Work breaks down for two main reasons.
Either there are too many steps, or you are stuck dealing with systems you cannot change.
Until recently, you had to live with that. Now you don’t.
AI can now read files, move through websites, click, type, test its own work, and loop until it gets something right. That changes what is practical. Instead of grinding through bad workflows, you can decide how to handle them.
Step 1: Find the real friction
Look at the process and ask:
Is the pain coming from too many steps?
Or from a system I don’t control?
That answer drives everything.
Step 2: Choose your move
Simplify
If you control the process, simplify it.
Cut steps. Remove handoffs. Collapse work into a loop that runs end to end.
Example:
On my own site, I can give an agent a task like “improve this page.” It pulls the code, makes changes, runs it, opens the page in a browser, checks how it looks, catches errors, fixes them, and repeats. What used to be a back-and-forth cycle of build, test, fix, review becomes one continuous loop that I supervise instead of execute.
The key shift is not speed. It is that entire chunks of work disappear.
Bridge
If you don’t control the system, build a bridge across it.
Let an agent handle the mechanics while you handle the judgment.
Example:
Filing taxes is a perfect case. I can’t change that system, and I’m not going to rebuild it. So instead, I:
- Put all my documents in one folder (returns, K-1s, business info)
- Gave the agent access to control the mouse and keyboard
- Told it: go through the tax software and fill this out as far as you can
- Let it move screen by screen, entering data, navigating menus, and progressing through the flow
- Had it stop and report back where it got stuck or needed input
Instead of me clicking through dozens of screens and copying values, the agent does the traversal. I step in only where actual decisions or missing info are required.
The system stays exactly as it is. The effort changes completely.
Replace in stages
If the system needs to go but can’t go all at once, replace it piece by piece.
Example:
Say you have a messy internal tool. You don’t shut it off overnight. You start by bridging into it so the manual work drops. Then you begin building cleaner paths for parts of the workflow. Over time, more of the work happens in the new flow and less in the old one.
You don’t wait for a full rebuild. You move gradually, without breaking things.
Leave it alone
If something is rare, unpredictable, or high risk, leave it manual.
Forcing automation where it doesn’t belong usually creates more problems than it solves.
The shortcut
Do I control this process?
Is the pain inside the steps or at the system boundary?
Can I replace this now or only over time?
The point
Most workflows were shaped by human limits.
Those limits just changed.
So the job is no longer to slowly improve broken processes.
The job is to remove them, route around them, or replace them.