What Agents Can Automate
Every entry here is real work that agents do today, not a demo. Each one lists the tools it needs and how much rope it should get, because the difference between a useful agent and an incident is almost always the autonomy level rather than the model.
Read the autonomy column first
Assist
The agent prepares, a person decides and does. Start every new use case here, whatever the sales deck says.
Approve
The agent does the whole job but stops before the irreversible bit. One click from a human, every time.
Auto
It runs unattended. Earn this with a logged track record and a way to undo, never on day one.
The catalog
How to pick your first one
- Pick something done many times a week, not the hardest thing you do once a quarter.
- Make sure a correct result is obvious to a person in under a minute. If nobody can tell right from wrong quickly, you cannot supervise it.
- Check the tools exist and have an API. Most agent projects die on access, not intelligence.
- Start at assist. Move to approve when the drafts stop needing edits. Move to auto only with a rollback.
- Write down what it costs today in hours. Without that number you cannot tell whether it worked.
The uncomfortable pattern. The entries with the highest risk are also the ones with the clearest payback: refunds, access requests, invoice matching, screening. That is not a coincidence, it is why they are still done by hand. Automate the preparation for those and keep the decision with a person, and you get most of the time back without owning the failure. Ready to build one? Assemble your own agent and run it here.
Free from AI School - no signup, everything runs in your browser.