Ask ten CEOs whether their company uses AI and all ten say yes. Ask what that means and the answers split: one pastes contracts into ChatGPT, another runs systems that research, deploy, and schedule sales meetings overnight. Both are "using AI." The phrase hides the distance between them.
The takeaway: AI adoption climbs four levels — chat, files, tools, triggers. Ad-hoc chat answers questions; file access adds memory; business-tool integrations add reach; scheduled and event triggers add initiative. Locate your company on the ladder before spending another dollar on tools.
AI Workshop for CEOs
Locating your company on these four levels — and mapping the climb to the next one — is exactly the mapping the workshop does with your team. Three hours live with a group of 8 CEOs, plus a 1-on-1 session to apply the levels to how your team actually works.
Reserve Your Seat →Level One: Chat
Level one is a research librarian on text message. The AI answers questions and assists on demand: web searches, deep research, examining a file you hand it, even writing and running a small program to parse your data.
The limitation: it's a one-to-one exchange. You make a request, you get a response. It anticipates nothing, and beyond general knowledge, it knows nothing about what you're trying to accomplish. Every scrap of background rides in on your prompt.
Direct ChatGPT or Claude to research cargo vans, and it's on you to say how you'll use them. The answer differs if you're delivering candy a few miles versus hauling heavy machinery a few hundred. Leave the detail out, and the model can't know which van you need.
Level one has telltale habits. Very long-running chats that carry a project — until the thread gets unwieldy and details fall out of the context window, the AI's short-term memory. Files copied into the chat for editing, then dragged back out to the file system or an IDE. Work happens, but nothing accumulates.
Level Two: Files
At level two, your AI tooling gets access to a file system or version-control system — a git repository, OneDrive, Dropbox, or Google Drive. The AI now gathers context from your files and contributes back: creating new files, editing existing ones. Claude Code is the canonical example. It can search an entire codebase, compile it, read the error messages, and add features to your website, mobile app, internal tools, or even write a book.
Because files persist, each prompt builds on the previous ones. The things you learn and the tools you make become part of the project. This is the beginning of a harness — the combination of information and tooling that lets the AI do new things and understand your needs better with every run.
The same shape works outside engineering: a marketing team gives the AI a shared drive of brand guidelines, past campaigns, and drafts — it reads them for context and writes new ones back.
Files also make the work shareable. Pull AI output into a repository and colleagues can collaborate on it; at the most sophisticated end of level two, teams work the same project simultaneously and resolve conflicts through GitHub, GitLab, or Bitbucket.
Level Three: Tools
Level three connects the AI to tools beyond the file system: analytics, website hosting, web search, deployment, email, calendar, scheduling. A website agent with access to Google Search Console and Google Analytics can research an SEO improvement, implement it, and deploy to production — pulling in up-to-date business information, considering it, and acting on it. Level three is also where operators direct AI to build custom software that plugs into those same business systems — monitoring a process, flagging exceptions.
The role changes here. With context at the business level rather than the project level, the AI becomes something closer to a chief of staff — advice, not just implementation, on design, sales, marketing, tactics.
A human still initiates the work. But the delegation gets long: a series of long-running tasks executed without asking the operator for anything along the way. A human usually reviews the changes — but some of the work ships unreviewed. Mature level-three shops supplement human review with automated tests and evals — their equivalent for AI output.
Level Four: Triggers
Level four adds the trigger. Instead of a human requesting work, the AI runs on a schedule or fires on an outside event.
These systems exist today — almost all in the less sensitive corners of the business. AI BDRs qualify leads and schedule sales meetings, and the fit works because nobody expects a BDR's outreach to be 100% successful — a miss costs little. Very few businesses run the sensitive surfaces — social posts, marketing campaigns, ad spend — without a human in the loop. Delegate autonomously where a miss is cheap; keep judgment in the loop where it isn't.
A completely autonomous business is the entrepreneur's dream — and like completely reliable passive income, very few exist.
Using the Levels
Most businesses use AI at several levels at once. If you're looking for a good deal on ink for the office printer, chat is fine. You don't need a new Claude Code project.
Tasks also graduate. A chat you thought was a quick question turns into a real project — ask the AI to summarize the chat into a file, drop that file into a project, and you've migrated from the chat level to the file level. Files graduate the same way: adding a tool connection to an existing project is work your AI can do for you.
The levels measure context, not business maturity — though most businesses will want a level-three setup for their complex projects. Match the level to the task: each step up adds capability and cuts the human busywork.
Return on Investment
One rule governs every level: the AI must cost you less time than the task it replaces. It makes no sense to automate a two-second task with an AI prompt that takes three minutes to run — unless it runs without supervision or review, in which case the three minutes are the machine's, not yours.
Sometimes, moving up a level will create a positive payoff on AI automation. Moving from files (level two) to tools (level three) can mean significant gains for tasks that aren't in the AI wheelhouse. A software tool can perform a complex calculation much more quickly and accurately than an AI, so directing Claude to write you an inventory trend calculator makes inventory predictions both faster and cheaper in tokens.
Other times, like finding a good deal on ink, leveling up is counterproductive. You spend more time on AI infrastructure than you will ever recover from finding toner deals.
Find Your Level
Two observations from advisory work with these levels:
- Companies aren't at one level — functions are. Engineering may live at level two while marketing is still copy-pasting into a chat window at level one. Diagnose per function, not per company.
- The climb is one level at a time. Each level's habits are the raw material for the next: the context you assembled for chat prompts becomes the project files of level two; the project files become the business context of level three; the delegated tasks of level three become the scheduled jobs of level four. A function that skips ahead — level-one habits wired to level-four triggers — automates work nobody has learned to review.
The whole diagnostic fits in four words — chat, files, tools, triggers — asked of each function: which is this team actually wired to?
Most companies discover they're further down this ladder than their AI spend suggests. Use that as a map: the gap between your level and the next one is where the leverage is.
Frequently Asked Questions
What are the four levels of AI adoption?
Chat, files, tools, triggers. Level one is ad-hoc chat — an AI assistant answering questions on demand. Level two gives the AI a file system to read and write, so work accumulates. Level three connects it to business tools — analytics, email, calendar, deployment. Level four adds triggers: the AI runs on a schedule or fires on an outside event without a human starting it. Each level adds a capability the one below lacks — memory, then reach, then initiative.
What level is using ChatGPT or Claude in a chat window?
Level one. Chat answers questions and assists on demand, but it's a one-to-one exchange: the AI anticipates nothing and knows only what rides in on your prompt. The telltale habits are long-running chats that carry a whole project and files copied in and out for editing. Work happens, but nothing accumulates.
Does every company need to reach level four?
No. The levels measure context, not business maturity — match the level to the task. A printer-ink price check belongs in chat; a complex project earns a level-three setup. One rule governs every level: the AI must cost you less time than the task it replaces. Autonomous level-four delegation fits where a miss is cheap — BDR outreach, not ad spend — and sensitive surfaces keep a human in the loop.