It’s easy to think of AI as a tool. Something you open, ask a question of, and close again.
It isn’t just a tool. It’s a whole new way of working, and the fastest way I’ve found to explain that new way of working is to relate it back to the way you work right now. You already know how an office runs. You know what a good brief looks like, what a specialist is for, and why you’d never expect a new starter to do great work on day one with no files and no idea how you like things done.
So here’s how I explain it. You’ve moved into a new rented office, and you’re the boss.
The building comes with a few things the landlord has fixed in place that you can’t change. Everything else (the desks, the filing, the keys [or combination locks], the specialists, the instructions on the wall) is yours to set up and distribute within your team, and yours to take with you when you leave.
When it comes to using AI, most people move in, sit at a bare desk, are reluctant to invest in their tools or training (that’s training for both the AI team they’re developing and themselves), and then wonder why the work’s so generic.
The office isn’t the problem. The empty office is.
Key Takeaways
- AI isn’t a single tool. It’s an ecosystem you’ll be doing business in, and every part of it maps to something you already understand from running an office.
- Your main AI tool of choice is, in the majority of cases, a rented building. You don’t own it, so don’t screw your business to it. Your knowledge files, skills and instructions are furniture you can take with you.
- Skills are your specialist staff, connectors are the keys you hand them, projects are your desks, the context window is the desk surface, and your AI Brain is the filing cabinet.
- The default setting in the AI office is amnesia. The filing cabinet, memory, saved briefs and artifacts all exist to get around it.
- If you wouldn’t expect a new starter to do good work with no brief, no files and no access, don’t expect it from AI either.
What does the AI office look like?
I find myself explaining these components to people all the time, one at a time, in workshops and on calls. The trouble with explaining them one at a time is that AI stays a weird, invisible concept you can’t quite wrap your head around. You learn what a skill is, then what a context window is, then what memory is, and none of it connects.
So I mapped the whole thing out in one place. Here’s the office at a glance, and the rest of this post walks through each room.
| In the office | In your AI setup | What it’s for |
|---|---|---|
| The building | Your AI tool (Claude, ChatGPT, Gemini, Copilot and so on) | Where the work happens. You choose it, you don’t own it |
| Fixtures | The tool’s built-in capabilities | What it can and can’t do. Not yours to change |
| Furnishings | Your knowledge files, skills and instructions | Everything you set up. Yours to take with you |
| The cafe down the road (no kitchen) | A second AI tool for one specific job | Covering a gap your main tool doesn’t fill |
| The boss | You | Setting the work, reviewing it, signing it off |
| Specialist staff | Skills | Experts who each deliver one specific outcome |
| The calibre of staff | The model (Opus, Sonnet, GPT and so on) | The grade of talent doing the work |
| The keys | Connectors | Access to the tools where the work happens, with guardrails on what they can do |
| A desk | A project | One area of work with its context attached |
| The desk surface | The context window | How much can be worked on at once |
| A job | A chat | One task, done at one desk |
| The brief | What you hand over | The problem, plus the context needed to solve it well |
| The filing cabinet | Your AI Brain | Your business knowledge, in plain files you own |
| What they remember about you | Memory | How you work, built up over time |
| The handbook | Your instructions | The rules, placed where they’re needed |
| The org chart | How your skills hand work to each other | Turning separate specialists into a team |
| What leaves the desk | Artifacts | The work you keep, share and update |
| The standing appointment | Scheduled tasks | Recurring jobs done without being asked |
The building: your AI tool
Claude, ChatGPT, Gemini, Copilot, Grok and the rest own the buildings. Anthropic, OpenAI, Google, Microsoft and SpaceXAI (Elon Musk) are the landlords. You choose which one to move into based on the capabilities you need, possibly also weighed up against a decision around which landlord you choose to trust.
Fixtures you can’t move
Every building has things fixed to the property. You can’t shift the toilet to where the kitchen is. Those are your tool’s embedded capabilities: how it reasons, what it’s naturally good at, and what it simply won’t do.
You work with them. You don’t get a vote.
That includes when they change. The landlord can renovate whenever they like. Features you’d built your whole workflow around can be deprecated, reworked or abruptly retired with little or no notice. Like the way Custom GPTs are being phased out. You’ll be told it’s an improvement. You won’t be asked.
Which is exactly why you don’t want your business depending on anything fixed to a building you don’t own.
Furnishings you can take
Everything else is furniture. Your desks, your filing, your instructions, your specialists.
As long as you don’t fix your furnishings to the property, they come with you when you leave. That’s why I am constantly banging on about the importance of building your knowledge files, your skills and your instructions as plain files you own and keep outside the tool. They’re usually plain little markdown files, which is exactly why any tool can read them. Do that and you can plug them into whatever building you’re in next.
