Build Log #2: What AI Actually Does for a Founder

I run a company with AI agents. They build analysis modules, monitor production, review errors, generate content, deploy code. I have many automated operational skills and several specialized agents. From the outside it looks like AI is doing everything.

It's not. And the distinction matters if you're building anything with AI.

AI Can Only Reflect

Here's what I learned at 1am arguing with Claude about my own pitch deck.

I asked it to write a tagline for my company. It gave me: "AI-powered statistical analysis reports. For businesses that need answers, not dashboards." That's the average of every SaaS tagline ever written. It's not wrong. It's just nothing. It's what you get when you average a million pitch decks together.

I pushed back. It gave me ten more. Some were better. None were great. The best ones were rearrangements of things I'd already said in the conversation: "ChatGPT with a knife," "your business data deserves better than a chat window," "the analyst you can't afford." Those were my ideas. The AI caught them, cleaned them up, and presented them back to me.

That's what AI does in my work. It often echoes the ideas and common words I give it. I use "reflect" as a simple name for this pattern. It is not a rule for all AI output. The AI can rearrange, combine, and polish, but the idea has to come from me.

The AI Obvious

I started calling it the AI Obvious. Left to itself, AI output tends to drift toward the most common way people write about a topic. That makes it useful for a first draft, and the part I have to add is whatever it would not produce on its own.

In my view, this is why AI-generated content so often sounds the same, and why many AI pitch decks have the same structure. It's why ChatGPT's marketing copy reads like it was written by a committee of everyone who ever wrote marketing copy, because it was.

The AI Obvious is fine for things that should be average: legal boilerplate, code documentation, data entry. It's death for things that need to be distinctive: your brand, your pitch, your story, your content.

If your marketing sounds like AI wrote it, it sounds like everyone else's marketing. You've paid money to be invisible.

So What Is AI Good For?

In my work, AI can help me sort an idea and make a draft when I give it a start and I check the result.

When I said "we're like Claude's nerdy cousin who actually studied statistics," the AI didn't come up with that. I did. But the AI took that one sentence and built it into a one-page pitch document with a usable structure, sections, consistent messaging, and formatting. I did not time it.

When I had a scattered idea about using our own analytics platform to test whether writing down goals changes outcomes, the AI didn't generate the idea. But it organized it into a blog section, connected it to the product story, and made it coherent. I did not time that either.

The AI holds all the pieces while you figure out which ones matter. It catches the good ideas as fast as you can throw them and puts them where they belong. It builds the structure around your insight at the speed of typing.

That's not intelligence. That's a very fast mirror. But a very fast mirror is an incredible tool if you know what you're looking at.

A Made-Up Timing Example

None of the numbers in this example are real. I did not time my own work. The script below makes up drafting times with a fixed seed, so you can run it and get the same result. It also shows the tagline idea going from my input to a generic draft, an AI-sorted draft and my edit.

import random

SEED = 23
rng = random.Random(SEED)

idea = "ChatGPT with a knife: challenge a business's numbers, not just summarize them."
generic = "AI-powered analytics that deliver insights for better decisions."
organized = "ChatGPT with a knife for your business data: challenge the numbers, then explain what they mean."
human_edit = "ChatGPT with a knife. Your business data deserves better than a chat window."

solo_minutes = [rng.randint(90, 150) for _ in range(5)]
assisted_minutes = [rng.randint(20, 45) for _ in range(5)]
mean_solo = sum(solo_minutes) / len(solo_minutes)
mean_assisted = sum(assisted_minutes) / len(assisted_minutes)

print(f"Simulated data, seed {SEED}")
print(f"Human input: {idea}")
print(f"Generic first draft: {generic}")
print(f"AI-organized draft from the input: {organized}")
print(f"Human edit: {human_edit}")
print(f"Simulated solo drafting times (minutes): {solo_minutes}")
print(f"Simulated AI-assisted drafting times (minutes): {assisted_minutes}")
print(f"Mean simulated solo time: {mean_solo:.1f} minutes")
print(f"Mean simulated assisted time: {mean_assisted:.1f} minutes")
print(f"Simulated time difference: {mean_solo - mean_assisted:.1f} minutes")

Printed output:

Simulated data, seed 23
Human input: ChatGPT with a knife: challenge a business's numbers, not just summarize them.
Generic first draft: AI-powered analytics that deliver insights for better decisions.
AI-organized draft from the input: ChatGPT with a knife for your business data: challenge the numbers, then explain what they mean.
Human edit: ChatGPT with a knife. Your business data deserves better than a chat window.
Simulated solo drafting times (minutes): [149, 139, 150, 108, 147]
Simulated AI-assisted drafting times (minutes): [22, 20, 38, 29, 33]
Mean simulated solo time: 138.6 minutes
Mean simulated assisted time: 28.4 minutes
Simulated time difference: 110.2 minutes

In this made-up test, my idea shows up in the AI-sorted draft, and I then edit it. The made-up mean for solo drafting is 138.6 minutes and for AI-assisted drafting is 28.4 minutes. Both are made-up numbers, and so is the 110.2 minute gap. They do not show how long a real pitch takes.

How This Changes How I Build

I stopped asking AI to generate ideas. I start the conversation with my idea, however rough, and use AI to develop it. In my work, a clear start often helps me get a better draft. It does not ensure a good result.

For the company, this means:

I write the first draft of anything that needs to be distinctive. The pitch, the positioning, the brand voice. Then AI structures, expands, and formats. Never the other way around.

I let AI own anything that should be average. Code documentation, SQL migrations, deployment scripts, operational monitoring. These benefit from being standard. The AI's tendency toward the average is a feature here, not a bug.

The agents handle the operational baseline. Automated skills run daily checks, error monitoring, QA rotations and traffic analysis. This is where AI shines: reliable, consistent, tireless execution of well-defined tasks. No creativity needed.

The creative work stays with me. Product positioning, marketing angles, strategic decisions, content that needs to be distinctive. AI accelerates the execution but the direction comes from a human.

The Takeaway for Anyone Building with AI

If you're using AI to generate your strategy, your marketing, or your product vision, you're getting the average. You're getting what everyone else with the same prompt would get. You've automated being unremarkable.

If you're using AI to execute your strategy faster: to organize your ideas, build your structures, run your operations, hold all the pieces while you think. You're getting help with the writing and sorting. These cases show AI helping me write and sort work. They do not show how much one person can do.

The difference between those two things is the difference between a company that sounds like AI built it and a company that sounds like a person built it with AI. I think customers and investors notice the difference.

Drop the stone above the average. Let AI catch it. That's the whole game.