1 / 40
← → navigate · N notes · F fullscreen
AI, explained for the people who run the business.
How it works, how to run it, and how it helps a small team operate like a big one.
learn.growwithkoda.co
Run a business that operates bigger than it is.
Every growing business has more work than people. AI doesn't replace your team. It fills the seats you haven't been able to hire for.
Same team. More seats filled. Work that gets done whether or not someone is at a desk.
You
Sales
Operations
Office
Follow-up
AI
Research
AI
Reporting
AI
First drafts
AI
PeopleAI, set up and directed by your people
That's the goal. Everything in this deck is how you get there.
What we'll walk through.
The brain
What a model is and how it reads
The body
Chat, Cowork and Code
Where it runs
Your computer or the cloud
Agents
Answering vs doing
Connections
Plugging into your tools
Context
The part you control
Your setup
From one person to a team
Each part builds on the last. By the end, you'll know what to set up first.
How we got here
Seventy years of AI, in one line.
1956
The field gets a name
Researchers at Dartmouth coin "artificial intelligence" and set out to make machines think.
1980s
Rules written by hand
"Expert systems" capture a specialist's know-how as thousands of if-then rules. Useful, but brittle.
1997
A machine beats a champion
IBM's Deep Blue defeats world chess champion Garry Kasparov.
2010s
Machines learn from data
Machine learning goes to work: spotting defects on factory lines, predicting repairs, flagging fraud.
2017
The transformer
Google researchers publish the design behind every modern language model.
2022
ChatGPT
AI you can simply talk to. It passes a million users in five days.
2025
AI works in your files
Claude Code connects AI to the folders on your computer. That's where this deck picks up.
Rules, then learning, then language. What's new is that anyone can put it to work.
The brain
A model is a brain that learned by reading.
You type
"Draft a friendly check-in to a client who went quiet after our quote."
The model
Predicts the most useful next words, one small piece at a time, based on everything it has read.
You get
"Hi Marisol, I wanted to check in on the quote we sent over. Happy to walk through any of it..."
Think of a brilliant new hire who has read almost everything ever written, but knows nothing about your business yet.
It knows the world. It doesn't know you. That's the gap you fill.
Company, product, model.
Company
The app you open
The brains inside
Anthropic
Claude
Fable 5.1Opus 5.5Sonnet 5Haiku 4.5
OpenAI
ChatGPT
GPT-6 AstraGPT-6 SolGPT-6 Luna
Google
Gemini
Gemini 3.1 ProGemini 3.8 FlashFlash-Lite
xAI
Grok
Grok 4.7Grok 4.3
Meta
Muse
Muse Spark 1.3
"ChatGPT vs Claude" compares products. The model inside does the thinking.
Not every model talks.
Jev, released this month by TypeSafe AI, never writes a sentence. You give it the allowed answers up front, and it returns one of them, plus how confident it is.
Answers built for software, not people. That's what automations run on.
A new email arrives
"Hi, we're opening a second location in March and need pricing for the build-out..."
Is this a sales lead?
Answer
lead: yes
urgency: high
confidence: 0.94
urgency: high
confidence: 0.94
No paragraph to read. Your system routes it to sales on its own.
Chat is one way to use AI. Automations use it to make small decisions all day long.
AI reads in tokens, not words.
A token is a small chunk of text, often a whole word, sometimes part of one. Everything is counted in tokens: what you send, what it reads and what it writes.
- Pricing is per token.
- Memory limits are in tokens.
Roughly how a model splits a sentence
Please schedule the site walkthrough for Thursday.
7 words, 9 tokens
1,000
tokens
≈
750
words
≈
1.5
pages
Clear and relevant beats long. Every extra word costs time, money and attention.
Every model has a working memory limit.
The context window is everything the model can see at one time: your instructions, your files, the conversation so far and its own answer.
When it fills up, older details get summarized or dropped.
Claude's biggest models hold up to about 1,500 pages at once. A focused window still gets better answers.
Context window
Instructions
Your files
The conversation so far
Its answer
Room left
It only knows what's in the window right now. What you put there matters.
Every platform has a big, medium and small brain.
Bigger models think harder and cost more. Smaller ones are faster and cheaper.
On a monthly plan you don't see a bill, but bigger models use up your limit faster.
