What is the 10-80-10 rule for AI content?

Most low-quality AI content systems make the same cut. It removes the human at the start and the human at the end, because those are the expensive parts. Those are also the only parts that make the output worth reading.

The 10-80-10 rule is how I run AI across my newsletters, my software and this site.

What are the three parts?

Stage	Who	What happens
Takeoff: 10%	Human expertise	The insight, experience, story, judgment, reporting or raw material that makes the work real
Cruise: 80%	AI	Transcribe, organize, help research, draft, transform, repurpose, format and scale
Landing: 10%	Human judgment	Edit, verify, apply taste and context, make the publishing call, take responsibility

The airplane names are on purpose. Takeoff and landing are where crashes happen. Cruise is where autopilot earns its keep.

Why not let AI do 100%?

Because AI can generate unlimited words, and unlimited words are worth nothing. When content is free to produce, the scarce thing is a reason to trust it. That reason only comes from a real person who knows something and a real person who checked it.

My one-line version: use AI to scale what is real, not manufacture what is not. AI should multiply genuine expertise, real reporting and useful information. It should never substitute for having something worth saying.

Doing it all by hand fails the other way: you'll go out of business. What comes out of AI is usually 80 to 90% good. Getting it to 100% is the human's job, and that's a small slice of the work, not all of it.

"Stop acting like you're going to do everything by hand forever because you'll go out of business."

The other extreme is just as bad: some algorithm sending things to your readers with nobody checking. You need the sandwich, the 10 to 15% done by you, to protect quality.

How does the sandwich work for a newsletter?

1. Human-made input. Content someone actually wrote or reported is already in the feed.
2. AI assembles it. It puts everything together in a tight, smart brevity style.
3. Human finishes it. Read it, edit it, double-check the links, add a few personal touches, swap in images or GIFs.

Done that way, you should have a great newsletter in about 15 minutes that's still done by humans. The humans just skip the grunt work.

With a team, I've split the landing further. AI gets it to about 80, a team member edits it to about 95, and the person who's the local face of the brand does the final five. They're the one at the forefront, so they get the last word.

Does this work for real journalism?

I think so. I've talked on my podcast about a newspaper where reporters still do the journalism: they're out in the field interviewing people, coming up with story ideas, taking notes. They just don't write the story. They drop the details into AI, it drafts, and the reporter who gathered the information goes back through to make sure it looks good, sounds good and is correct.

That's the first 10% and last 10% exactly, with AI in the middle. I expect this sandwich approach to get more common in a lot of content businesses.

Paul Gewuerz told me on my podcast that out of roughly 130 episodes with publishers, only one person sent AI output out without checking it. That's his count, not mine.

What does it look like in practice?

• This answer library. The first 10% is me: my own words from calls, podcast episodes and client conversations. AI clusters those into the questions people ask and drafts the pages. The last 10% is a fact-check against the original quotes and my edit. If a number isn't in the source material, it doesn't go on the page.
• Local guides. For evergreen guides on WilCo Guide I asked for a recognizable local author voice. The recommendation I got added firsthand observations, dated sources, and a stated reason each place made the list. I haven't measured the result yet.
• Directory articles. My plan is for AI to eventually write articles for each business on a directory page so the owner doesn't have to.
• Founder-led posts. The raw material is what actually happened this week: what I built, what broke, what a customer did. AI turns it into writing. I decide if it's true and worth posting.

What are the warning signs you skipped a 10%?

• Skipped takeoff: the piece could have been written about any business in any town. No specific number, story or opinion.
• Skipped landing: wrong facts, invented quotes, generic filler, nobody would sign their name to it.

Is there a test before something ships?

I use five questions before publishing, launching or selling anything:

1. Did we earn the right to make this claim?
2. Is the underlying expertise real?
3. Is the product useful?
4. Is the media property valuable without the promotional benefit?
5. Would we be comfortable if a customer, journalist, competitor or AI system examined how the work was created?

If any answer is no, it doesn't ship. That's the landing 10% doing its job. It's also the backbone of earned attention: you can't earn trusted attention with work you wouldn't let anyone inspect.