Think about the work you'd most like to hand to a bot. It probably isn't hard because of the clicking but because of the small decisions inside it: which lead deserves a reply today, which invoice looks off, which vendor always pads the estimate, which email can wait and which one can't. You make those calls all day without noticing, and they're the reason the work still lands on you.
Getting a bot to move around a web app and fill in fields isn't the impressive part anymore, and the hard part is getting it to make the call you would have made. My view is that most automation falls short right there, and not because the software is dumb but because nobody handed over the judgment.
That gap is where your time goes, because if your automation only works when nothing unexpected happens, you haven't handed anything off. You've built something that makes new messes for you to clean up. You get your time back when you can trust the bot with the odd case, the way you'd trust a good colleague.
So why hasn't everyone done this? The reason is that explaining your judgment is miserable. Nobody wants to write a long document about how they think, and most people can't. Ask someone who's excellent at their job why they flagged something, and the honest answer is often some version of "it just looked wrong." The knowledge is real, but it lives in doing the work rather than describing it.
That's also why the usual fix, a better prompt, doesn't get you far. A prompt is you guessing in advance what the bot will need to know. You'll cover the things you think about and miss the things you've stopped thinking about, which are often the calls that make you good at the job.
Think about how you'd teach a sharp new hire. You wouldn't hand them a manual and walk away. You'd have them watch while you work, and you'd talk as you go. "I'm skipping this one because they never pay on time. I'm flagging this because the number is way off from last month." That narration is the valuable part, because it's the reason, said in the moment and attached to the exact thing it explains.
That's what I built LIT for. You record your screen while you do a task and talk through it, and that recording and its transcript become how your bot learns the task, including the calls that need your judgment. You teach it the way you'd teach a person: by doing the work and saying why.
The fair objection is that one recording is one Tuesday. It won't contain every edge case, and some of your reasoning you won't say out loud. That's true, and a new hire who watched you for an afternoon doesn't know everything either. But the reasons they heard are tied to real moments in real work, which is more than a prompt written from memory gives you. And when a new odd case comes up, you do what you'd do with a person: you show them that one too.
I'm looking for 20 beta testers who have real work they're tired of doing and judgment they'd like to hand off, and I'll cover your tokens. If that's you, request access at LIT and bring the task you think is too nuanced to automate, because that's the one I want to see.