Different buildings are different, but most let you do similar work. Moving is a PIA, not a disaster. There will always be moving costs. Good setup is what keeps them small.
Where people come unstuck is building everything into the fabric of the building. When the tool changes or stops suiting them, they’re trapped. Not because moving is hard, but because they never set it up to be moveable.
When your building has no kitchen
Sometimes the building doesn’t have the room you need, so you go elsewhere for that one job.
Claude isn’t great at generating images, so I use ChatGPT for that. ChatGPT is the cafe down the road that serves me lunch. I don’t move out of the building because it has no kitchen. I just know where the good food is and go there.
That’s why so many business owners end up with more than one AI tool, and why that’s a sensible answer rather than a failure to commit.
The people: you, your specialists and the calibre of talent
You: the boss
You set the work, you review it, you sign it off. Nothing leaves the office without you.
That’s not a limitation of the technology. It’s how you’d run any team, and it’s the whole point of being human-led.
Your staff: skills
Skills are your expert specialists. Each one is brilliant at achieving one specific outcome.
That’s why I give mine human names with the job outcome built into the name. Cooper Copy Crafter. Maggie Magnet Maker. Barney Brain Bank. I know exactly who I’m calling on and what I’m calling them for based on their names.
A skill isn’t a manual. A skill is the expert who holds the manual. They’ve read it, they understand it, and more to the point they know how to apply it to the job in front of them. That’s why the work comes back consistent in standard and shape without being word-for-word identical. The standard is embedded in them. (If you’re new to skills, here’s what AI skills are and how they work.)
Each specialist comes in when the job needs their particular expertise, sits down at the hot desk beside you, does the work, and heads off again until you need them next. That’s why the office never feels crowded, even when you’ve got eighty of them on the books.
Three rules matter here.
Don’t let them fight over a job. If two specialists have overlapping remits, your tool won’t know which one to call in. Clear, distinct roles aren’t tidiness. They’re what makes the right expert turn up.
Don’t bloat them. You wouldn’t take an employee out to lunch, wine and dine them, stuff them full of food and then expect sharp work that afternoon. Give them the fuel they need for the job in front of them. Nothing less and nothing more. The temptation is to build one specialist who can do everything, because it feels efficient. It isn’t. A skill loaded up with ten jobs does all ten of them worse than a skill built for one. Discrete beats comprehensive every single time.
Hire one at a time. Build one, prove it earns its keep, then hire the next. Twelve half-built specialists are worth less than one that’s pulling its weight.
And realise that when you haven’t called a specialist into your chat, you’re just working with the general staffer directly. That’s fine for simple jobs, but you shouldn’t be doing most of your work like this or you’ll never get past average AI outputs.
The calibre of your staff: the model
The model (Fable, Opus, Sonnet, Haiku and so on, they seem to release a new model every other day) is the grade of talent filling the seat.
The same specialist performs differently depending on who’s in the chair that day. Run a skill on a junior model and you’ll get a junior job. Choose the grade to match the work and what you want to spend, exactly as you would when hiring.
Don’t be cheap, choosing a model that gives you fast and quick answers when you want to do deep research. And similarly, don’t pay over the odds for models that are designed for complex work and tough challenges when you’re asking for the answer to a simple question that you were just too lazy to Google.
The keys: connectors
Give your AI the keys to the tools they need, with guardrails in place, so the work gets done instead of handed back to you.
Your specialists can be brilliant, but if they can’t get into the rooms where the work happens, all they can do is hand it back to you.
“Here’s what I suggest you do.” “Here’s the copy to cut and paste.” That leaves you as the courier between tools, and that’s not the best use of your talents.
Connectors are the keys. They’re basically plugs that connect your AI to the other tools you use: your email, your calendar, your CRM, your project management tool, your Notion. Once they’re connected, your AI can do the next step for you instead of handing it back.
I have a whole heap of connectors in my account: Asana, Gmail, Google Calendar, Google Drive, Lead Activator (my CRM), Notion, Plaude, and Slack. Every day, Chandler, my AI Chief of Staff, goes into my Asana and my calendar and gives me a daily brief on where to focus.
Here’s what that looks like in practice… Someone told me my opt-in page for one of my Content Jumpstart training was out of date, as it still showed the live training date, which had passed. I asked Claude to make the page evergreen, checked and approved what it suggested, then asked it to update the page. Because it’s connected to Lead Activator, which that landing page is built through, it went and did it.
Before having my current AI office setup, a change like this would’ve taken me at least an hour to locate the changes, decide on the copy that needed updating, and then making the changes. My involvement was about two minutes.
Keys with a lockbox
This is where people freak out on me. Handing over the keys sounds scary.
But think about the first time you gave someone the keys to let themselves in and open up your office. You were probably a bit nervous, wondering whether you could trust them. So you put boundaries in place, maybe a lockbox for them to access the keys so they didn’t lose them or so that someone else could open up if they didn’t show up.
