FOUNDATIONS
When to Use AI (and When Not To)
Use AI for fuzzy tasks, where many different answers would be fine, like summarizing, classifying, or drafting. Use plain code for exact tasks, where there is one right answer and almost right is wrong, like validating, calculating, or querying.
What the video explains
Software is built in layers
AI can do a lot, but it is not the right tool for every job. To choose, the video looks at where AI sits inside a piece of software. It starts with programming languages like Python, which are how people tell a computer what to do. They are not the bottom. Under them is machine code, which is just ones and zeros, and under that is hardware, where electrons move.
At first there was only machine code. Then programming languages arrived, new and not yet trusted, and over time they became solid. Each layer is added on top of the one below it to go further. AI is the newest layer, where programming languages once were: young and still proving itself. It sits on top of the others and does not replace them.
AI is a thin layer
Looking at how thick each layer is, the AI layer is thin. In a good application, AI is only about ten percent of the logic, and everything beneath it does the rest. The rule given is that past about ten percent AI, a solid app tends to become an unstable one. So AI is a thin layer on top, not the foundation.
Fuzzy tasks and exact tasks
Every job in an app falls into one of two kinds. Fuzzy tasks have no single right answer, and many answers would be fine. Exact tasks have one right answer, and it must be correct every time.
For fuzzy tasks, AI is the right layer. A customer sends a long email: you could summarize it, classify it as a complaint or a question, or draft a reply. There is no single correct wording, and many versions work. For exact tasks, use programming languages. A signup form needs an email address validated, an order needs its price calculated, a database needs queried for one customer. Each has one right answer, and almost right is still wrong.
When AI writes the code
AI can also write code. A fuzzy request, like "write me a script," comes out as exact code. The video marks that as coded by AI, and it sits right on top of exact tasks. Fuzziness generates exactness, but the result is still an exact task.
The whole picture: AI is a thin layer, about ten percent, on top of solid code. Use AI for fuzzy tasks like summarize, classify, and draft. Use code for exact tasks like validate, calculate, and query. Ten percent AI, ninety percent everything else.
Where AI sits inside software
Software is built in layers. Electrons move through hardware, machine code runs on the hardware, and programming languages like Python sit on top of machine code. Each layer was added on top of the one below it, and none of them was replaced when the next one arrived.
AI is the newest layer. It sits on top of programming languages and does not replace them. Programming languages were once new and people did not trust them yet; AI is at that stage now, still proving itself.
Sort every task: fuzzy or exact
Every job in an application falls into one of two kinds. A fuzzy task has no single right answer, and many answers would be fine. An exact task has one right answer, and it must be correct every time.
The test is what happens when the answer is slightly off. If a slightly different answer is still acceptable, the task is fuzzy. If almost right is still wrong, the task is exact.
Use AI for fuzzy tasks
Take a customer who sends a long email. You can summarize it, classify it as a complaint or a question, or draft a reply. There is no single correct wording for any of those, and many versions work, so this is a good job for a model.
Summarizing, classifying, and drafting are the typical fuzzy tasks. A language model handles them because it produces a plausible, useful answer from messy input, and that is all these tasks ask for.
Use code for exact tasks
Take a signup form that asks for an email address. Validating that address, calculating the price of an order, and querying a database for one customer each have one right answer. A result that is almost right is still wrong, so these belong to programming languages.
Code gives the same output for the same input, which is what an exact task needs. A prompt can make a model's output more consistent but does not make it guaranteed. Anthropic's guide on output consistency points to structured outputs when you need guaranteed schema conformance, which is a case of using a tool built for an exact requirement instead of relying on prompt wording.
When AI writes the code
There is one more path. AI can write code: a fuzzy request like "write me a script" comes out as exact code. The request was fuzzy, but the result is an exact task, because the script has to run correctly every time.
Treat that output as code, not as AI. It sits on top of the exact tasks and gets reviewed and tested like any other code. Fuzziness generated the exactness, and the exactness still has to hold.
How much of an app should be AI
The rule of thumb in the GenLucid video is that AI is about ten percent of the logic in a good application, with everything beneath it doing the rest, and that past roughly that point a solid app tends to become an unstable one. This is a heuristic from the video, not a measured threshold, so treat it as a way to keep AI a thin layer, not as a number to hit.
The same idea appears in Anthropic's guidance on building with LLMs: find the simplest solution possible, and only increase complexity when needed. Agentic systems often trade latency and cost for better task performance, and the tradeoff is worth checking each time, which can mean not building them at all.
FAQ
- When should you not use AI?
- When the task has one right answer that must be correct every time: validating input, calculating a price, querying a database for a specific record. Code does these exactly, and a model that is almost right is still wrong.
- Is classification a job for AI or for code?
- It depends on whether the categories can be written as exact rules. Sorting emails into complaint or question has no clean rule and many borderline cases, so it is a fuzzy task. Checking whether a number is above a threshold is exact and belongs in code. A decision model is one way to handle the fuzzy kind without a full LLM.
- What is the ten percent rule for AI?
- It is a rule of thumb from the GenLucid video: AI is about ten percent of the logic in a good application, and past roughly that point a solid app tends to become unstable. It is a heuristic for keeping AI a thin layer, not a measured threshold.
- If AI writes the code, is that an AI task or a code task?
- Both, in sequence. The request is fuzzy and the model handles it, but the output is code, which is an exact task. It still has to run correctly every time, so it gets reviewed and tested like any other code.
- Does using AI for fuzzy tasks mean the output needs no checking?
- No. A fuzzy task accepts many answers, but not every answer. A summary can drop the one fact that mattered, or a draft can miss the tone. The point of the split is that these errors are tolerable and reviewable in a way that a wrong calculation is not.