Most AI assistants still feel like cruise control. Helpful, not trustworthy. The path to true AI autonomy has a ladder similar to full self-driving. In self-driving, there’s a legal five-level scale, from Level 0 (no automation) to Level 5 (full autonomy with no human involvement). You climb each level by proving reliability without breaking safety or confidence. 

I see personal intelligence in the same way. Each level builds on the last, and you can’t skip the foundation and expect everything else to work. I want to walk through what I consider to be the Four Levels of Personal Intelligence.

Level 1: Understanding Requests

If an AI cannot understand you, nothing else matters. Imagine sitting down for the SAT and having to read the instructions for each section. You need to fully grasp what the examiner is asking before you even think about bubbling in an answer. Level 1 is the same. The system should consistently grasp not just your words, but your intent, slang, and shorthand without you slowing down or rephrasing.

Checkpoint: 97 percent first-try intent on your top 50 verbs, including slang and shorthand.

Success Metric: first-try understanding rate = intents understood on the first pass / total intents.

Today, most assistants hover around 90 percent accuracy on first-try intent for common verbs (except Siri). That final 10 percent matters because it’s where trust starts to form.

Level 2: Taking Actions with Oversight

Understanding is one thing. Doing is another. Level 2 is where the AI system acts on your request using tools, APIs, and frameworks. Currently, frameworks like Apple’s App Intents or Anthropic’s Model Context Protocol (MCP) are the primary way to introduce actions and a certain level of agency to an LLM. A user sets guidelines for what the assistant can do and which applications it can work with, forming what I call a trust contract.

Example Prompt: Book flight from SFO to ORD next Friday, window seat, aisle if needed, under $450, land before 6 pm, add to calendar, share to Family chat.

A great Level 2 assistant shows two or three options with tradeoffs in plain language. A user will approve the final choice when the system reliably respects the trust contract. At this level the assistant does not speculate. It executes within the bounds of the trust contract, explains, and asks.

Why this isn’t solved yet: People do not fully trust the choices an AI presents. If it recommends a hotel or flight, there’s often a nagging thought: “I could probably find something better or cheaper.” Even when the choice is optimal, a perceived lack of transparency keeps you from letting it finalize without a check.

The most effective fix is to introduce exit ramps: prompts that allow a user to make AI actions reversible within a grace window with readable receipts. Buttons like “Let me see,” “Let me change,” and “Let me undo.”

Success Metric: end-to-end task success = tasks completed within constraints without user changes / total tasks attempted. Count the full chain.

Level 3: Prompting A User

At this stage, the assistant starts the interaction. It learns your patterns, makes smart suggestions, and sets you up for next steps. Suggestions must live inside the trust contract and must not sound like an ad. The moment incentives enter, trust exits.

Example: You have edited one wedding photo for 58 minutes. The assistant asks, “Want me to apply these edits to the entire album and export the images?” Three buttons sit under the suggestion: “Yes.” “Preview First.” “No.”

A great Level 3 assistant makes suggestions feel like a genuine favor, not an interruption. The assistant cites where learning came from and how to stop it.

Caution: Too many systems turn suggestions into promotions. A coupon for a McCrispy inside Waze while you are stuck in traffic is icky. One ad-shaped nudge erodes all trust established in Levels 1 and 2.

Achieving Level 3 also inherently raises Level 1’s understanding from 90 percent to 97 percent. The skill gap is learning, over time, which suggestions you welcome and which feel intrusive, and adapting so those prompts become more personal and less generic.

Success Metric: 30-day suggestion acceptance = suggestions accepted after 30 days of learning / suggestions shown. Track “stop suggesting this” decay.

Level 4: Proactive Autonomy

This is the holy grail of personal intelligence. An AI system knows your habits, goals, and upcoming tasks well enough to move before you ask.

Two examples:

Morning: an urgent meeting appears. The assistant reschedules your gym, books the same trainer for tomorrow, updates your calendar, and prepares a slide deck for that meeting with notes drafted in Notion.

Travel day: the AI checks you in to your flight, orders a meal near your gate, queues a Waymo ride at the right time after landing, and unlocks your door when you get home.

A great Level 4 assistant leverages an autonomy budget. You allow an assistant to have a certain amount of agency that is custom to you. (Ex. You can allow up to 20 auto-actions per day. Cap any transactions at $50 each. Require receipts to be texted to you each time.)

When in doubt, the AI falls back to approval if confidence drops below 85 percent or when a ceiling is hit. 

Success Metric: auto-action acceptance = actions not undone within the grace window / auto-actions taken. Also track post-hoc undo rate.

This remains hard even if Levels 1 to 3 hit 95 percent, because anticipation needs deep, live context. Without it, proactivity turns into guesswork. The work is scaling personalization so actions keep feeling right as your life changes.

Achieving true personal intelligence requires a significant scaling of compute. In the long term, all personal intelligence goes local first with cloud assist. Phones and laptops hold your ambient context and learned preferences. The cloud bursts for heavy models, shared knowledge, and multi-device sync. This keeps latency low, privacy intact, and cost predictable. 

This AI arms race feels so intense because companies are fighting tooth and nail to be the first to achieve AGI or personal intelligence. As a result, many take the posturing playbook and tout massive improvements solely for PR. They jump the gun and overpromise features from a higher level.

No one can realistically jump from Level 1 to Level 4 just because they built a fancy new framework. Take Apple: as shipped today, Siri still misses Level 1 in basic moments, so the announced Level 2 and Level 4 features of Apple Intelligence at WWDC 2024 felt unearned. If Level 1 wobbles, everything above it inherits the wobble. Iterate on Level 1 until it is boring. Then let Level 2 be boring. Boring is trust.

The next decade is not a token race. It is a trust and permission race. Win that, and Level 4 is inevitable. Beyond Level 4, the limit is not what the AI can do. It is what you dare to attempt.