Do the same here. In Claude, you can set permissions for each action a connector can take. Take my Slack connector. The read-only actions are set to “go for it”, because AI needs that context to do good work. The write and delete actions were set to need my approval. After checking and approving its messages for a while, I thought, “Why do I need to check this?” So I gave it more leeway to write. It still doesn’t get to delete anything.
Start with the lockbox, then loosen the guardrails as it earns your trust.
The workspace: desks, the desk surface, jobs and briefs
The desk: your project
A project is one area of work, with its context attached. Different desks for different jobs.
A desk isn’t bare. It comes with a drawer, holding the files that only matter to this body of work. It has instructions pinned to the wall, the standing rules for anything done at this desk. And it has a surface, which is where it gets interesting.
The desk surface: the context window
Only so much fits on the desk at once.
As you pile more on, things get messy and start sliding off the edges. Not all at once, and not obviously. What you notice is that the work stops being as good as it was an hour ago. You’re repeating yourself. Instructions you gave earlier are being ignored. You’re getting frustrated because you know it did this properly before.
It’s the same as overloading an employee. Give someone forty instructions across three hours and they won’t hold all forty. They’ll hold the recent ones and the loud ones, and quietly drop the rest.
When a conversation starts going backwards, that’s your signal. Clear the desk, start fresh, and carry over only what matters. (I’ve written more about how the context window eats your usage if you want the detail.)
The job: a chat
A chat is one task, done at one desk.
When you walk away, the conversation is gone unless something got saved.
Which is why you set your office up so the important things are saved to the right place as you go. Usually that means somewhere outside the office entirely: a file on your computer, a record written to your CRM, a row in a Notion database, depending on what the output is. The keys are what let your specialists put it there themselves.
You train your specialists and set up your desks so it’s clear where things belong, so the knowledge is captured and can be pulled back out and reused whenever you need it. Get this right and the investment you put into the setup compounds. Get it wrong and you rebuild the same thing every few weeks.
The brief: what you hand over
Not “some files.” A brief.
A good brief says what problem you’re solving and what context is needed to solve it well. The common mistake is handing over the problem, skipping the context, then judging the work that comes back.
The upgrade worth making: get AI to help you write the brief. Tell it what you’re trying to achieve and ask what it needs from you to do the job properly. It’ll tell you. Better brief in, better work out, and you’ll learn what you’ve been leaving out.
The Filing Cabinet: your AI Brain
You need a proper filing cabinet in your building.
Your AI Brain is a folder of plain text files on your own computer, structured so any AI tool can read it. It has three parts. The archive of originals holds everything as it came in, never altered: your knowledge files, your frameworks, transcripts of you teaching, your best emails, the questions clients actually ask. The reference section is linked, searchable pages built from those originals, with a contents page listing the lot. That’s where it stops being storage and starts being knowledge. And the label on the front of your filing cabinet sets out what belongs in each drawer, what each folder is for, what may be done with it, and what must never happen to it.
One drawer is special. A small set of files, around ten at most, read in full before anything goes out in your voice: your Brand DNA, Brand Voice Guide, Ideal Customer, Offers, Positioning and Personal DNA. The test is “do I want this read every single time?” Go much beyond ten and nothing gets proper attention.
Separate libraries for separate purposes. Your business knowledge lives in one. What you’ve learned from other people lives in another, kept deliberately apart so you never accidentally claim someone else’s thinking as your own. Split by how you think, not by topic, and don’t over-segment.
What stays out: passwords and payment details, live operational data (your CRM and project tools hold that), and other people’s personal information.
And the part that matters most for the building: it’s plain files, on your drive, readable by whatever tool is best next year. Put your filing cabinet on a trolley and roll it into the next building if you decide the one you’re in no longer suits your requirements. (Here’s how to build your own AI Brain.)
Memory: what your AI remembers about you
Memory is separate from the filing cabinet, and often confused with it.
The filing cabinet is about the business, and you curate it. Memory is about you and how you work, and it builds up as you go.
It works at more than one level.
| Level | What it holds |
|---|---|
| The whole building | Applies everywhere you work in this tool. How you like things done, what you’re working on generally. This is your system-wide memory. |
| This desk | Scoped to one project. What’s true for this body of work and nowhere else. |
People often talk about a third level, memory within a chat, and it’s worth being precise here because it causes real frustration. Within a single conversation, nothing is being remembered. That’s the desk surface, and it degrades as it fills, exactly as described above. What feels like forgetting isn’t a memory failure. It’s an overloaded desk.
Also, your AI writes to memory on its own, but you’re not locked out of it. You can tell it to remember something, correct something it’s got wrong, or tell it to forget something entirely, at any point. You can also read what it’s holding, and it’s worth doing every now and then.