Claude: cost to write 1 million tokens
Pay-per-use prices, September 2026
Fable, Opus, Sonnet: 1M-token windowHaiku: 200K-token window
Same pattern elsewhere. OpenAI: GPT-6 Astra, Sol, Luna. Google: Gemini Pro, Flash, Flash-Lite.
The biggest model isn't always the best choice. Match the brain to the job.
Pick the brain for the job.
The senior partner
Biggest: Fable, Opus
Strategy, tricky judgment calls, a proposal that has to land, long complex projects.
The manager
Middle: Sonnet
Most everyday work: drafts, summaries, research, cleaning up a spreadsheet.
The fast assistant
Smallest: Haiku
Sorting, tagging and quick answers, thousands of times a day, for pennies.
Start in the middle. Move up when the work is hard, down when it's repetitive.
The harness
The model is the brain. The harness is the body.
The harness
Files
Reads and writes your documents
Tools
Email, calendar, your CRM
Browser
Looks things up on the web
The model
thinks and decides
Memory
Remembers across sessions
Permissions
What it may do without asking
A computer
Runs programs and tasks
Most differences between AI products aren't the brain. They're the body.
How AI got a body.
2022
Chat
ChatGPT arrives. You copy and paste into a chat box, and it only knows what you pasted.
2024
Projects
You upload files to a workspace so every chat can refer back to them.
2025
Claude Code
Coders were tired of pasting code into chats, so the AI was connected straight to the folders on their computer.
2026
Cowork
The same idea in a desktop app, built for everyone. No code, no terminal.
Each step gave AI more of your context, with less copying and pasting.
Claude Code isn't only for code.
A software project is just a folder of files. So is your business.
- It reads, writes and organizes files in the folders you point it at.
- People now use it for proposals, reports and operations, not just software.
- Cowork packages the same idea for teams who never want to see code.
A software project
my-app
checkout.js
login.js
README.md
tests
=
Your business
my-company
clients.csv
pricing.md
proposals
how-we-work
Forget the name. It's an AI that works inside your folders.
Same brain, different bodies.
Claude
models
Chat
Talk it through, draft, ask questions. Browser, desktop or phone.
Cowork
Works on your files and apps from the desktop app, and handles multi-step tasks.
Code
Builds tools and automations. The most power and the most control.
OpenAI
GPT models
ChatGPT
The chat window, in the browser, desktop or phone.
ChatGPT Work
An agent that works across your apps and files and hands back finished work.
Codex
OpenAI's agent for building, in an app, a terminal or the cloud.
The brain is the same in all three. Pick the body that fits the job.
Where it runs
On your computer or in the cloud.
On your computer
- Sees your files and apps.
- Uses the logins you already have.
- Nothing extra to host or pay for.
- Stops when your laptop sleeps.
In the cloud
- Works around the clock, laptop closed.
- Runs on a schedule or when something happens.
- Costs money to host and run.
- Walled off: sees only what you connect to it.
Local is your assistant at your desk. Cloud is the employee who works the night shift.
Every setup trades control for convenience.
They run it for youYou run it yourself
Meta
Muse
A personal agent on its own computer in Meta's cloud. Connect your apps and delegate.
xAI
Grok Bot
A managed team of always-on bots, each with its own cloud computer.
Anthropic
Claude Cowork and Code
Run on your computer, with your files, under your rules.
Open source
OpenClaw, Hermes
You host them on your own computer or a server and plug in any model.
Pay for a service: less to customize, less to maintain, less responsibility.
Run it yourself: full control, and full responsibility for the machine, the security and the upkeep.
There's no right answer, only a tradeoff. We start in the middle: your computer, your files, your rules.
Agents
A chat answers. An agent acts.
Chat
You ask
It answers
You do the work
Copy, paste, send, update, file
Agent
A goal
Plan
what to do next
Act
uses its tools
Adjust
fix and retry
Check
did it work?
Done
It reports back with the finished work
With chat, you're the hands. With an agent, you're the manager.
What makes it an agent.
A model
The brain that decides what to do next.
+
Tools
Its hands: email, files, your CRM, a browser.
+
Context
What it knows about you and this job, plus memory.
+
Permission
Allowed to act, on its own loop or on a schedule.
=
An agent
Take away any one of these and you're back to a chatbot.
