Zhang Pinpin · Hands-On AI Coach
← Back to home

Don't Use a Cannon on a Mosquito: The Five Levels of AI in a Business, and Which One You Actually Need

Long-Form Video · EP0087 August 26, 2026 05:09
What this episode covers

The easiest trap for a company adopting AI is using a cannon to swat a fly—jumping straight to training or fine-tuning a model when the actual business need could be met with something far simpler and cheaper. Andrew Ng once said that 75% of teams doing fine-tuning are working on problems that don't require fine-tuning at all, and might not even need complex prompts. Yet plenty of companies have walked right into that pit and burned real money.

Here's something that seems strange on the surface: more than half of all compute worldwide is now going to AI coding. Not everyone is a programmer, but coding is eating half the compute, and every major tech company is racing down that track. Think about what actually happens when we sit at a computer all day. The machine turns code into text and images we can read on screen; we edit those documents with a mouse and keyboard; and our edits get converted back into code that operates on the files. The step that actually produces the deliverable is the code step. Tools like Claude Code and Codex skip the part where we operate the computer—they take our intent, write it directly as code, and the work is done. Writing a proposal is writing code. Building a spreadsheet is writing code. Making a web page is writing code. AI coding isn't serving programmers—it's using code to produce nearly every kind of deliverable a knowledge worker can make.

The easiest trap for a company adopting AI is using a cannon to swat a fly—jumping straight to training or fine-tuning a model when the actual business need could be met with something far simpler and cheaper. Andrew Ng once said that 75% of teams doing fine-tuning are working on problems that don’t require fine-tuning at all, and might not even need complex prompts. Yet plenty of companies have walked right into that pit and burned real money.

Before we get into the levels, consider something that seems strange on the surface: more than half of all compute worldwide is now going to AI coding. Not everyone is a programmer, but coding is eating half the compute, and every major tech company is racing down that track. Why?

It clicked for me one day. Think about what actually happens when we sit at a computer all day. The machine turns code into text and images we can read on screen. We edit those documents with a mouse and keyboard. Our edits get converted back into code that operates on the files. The step that actually produces the deliverable is the code step. Tools like Claude Code and Codex skip the part where we operate the computer—they take our intent, write it directly as code, and the work is done. Writing a proposal is writing code. Building a spreadsheet is writing code. Making a web page is writing code. AI coding isn’t serving programmers—it’s using code to produce nearly every kind of deliverable a knowledge worker can make.

So how did things get to this point? We can walk through the whole arc from the beginning.

When GPT first appeared, AI was just a large language model—something that could take text in and put text out. What made it remarkable was that it had pre-learned the world’s knowledge and methods. Every time you asked a question, it could quickly sift through all of that, find the most relevant answer, and articulate it back to you in coherent language. At that stage its role was essentially a talking encyclopedia—already impressive, and a huge boost to the speed of processing, filtering, and synthesizing information. A lot of work can still be done with nothing more than a large language model. The later levels don’t make this one obsolete.

The second level was the chatbot. People started learning about prompting. A chatbot takes the base language model and uses prompts to give it a specific scenario and identity, usually with a private knowledge base plugged in. Define it as a customer-service agent and it talks like one. Feed it your product manual and sales scripts and it can answer questions like a seasoned employee. Back in 2023 I ran an experimental project at Tencent—an AI-driven ad-review assistant that let small and mid-size advertisers do their own compliance pre-checks on ad content. To this day, the chatbot format is probably still the most common way AI actually gets used inside companies.

The third level is the workflow. By this stage, model capabilities had taken a meaningful jump—reasoning got stronger, multimodal abilities started emerging. A workflow breaks a complex business process into many parallel or sequential nodes and runs them like an assembly line, with conditional checks along the way. The output is extremely stable, but also relatively rigid. Almost nobody builds complex workflows anymore, yet simple workflows are still used constantly in day-to-day operations because they can churn out stable, high-volume deliverables with extreme efficiency.

The fourth level is Skills. As models kept getting more capable, there was no longer a need to hard-code every step. Instead, you write your workflow SOPs as documents the model can read directly. One document equals one skill; load whichever skill the task requires and call it on demand. But even today, a lot of people treat Skills like gospel—convinced that once their agent is loaded with a hundred expert-written Skills it’s unstoppable. Think about what a Skill actually is, though. It’s a prompt that gets loaded flexibly based on the scenario. That’s it. Useful? Sure. Magic? No.

The fifth level is the Agent. Honestly, there’s no strict consensus definition. Different people picture different things when they say the word. Earlier this year “小龙虾”—the Chinese nickname for the open-source Agent project OpenClaw—was everywhere, and for a while the name became almost synonymous with Agent itself. An Agent is a subject that runs autonomously over long periods, holds long-term memory, and can even develop personality traits. It doesn’t follow a fixed instruction manual. It decides on its own which tools to invoke and which path to take toward a goal. Through extended interaction it learns the user’s values, behavior patterns, and decision-making preferences, gradually becoming a working extension of that person. Agents are extremely flexible—but that flexibility comes at the cost of massive compute consumption, lower efficiency, and a higher error rate.

AI has arrived at this point one step at a time, but that doesn’t mean every company’s operations need the most advanced or most complex technology available. When the quality of the deliverable is the same, cost-effectiveness and efficiency deserve serious weight. If you’re running a business, the high-frequency, labor-intensive, high-value work on your plate should be matched to the right level of AI capability—and by now, you should have a working sense of which level that is.

AI coding isn't serving programmers—it's using code to produce nearly every kind of deliverable a knowledge worker can make.