AI for daily, open-ended knowledge work.
Different from well-scoped, high-frequency, SOP-like tasks, e.g. processing claims via a documented process. Also different from SWE coding work that pushes code to production.
This guide is purely functional. We will not talk about the personal, emotional and psychological journey that accompanies this ladder. Seek your own help.
Your work will be smoother and less stressful, and you will have more time for higher-impact, creative and strategic work.
Very soon, not being familiar with how to use agents for daily knowledge work will be as much a basic prerequisite as being fluent with Word, Excel and PowerPoint.
Level numbers are linear, but much like a job ladder, the effort required to go from one level to the next—as well as the impact and benefits of doing so—is exponential.
Copy-paste corp info in and out of a personal ChatGPT account.
Using ChatGPT as a proxy for any mainstream LLM agent product. Could be Claude, Grok, Copilot, etc.
Nightmare of every CISO.
Copy-paste corp info in and out of a ChatGPT Enterprise account.
Your IT admin whitelisted OpenAI so that your enterprise ChatGPT instance can connect to your M365 or Google Workspace, so ChatGPT can at least see your email and SharePoint or Google Drive files without you manually uploading them.
The hero who fought the red tape to make this happen—give them a medal.
Scheduled tasks. Morning brief!! Skills. Maybe even MCP for your common SaaSes like Figma.
Like a toddler discovering that two Lego blocks can actually snap together!
This level is the attitude shift after daily use of Level 1. Start trying AI for all tasks. Try AI first.
You’re building up a folder full of SKILL.md files. You’re building a mental model of the jagged frontier. You get excited by new model releases. ChatGPT outages are a bummer.
Beginnings of shared structure. You will start seeing shared skills for things like “company voice and style guide.” People start sharing procedural skills like “interview me to draft a PRD.” Mostly led by AI super users evangelizing.
You don’t have to write SQL by hand to get sales numbers by region and time period!! Maybe even make your own dashboard! You can make sense of that 2,000-row × 47-column spreadsheet even though you’re not an Excel jockey.
This is a distinct stage because it leads to internal demand for…
IT and data teams realize that it will actually reduce their incoming requests if they bite the bullet and stand up internal MCP servers for high-use, high-demand data sets—typically operational or commercial data. People will connect this to ChatGPT and do their own bespoke analyses, dashboards, etc.
More internal MCP servers. Realization that they need more granular auth and permissions: only some people should see sales numbers, and very few should have write access.
Articulate, capture, and then serve and manage company-wide processes. Write out tacit knowledge. “Run the skill for it” becomes a common utterance.
Culture shift: for new tasks, right away think “how would I capture this in a skill?” There is a healthy proliferation of skills, the good ones naturally getting blessed into the org- or company-wide skill repo.
A.k.a. articulating and measuring what good looks like. Realization that skills/processes plus evals are core IP.
Happy side effect: token cost reduction, since you find that many tasks can be done with models a few steps behind the (expensive) frontier.
A.k.a. self-improvement. Toyota kaizen. Notice issues, encode back into process/skills and evals/rubrics.
Like adding a unit test when you fix a bug in code so it doesn’t regress.
Automate the virtuous flywheel.
Live on UBI and make art, philosophy and (hopefully) babies.