The six levels of working with AI - and why I operate at 3 to 5
Most companies chat with AI and call it adoption. I use a six-level ladder to show what real AI work looks like. Each level roughly doubles output, and knowing your level tells you what to build next.
Ask a room of marketers who uses AI every day and every hand goes up. Ask who has AI doing real work while they sleep and most hands drop. Both groups say they have adopted AI. The word covers a very wide gap.
This is how I see working with AI: a ladder with six levels. Each level has a rough output multiplier, meaning how much more one person gets done compared to working without AI. Below I walk through each level and where I stand on it.
Talking to AI
Chat - "Google Search 2026"
Prompt engineering - you give it a role
Where you should operate
Persistent context - AI knows your company
AI with tools - real actions, your route
Agent with a goal - picks its own route
Frontier labs only, for now
Full autonomy - plug it in and it works
Bar length = output multiplier. Directional, not measured to the decimal.
The numbers are a guide, not a lab result. The shape is what matters: each level roughly doubles the one before it, and each step up is harder than the last.
Level 1: chat (1x)
This is where most companies are. A chat tab is open, questions go in, answers get copied out. It is useful in the same way a search engine is useful, which is why I call it Google Search 2026.
It gives you no edge. Your competitors have the same tab open.
Level 2: prompt engineering (2x)
At level 2 you give the model a role, rules and examples. “You are a senior media buyer. Review this campaign. Flag anything that wastes budget.” The output gets clearly better. Most AI training for teams ends here.
The limit is that every conversation starts from zero. You are the context, and you type it in again every time.
Level 3: persistent context (4x)
This is the first level that compounds. Brand guidelines, clients, tone of voice, product facts: written down once, stored where the AI reads them, and loaded before you type a word.
At level 2 you brief a new intern every morning. At level 3 the intern remembers. A short prompt now gives long and correct output, because the hard work sits in the context. My rule of thumb from daily use: when the context is rich, the prompt can be one sentence.
Level 3 needs no engineering. It needs writing. That is also why many teams get stuck here. To write down how your company sounds and what it sells, you first have to agree on it.
Level 4: AI with tools (7x)
Level 4 gives the model tools. It takes real actions in real systems, on a route you designed. Scrape this ad library. Search LinkedIn for these roles. Read this sheet. Email the report. The AI does the steps, you decided the steps.
Here the output stops being text and becomes finished work. One example from my own work is a competitor teardown: one Facebook Ad Library URL in, 86 Żabka ads scraped, grouped into themes, read visually and mapped to the funnel. It took a couple of minutes and cost about 10 cents. An analyst would need a day or two.
For most marketing teams, level 4 has the best return for the effort. The tools are ready - connectors and MCP servers exist for most data sources a marketer uses - and you keep full control of the route.
Level 5: agent with a goal (10x)
At level 5 you stop designing the route and give a goal. “Find companies in this sector that fit our profile, qualify them, and add the good ones to the sheet.” The agent chooses which tools to use, in what order, and when it is done.
I run agents at this level in production. The honest part: level 5 costs you in reliability work. An agent that picks its own route will sometimes pick a wrong one, with full confidence. So every agent I build has hard limits, a human approval step before anything leaves the building, and a log of what it did. None of this shows up in a demo. In production it is most of the work.
Level 6: full autonomy (20x)
Plug it in and it works, with no human in the loop. Today this level belongs to OpenAI, Anthropic and Google. It does not belong to me, to your agency, or to the vendor whose deck says “fully autonomous”.
Anyone selling level 6 today is either marketing or taking a risk with your business. For a company that is not a frontier lab, the real ceiling is a well-run level 5 with human checks in the right places. That is plenty. The gap between level 5 and the market’s level 1 is already about 10x in output per person.
How to use the ladder
I work at levels 3 to 5. Every agent on my agents page sits there: it knows the brand, uses real tools, works toward a goal where that is safe, and waits for a human where it matters.
If you want help finding your level, or climbing the next one, get in touch.