Connections
AI plugs into the tools you already use.
An agent's tools are your other software. A connection lets it read from and act in those apps, like sending the email or updating the deal.
There are two main ways to connect: an API and MCP.
Email
read, draft, send
Calendar
find time, book it
CRM
look up, update deals
Drive
find and file documents
Your AI
Team chat
post updates, answer
Accounting
invoices, payments
Databases
your records and lists
Websites
forms, bookings, orders
No connection, no hands. Connections turn advice into action.
An API is the software's service window.
Most software has an API: an official window where other programs can ask it for data or tell it to do something. Nobody walks into the kitchen. They order at the window.
- An API key is a password for software. It proves the request really comes from you.
- Guard it like a bank PIN. Never paste it into a chat or an email.
Your automation or agent
asks: "list the deals closing this week"
API key
attached to the request, proves who's asking
The API
the app's official service window
Your CRM
hands back the deals, or makes the change
APIs are how automations talk to your software when nobody is at the keyboard.
MCP is a universal plug for AI.
MCP, short for Model Context Protocol, is a standard way to plug a tool into an AI app. Think of it as USB-C for AI.
- In Claude, these plugs are called connectors.
- You click connect, sign in with your normal account and choose what it can access.
- Anthropic started it in 2024. The major AI apps now support it.
Claude Chat
Cowork
Claude Code
Other AI apps
MCP
one standard plug
Gmail
Google Drive
Your CRM
Slack
Built once, a plug works in any AI app that supports it.
What you'll do to connect.
A connector (MCP)
Best for everyday work in Chat and Cowork
- Pick the app from Claude's connector list.
- Sign in with your normal account for that app.
- Choose the access: what it can read, and what it can change.
- Check with your admin on a team plan. They control what's allowed.
An API key
Best for automations, cloud agents and custom builds
- Create the key in the app's settings, usually under API or Developer.
- Store it safely: a password manager or a secure settings file, never a chat.
- Give it the least access that does the job. Read-only if that's enough.
- Replace it if it leaks or someone leaves. Delete the old one.
Treat every connection like a login you're handing to a new employee.
At the controls
The controls you'll actually touch.
Model picker
Which brain: big, medium or small.
Effort
How hard it thinks before answering, from low to max.
Permission mode
How much it can do before it has to ask you.
Project or folder
Which set of files and instructions it works from.
Four settings. Get these right and most of the results follow.
Permission modes: how much leash.
Asks about everythingAsks about nothing
Plan
Reads and proposes. Changes nothing until you approve.
Manual
Asks before each edit or action. The safe default.
Edit automatically
Edits files on its own, asks before anything bigger.
Auto
Keeps going, with a safety check on each action.
Bypass
No checks at all. Only inside a sealed-off test machine.
Claude Cowork
Asks before acting by default. Your admin can allow an automatically-approve setting.
OpenAI CodexRead-only, Auto (the default: works inside the project folder and asks before going outside it) and Full access.
Start with more questions, not fewer. Loosen the leash as it earns your trust.
Chat, Cowork or Code: which one to open.
Think something through, draft an email, ask a question
Chat
OpenAI: ChatGPT
Work across your files: organize, summarize, fill templates, prep a report
Cowork
OpenAI: ChatGPT Work
Build an automation, a script or a tool that runs on its own
Code
OpenAI: Codex
Start in Chat. Move to Cowork when your files are involved. Reach for Code to build.
Context
AI is only as good as what you tell it.
New hire, day one, no binder
- Guesses at your prices.
- Writes in a voice that isn't yours.
- Doesn't know your customers or your process.
- Asks you everything, or worse, asks nothing.
New hire, day one, with the binder
- Knows what you sell and what it costs.
- Has examples of what good looks like.
- Knows who the customers are and how you quote.
- Knows when to check with you.
Your job isn't to prompt harder. It's to hand it the binder.
How much does a new hire need to know?
Before any task, ask yourself: what would I hand a sharp new employee so they could do this one job well?
Give it that. Not less, and not the whole filing cabinet.
Too little
It guesses, and you spend your time fixing the guesses.
Too much clutter
It gets lost, misses what matters and costs more.
Organized and relevant
It does the job right the first time.
Right-size the context to the task.
Your files are the training manual.
