Choose the right kind of AI for a task
Distinguish rules, classifiers, embeddings, and generative models by what they return and how you can check it.
Start with the output
“Use AI for the library” leaves the job open. Start by writing what goes in, what comes out, and how you’ll tell whether it worked. Choose a tool that produces the output you need and fits the inputs you can provide.
| Job | Tool to try | How to check it |
|---|---|---|
| Send a reminder three days before an explicit due date. | A date calculation and a template. | Test calendar boundaries, missing dates, and time zones. |
| Sort messages into return, renewal, or equipment reservation. | A classifier trained on labeled messages, or a language model with a fixed output schema. | Score each category and check ambiguous requests. |
| Find notes about “borrowing longer” when the handbook uses “renewal.” | Semantic retrieval with an embedding model. | Check where the expected passage ranks. |
| Explain a handbook policy in plain language. | A chat model supplied with the relevant passage. | Compare each claim with its source. |
Combine components when a job has multiple steps
A scanned handbook may need optical character recognition to recover text. A search index can find the passages for a question. A model can turn those passages into an explanation. Each step has a different output and can be tested independently.
A supervised classifier learns from examples paired with labels. An embedding model produces a vector that search software can rank against other vectors. A generative model produces new content, such as a sentence or image. You can run each locally if you have compatible weights, software, and hardware.
Try it
- Write one job you want your computer to do, such as “find every note that mentions the invoice number I entered.”
- Write one example input and the output you expect. Add a case where the answer should be missing or where the input is ambiguous.
- Choose a tool from the table. Decide which result you’d measure before trying another tool.
You can name a concrete output for each step and a test that would catch a mistake.
Check your understanding
You need to find a unique reference number exactly as it appears in a file. Which first approach can be tested without a model?
Search for the literal identifier and verify the matches in their source files.
Sources
Progress is saved in this browser.