It works while you sleep, and you open the laptop to finished work.
The five layers
They stack in order, or the top ones float. One by one:
- Context. The system knows your business: counterparties, contracts, numbers, the history of decisions. Permanent, structured memory. "I dropped the files into a chat" doesn't count – tomorrow the chat forgets all of it.
- Data. Emails, chats, bank statements, accounting feeds land in one place on their own, no copy-paste by hand.
- Intelligence. Models run on top of the data: sorting the inbox, hunting for risk, reconciling documents. The easy stuff goes to a local model, the hard stuff to the cloud.
- Automation. The report builds itself, the digest shows up in the morning, an expiring contract surfaces before it turns into your problem.
- Assembly. The system grows: a new process is a new module in a few days, not a six-month project.
What an AIOS is not
The word is new and sits next to similar ones, so let me draw the lines. Not a chatbot: it waits to be asked and forgets the conversation, while an AIOS runs on a schedule and remembers. Not SaaS: in SaaS you enter the data by hand, an AIOS gathers it itself. Not an AI agent: an agent does one task in the moment, an AIOS lives between tasks as a five-layer system with memory. Not RPA: that repeats clicks by a fixed script, while an AIOS understands the context and judges by it – the final call stays with the human.
What it gives you in practice
I track it as "autonomy hours": how much of the day the processes run without me at the screen. About two hours now, four by December. Those hours used to disappear into sorting mail, reconciling documents and the eternal "so where do we stand with this counterparty".
The second metric is duller, and accountants love it: money. The whole system spends around half a dollar a week on LLM calls, because a local model on my own Mac handles 81% of them. I wrote up how that works in a separate post.
Where to start
Not with the model. A system that doesn't know your business writes pretty text about nothing, no matter how clever the model behind it. So the first step of any rollout is the boring one: build the context. Who you are, how your processes run, where the data sits, who makes the calls. The model plugs in last, and you can swap it later without touching anything else.
FAQ
- What is an AIOS in plain terms?
- An AIOS (AI operating system) is a software layer around a person and their work: it sorts incoming mail, reconciles bank and ERP numbers, remembers contracts, and surfaces only the decisions that need a human. Unlike a chatbot, it runs on a schedule and on events, not on request – the person leads, the system works. It is built from five layers: context, data, intelligence, automation, assembly.
- How is an AIOS different from a chatbot?
- A chatbot answers when asked and forgets the conversation. An AIOS runs on schedule and on events: it sorts mail at six in the morning, recalculates the cash position after every statement, and raises an overdue contract on its own. Persistent memory of the business instead of a context window.
- How is an AIOS different from SaaS, an AI agent and RPA?
- SaaS is a tool you enter data into by hand; an AIOS gathers the data itself and works on top of it. An AI agent does one task in the moment; an AIOS is a permanent five-layer system with memory between tasks. RPA repeats clicks by a fixed script; an AIOS understands the context and takes on the routine, while the final call stays with the human.
- How much does running an AIOS cost?
- My system spends about half a dollar a week on LLM calls: 81% of requests run on a local model on a regular Mac, and only the hard tasks go to the cloud. The expensive part is not the models but the rollout: building the context layer for a specific business.
- Where do you start an AIOS rollout?
- With the context layer, not with model selection. The system first has to learn the business: counterparties, processes, where the data lives, who decides what. The model plugs in last and can be swapped without redoing the rest.