The same folders that would train a new employee train your AI. Organize them once and every person and every agent reads from them.
- Company: who you are and how you sound.
- Processes: how the work gets done.
- Templates: what good looks like.
- Data: the facts it needs.
A context library
company
about-us.md
voice-and-tone.md
processes
how-we-quote.md
new-client-onboarding.md
templates
proposal-template.md
data
clients.csv
price-list.csv
Much of "using AI well" is really organizing your business.
Pick a folder, pick the job.
In Cowork or Code, choosing a folder is handing the AI one project's files.
- It works inside that folder, not across your whole computer.
- It reads the folder's standing instructions first: a Cowork project's instructions, or a "read me first" file called CLAUDE.md in Code.
- In Chat, a Project does the same thing.
Marketing
Client: Harbor DentalAI works here
Client: Northside Build
Operations
Finance
Like handing an employee one client's file, not the whole cabinet.
Templatize the business.
Before
Every proposal written from scratch.
Different every time. The good version lives in someone's head.
After
# Proposal for {client}
## Scope {scope}
## Price {price}
## Timeline {timeline}
AI fills it from your notes in minutes, your way, every time.
If you do it twice, template it. Templates are how AI learns to do it your way.
AI reads some files better than others.
Native
Markdown, CSV, plain text
Clean text with nothing to decode. It reads these perfectly.
Works
Word, text PDFs, Excel
Readable, but some formatting and structure gets lost.
Hard
Scans, photos, tables in images
It has to read a picture of text, so it guesses more.
When you create something new for AI to use, make it Markdown or CSV.
Markdown is just text with structure.
What you type
# How we quote ## Before you quote - Confirm the site address - Ask for drawings or photos **Rule:** never quote without a site visit.
What it looks like
How we quote
Before you quote
- Confirm the site address
- Ask for drawings or photos
Rule: never quote without a site visit.
People can read it, AI reads it perfectly, and any app can open it.
The setup
Start here: a second confidant on your computer.
Before any automation, get one person working well with AI on their own computer.
- Ask it questions like you would a trusted advisor.
- Let it draft, organize and double-check your work.
- Build the context library as you go.
You
set direction
Your AI
Cowork or Code, on your computer
Your library
your business, in folders
The first win is personal: a second brain that knows your business.
Then: give the whole team the same setup.
Sales
+ their AI
Operations
+ their AI
One shared context library
Same playbook, same templates, same facts
Office
+ their AI
You
+ your AI
Everyone gets a smart assistant, and every assistant works your way.
Then: agents for the roles you haven't hired.
You and your team
Sales
Ops
Office
Inbox
agent
Follow-up
agent
Research
agent
Reports
agent
Run in the cloud or on a server you own, with tools like Hermes, OpenClaw or Grok Bot.
This is how KODA runs
Atlas watches our client software around the clock and fixes breaks with a one-tap approval from a phone.
Willow tracks every follow-up from our calls and sends a weekday briefing.
Agents fill the seats. The context library tells them how you work.
The stack, built bottom up.
Context library
Your business, organized into files
Local AI for your people
Cowork or Code on every computer
Cloud agents for the open roles
Working 24/7, reporting to people
A business that operates bigger than it is
Every layer depends on the one below it. Start with the library.
Working with AI
Give goals, not directions.
Directions
"Open the customer spreadsheet. Sort by last order date. Find everyone older than 90 days. Write each one an email."
It follows the steps, and stops there.
The goal, and why
"I want to win back customers who've gone quiet. Repeat work is our best business and I'd like them to hear from us before a competitor does. Find who's gone quiet and draft a warm check-in for each, in our voice."
It makes the calls you would, because it knows what matters.
Explain the what and the why. AI is good at the how.
Use it as your teacher.
"I want to set up ______, but I don't know how. Walk me through it, one step at a time."
That's a complete prompt. Anything it can do, it can teach you to do.
Ask it how
Stuck? Ask it to explain, then ask again simpler.
Give it the binder
Context like you'd give a new hire.
Review before it acts
Keep a person in the loop on anything that matters.
Save what works
A good result becomes a template.
You don't need to know how. You need to know what you want.
Organize the context. Direct the work. Operate bigger than you are.
Bobby Helf · bobby@growwithkoda.co
growwithkoda.co