Work out which parts of your job are still yours
Agents are not a phase. They are going to keep getting better at some things and stay bad at others, and almost nobody can tell you which is which for the work they personally do.
The question everyone asks is the wrong one
“Will AI take my job” has no useful answer. Every tool built around it hands you a percentage and stops, which tells you how to feel and not what to do. Jobs are not atomic anyway. A job is a bundle of thirty or forty distinct activities, and they are not equally exposed: an agent can produce most of a first draft, most of a data reconciliation, most of a research summary. It cannot hold a difficult conversation, carry the consequence of a decision, or be in the room.
Scored at the level of the job title. Returns a single percentage, which is a feeling rather than an instruction. Nothing in it tells you what to do on Monday.
Scored at the level of the activity. Returns a list you can act on, one row at a time, and it names the parts that are not going anywhere.
So the honest unit is the task, not the job title. Break the work into what it is actually made of and the picture stops being frightening and starts being useful. Some of it goes. Most jobs have a real amount that does not, and that part is worth knowing precisely rather than vaguely.
What we built
Every other directory of AI tools is indexed by vendor, which is why they all end up as the same alphabetised categories and the same handful of paid slots. Ours is indexed by work. We took the public occupational database the US Department of Labor maintains and broke it apart.
How much of each activity’s output software can produce today. Estimates, not forecasts.
You can enter from any direction. Your job title, a single task you are sick of doing, or a tool you are evaluating. They all lead to the same place.
What the index actually says
Here is the shape of the whole catalogue, which is not the shape most coverage of this subject would lead you to expect. The largest group by a wide margin is the work software barely touches.
Why the name
“Human in the loop” usually describes a person stationed at the end of an automated process to approve its output. A rubber stamp with a job title. We mean it the other way round.
A person is stationed at the end of it to approve what comes out. The work is the machine’s and the signature is yours.
You decide what to hand over and you keep the parts that were never really about producing output.
That is not a consolation prize. Judgment, accountability, presence and trust are most of what people are actually paid for, and they are the parts that get more valuable as the rest gets cheaper.
What we will not do
No vendor has paid to appear, to rank, or to be recommended. When we name one tool as the thing we would use, it is because it is the cheapest thing that does the job, not the best funded.
1,059 of the 2,083 activities we catalogued are barely automatable, and their pages say so plainly instead of padding a list with tools that would not help.
They describe how much of a task’s output an agent can reliably produce today, assuming a person still approves it. They are not forecasts of job loss and we do not present them as such.
The occupations, activities and task descriptions all come from O*NET, which is public domain, and every page links back to it. The whole index is published as an open dataset. Where a judgment is ours, it is labelled as ours.
Who is behind it
A very small independent project, built and maintained by one person alongside a handful of other things. That is worth knowing because it sets expectations in both directions: there is no support desk, and there is also no growth team, no investor deck, and no reason to tell you a tool is good when it is not.
If something here is wrong, telling us is genuinely the most useful thing you can do. The contact form goes to a real inbox and gets a reply from a person.