The instructions: a handbook spread through the building
This is the one thing structured completely differently to a traditional office.
There’s no single staff handbook sitting in a drawer. The handbook is distributed through the building, showing up where and when it’s needed.
| Where | What it governs |
|---|---|
| Global instructions | Everything, everywhere. How you want work done across the whole building. |
| Project instructions | This desk. Pinned to the wall above it. |
| Skill instructions | This specialist. What they know and how they do their job. |
| What you say in the chat | This job, right now. |
Most specific wins. If a layer is silent on something, it falls through to the next one out.
The practical instruction: put instructions where they’re needed. A rule that only applies to one client has no business in your global settings. A rule that applies to everything you write shouldn’t have to be repeated in every chat. Getting this right is most of the difference between a system that works and one you’re forever correcting.
The org chart: building for workflows, not just jobs
Here’s where a lot of business owners stall. They hire, and hire, and never draw the chart.
Twelve skills, four custom GPTs, a handful of automations, and no idea which one does what, which two overlap, or what happens when one gets it wrong. That’s not a team. That’s a room with people in it.
A subscription isn’t a strategy. What’s usually missing isn’t more technology. It’s management. (And if you’re still rationing that subscription, read why your AI subscription is the cheapest team member you’ve ever hired.)
Named roles with defined jobs. That’s what a skill already is, if you’ve built it properly: a clear outcome, a clear remit, and a name that tells you what it’s for.
Governance you decided on. You set the boundaries when you build the specialist: what they do, what they don’t touch, and what they have to check with you before doing. That’s not something you bolt on later. It’s part of hiring.
Build for workflows, not just for jobs
This is the piece that makes the biggest impact on your workload.
Build your specialists knowing who they hand off to. One specialist’s output becomes the next one’s input. A researcher’s findings feed the writer. The writer’s draft feeds the editor. The editor’s approved copy feeds whoever schedules it.
If you build each skill in isolation, you end up as the courier: copying output out of one chat, reformatting it, pasting it into the next, and filling in the gaps every time. You’ve automated the thinking and kept all the carrying. That’s not the best use of your time and expertise. Connectors are what let the last specialist in the line put the work where it belongs, instead of handing it back to you.
Design the handoffs deliberately. Decide what each specialist produces, in what shape, and who needs it next. Name the next specialist in the line inside the skill itself so the chain actually connects.
A flat list of roles is where the work stops flowing. The value shows up when the work moves between them without you in the middle.
A coordinator, once you’ve got more than a handful
The AI coordinator is the one you actually talk to. They hold the whole picture and route the job to the right specialist.
Otherwise you become the switchboard, and you’ve just given yourself a new job.
If you can’t draw your org chart on one page, you don’t have an AI team. You have tools you’re hoping will help.
What leaves the desk: artifacts and schedules
Artifacts: the work you keep
The chat is your conversation. The artifact is the thing that comes out of it and outlives it.
Documents and reports. Dashboards and trackers. Checklists, run sheets and SOPs. Calculators and quote tools. Proposals and one-pagers. Slide decks. Planning boards. A reference page your team keeps open in a tab. A landing page mockup you want to show someone before it gets built.
Each one has its own home. You can hand it to someone else, pin it on the wall, or pull it back out next quarter and update it rather than starting again.
The conversation is disposable. The artifact is what you keep. If it only ever existed inside a chat, you’ve lost it.
Schedules: the standing appointment
A scheduled task is the recurring entry in the diary. Every Monday at 9, do this job, without being asked. Your specialists clock in and get on with it.
The one real difference: amnesia
Everything above is a fair reflection of how an office works, with one exception worth naming out loud.
Nothing in a real office forgets the moment you leave the room. The default in the AI office, and with AI tools in general, is amnesia. The filing cabinet, the drawer, the memory, the artifact, the saved brief: every one of them exists to get around that.
Once you see that, the setup work stops feeling like admin and starts looking like the actual job.
Start with the filing cabinet
If you wouldn’t expect a new starter to do good work on day one with no brief, no files, no keys and no idea how you like things done, don’t expect it here either.
Set the office up properly and you’ve got a team that knows your business.
Skip it, and every day is day one. A capable stranger, no brief, starting from scratch, every single time you sit down.
So where do you start when fitting out an empty AI office?
With the filing cabinet. Every other part of the office reads from it. Your specialists do better work when they can pull from it, your desks carry it into every job, and it’s the one piece of furniture that comes with you to whichever building you’re in next.
That’s exactly what my Build Your AI Brain training walks you through: setting up your own filing cabinet of plain files on your own computer, getting your business knowledge into the right drawers, within the right folders, and the ten-minute weekly habit that keeps it current. It’s the first piece of furniture to move in, the one everything else depends on, and the one piece of furniture you take with you when you run from the burning building (let’s hope that never happens).

