What I delegate to AI, and what I still check myself

AI does a large part of the work in my business. I use it to write code, build websites, prepare content and work through technical problems.

That makes my decisions about what to check more important.

A page can load and still be wrong. A clip can export cleanly and still show the wrong speaker. An agent can finish its instructions while the thing I actually needed remains unfinished.

The boundary I care about is between doing the task and proving the result.

I delegate work with a clear finish line

A useful task has an output I can inspect.

For my personal website, that has included building pages, fixing mobile layouts, adding spam protection and connecting a contact form to email delivery. The agent can inspect the code, make the change, run checks and prepare a deployment.

I still need a clear answer to a basic question: what would show that this worked?

For a layout fix, that means opening the page at the size where it broke. For a backup, it means checking the destination. For a contact form, it means submitting a message and finding it in the inbox.

“Implement a contact form” leaves too much room for interpretation. “Save the inquiry, deliver the notification, reject invalid spam checks and keep the receiving email out of the public page” gives the work a finish line.

I check the claim behind the success message

The contact form on this site is a useful example.

There were several different kinds of success during setup. The code passed local tests. The website built. The deployment completed. The form accepted a submission.

Each proved something different.

The result I needed was an inquiry saved in the database and a notification delivered to my inbox. We checked both. The test message had a saved record, a notification reference and a matching email in the inbox.

Then mobile review caught another problem. The spam-check widget made the form too wide on a narrow screen. Working delivery had not proved a working mobile layout.

That is why I ask for evidence that matches the claim. A successful build tells me the build worked. It does not tell me the visitor can use the page.

Content needs a different kind of review

I use AI in podcast content work too. It can help find candidate moments, prepare captions and organize material for review.

The checks change when the output represents a real person.

A quote needs to match the source. A clip needs enough context to preserve the speaker’s meaning. A camera cut needs to show the right person, especially when several people are talking.

An exported file proves that a file was produced. It does not prove that the editorial choice was sound.

My standard for that work includes watching the actual result with sound. A transcript or a few still frames cannot establish every speaker handoff.

I delegate preparation and production. I keep responsibility for whether the finished piece says what it should.

Permission is part of the task

I also separate actions by their consequences.

Reading a file, preparing a draft and testing a local change leave room for review. Publishing a page, changing a domain setting or sending a message affects something outside the working session.

I want those boundaries stated before the agent reaches them.

While building my personal site, one boundary was simple: keep the PGS website separate. Access to both projects was not a reason to change both. Another was cost: setting up the form was not permission to buy a subscription.

The agent needs enough freedom to make progress inside the task. It also needs to know which decisions remain mine.

My job is to make “done” specific

I do not need to watch every command. I need to understand the result well enough to accept it.

The questions I come back to are short:

• What changed?
• What was checked?
• Where can I inspect the result?
• What remains unverified?

Those questions make delegation useful. They let me spend less time supervising the steps and more time judging the work.

I want AI to carry more of the execution. I still own the decision about whether that execution produced something ready to use.