Reviewing AI Drafts Is the Bottleneck: How I Moved Approval Off My Desk

An AI can draft faster than I can read. That is the whole problem in one sentence, and it took me longer than it should have to see it.

I am a licensed counselor. I run a nonprofit counseling organization, I hold a small caseload, and I am responsible for a handful of other companies. Almost everything I make now starts as a draft from a machine. None of it goes anywhere until I have read it and said yes.

So the machine got quicker and I did not. The queue of things waiting for me to read them grew, and every one of them was waiting on the same person.

The step only I can do

There is a step in this work that cannot be handed anywhere. Somebody has to read the thing and be willing to put their name on it. For anything clinical, that somebody has to be licensed. For anything about my companies, it has to be the person who is accountable for them.

That step is mine. It will stay mine.

What I could change is where I do it and what it costs me to do it.

What it cost me at a desk

The cost was more physical than I expected when I started paying attention.

Reading and approving drafts at a desk meant another hour sitting. More neck pain. More back pain. No calories burned, which matters to me, because managing my weight is a real goal and not a decorative one. The work got done. It got done in a way that was quietly costing me something every time.

It was not that the hour was miserable. It is that it was another hour of sitting still, added to a job that already has a lot of those.

What changed

Now a draft goes to a reading app on my phone. It reads the draft aloud to me at double speed while I walk. When I want to say something I stop, mark the sentence, and type the comment, and the playback waits for me. When I am done, the marks get pulled back into the file the draft came from, each one attached to the sentence that provoked it.

I get to get up and move. Tech neck goes down, back pain goes down, and the movement counts toward something I care about. The work is getting done in a healthier way. That is the trade, and it is not a productivity trick.

There is a second reason underneath it. Long drafts arrive as walls of text and I get text wall overwhelm. Something read aloud goes in where something stared at does not. Moving while listening is a strategy I learned in order to use my ADHD as a strength rather than to manage it as a deficit, and this is the clearest case of it in my working life.

What it is actually made of

Three pieces, and none of them is clever.

• A push that takes a file, strips the working notes out of it, and sends it to the reading app as a document I can listen to.
• A pull that fetches every highlight, the comment attached to it, and any note left on the document as a whole, and folds them back into the file the draft came from.
• Playback at double speed, which pauses whenever I stop to type.

The only part that took any thought is the third bullet in the second one. A comment has to come back attached to the sentence that provoked it, not into a separate pile of notes. A pile of notes is a second document to read, and a second document to read is the problem I was trying to leave.

Where it stands

In daily use, and it is the only way long drafts get reviewed here. There is no complaint left in it. The one thing I want is more highlight colours, and that is a request to the people who make the app rather than a fault in the loop.

What the numbers actually say

I went and counted this week, and the honest picture is not the flattering one.

Since 2026-06-25 I have sent 182 of my own documents into that app. I have finished 84 of them. I have never opened 47 of them at all. Measured by words rather than by documents, I have read about eighteen percent of what I sent myself.

Set that against the other half of my library. The books and long articles I am reading to feed my knowledge base, 124 of them, run at about sixty-two percent by the same measure, and I have finished fifty-six.

So the pattern is not that I read slowly. It is that I finish books and I skim my own drafts. Which is roughly correct, since a draft is read to be judged and a book is read to be learned from, and the parts of a draft that need judging are not evenly spread through it. But forty-seven untouched documents is not skimming. That is a queue I have not gotten to.

There is a reason and it is a choice rather than an accident. Business urgency has me in a drafting regimen rather than a reading one. More goes out than comes back read, on purpose, for now. The writing projects I am actually excited about need the opposite, and when I turn toward them the ratio has to flip.

I had never looked at any of this before this week. I had been counting what went out and had never asked how much came back read. It is a productivity measure now and I intend to keep an eye on it.

What it does not fix

The reading still takes as long as it takes. Double speed helps and it is not magic, and a dense paragraph still needs a second pass. Nothing here makes me a faster reader. It moves the reading to a place where it costs me less to do it, and it makes the marks easier to act on afterward because they come back attached to the sentences that caused them.

It also does not create time. Forty-seven unopened documents is the proof. A loop that makes reviewing cheaper still does not review anything by itself.

The general shape

If you are running AI-drafted work past your own judgment, find the step only you can do, and then ask what performing it costs you. Not whether you enjoy it. What it costs.

Mine cost me another hour in a chair, on top of a job with plenty of those. Moving that hour onto my feet did more for how sustainable this work is than any change I have made to how the drafts get written.

Then count what actually comes back read. I had not, and the number was not what I expected.

• 182 of my own documents sent for review since 2026-06-25
• 84 finished, 47 never opened
• About 18 percent of my own drafts read, measured by words
• About 62 percent of my books and long articles read, measured the same way
• Roughly 556,000 words read across the whole library in that period