Leveling up with AI
Four shifts that change what's possible
When I read accounts of people saying AI is overhyped or complaining it can’t do math, I look at how they’re using AI. The answer almost always explains the disappointment. There are levels to AI capability and disappointment often reflects a lack of awareness of these levels and how to tap into them.
So today I’m sharing how I think about the levels of AI use, feedback welcome.
Level 1: Access
The first level up is simply to pay. Free AI uses weaker models that hallucinate more, have smaller working memories, and reason less carefully. The difference between free and paid isn’t marginal. I recently gave the same file of survey data and the same prompt to a range of models. Free Claude’s best model got less than 30% of the answers correct. The paid models I tested got 100%.* Across the board, free models are worse in ways that matter. No one likes another bill, but if you’re still using a free tier, you’re getting worse output.
Level 2: Foundations
Paying gets you in the door. The next level is learning to direct what you’re paying for. Paid AI has features most people never touch — model selection, memory, project instructions, connectors, etc. — and each one meaningfully changes what AI can do for you. They compound in their impact when working together. As part of our Foundation Series, a recent You & AI participant connected her AI to her Google Calendar, Gmail, and Docs, then asked it to analyze what she’d focused on over the past four weeks against her quarterly priorities, which she’d stored in memory after a previous session. Five minutes later, she had a clear picture of how her time was aligning with her goals, and what she’d accomplished. That’s the difference between AI that gives you a decent first draft and AI that functions as a genuine thought partner.
Level 3: Workspace
The third level is using AI with professional-grade tools designed for real work. In specialized fields, purpose-built AI tools are already delivering outsized value. A friend at a private equity firm recently told me his legal team estimates they save $20,000 per question (!) using Harvey, an AI designed for legal work. For general-purpose knowledge work, Claude is currently the only model with a differentiated offering for work: Claude in Excel, Claude in PowerPoint, Claude Skills, and Claude Cowork offer a full additional level of capability beyond skillfully wielded general-purpose AI. Less chatting with AI, more AI designed to succeed at your common workflows.
Level 4: Agency
The fourth level is building your own tools, automations, and agentic workflows. This still requires some willingness to persist in the presence of code, and not everyone needs to go here. But I’d encourage even those who don’t intend to build to get exposure to what’s possible at this level. I recently created a slash command that processes contracts: Claude Code adds all key dates to my calendar, updates the business development pipeline tracker to reflect deal status and terms, and updates the financial model for the year, placing expected revenue in the appropriate month. Understanding what can be built — even if someone else builds it — changes what you can imagine.
These levels aren’t stages you graduate from. Level 1 still matters even if you’re at Level 4, and that is true every step along the way. Each level changes not just what you can do, but what it occurs to you to do, expanding your practical imagination.
* Except Copilot. YMMV.
