Why hire humans?

unpolished thoughts:

we talked about this before but like it's kind of a joke how fast models improve vs humans

the only limits rn is that you are capped by time and money, let's say 100 human decisions per day, 5 will be meaningful, then only so much cash to spend on Claude/Codex plans.

 

things that are hard for agents :
1) sales calls
2) videos (like creating influencer style videos)

things its very good at :
1) scraping/understanding large data
2) handling support (given an SOP)

 

3) creating static images

4) writing emails

 

types of business its not good at :
1) sales heavy
2) influencer heavy where its mostly video

3) requiring ongoing R&D
4) trying to find unvalidated product ideas

 

products that its good at :
1) google ads / meta ads
2) validated product that doenst require constant R&D
3) low human touch post purchase

a lot of problems i think should be looked through the lens of trying to be solved by humans and solved by AI. One obvious task that is great for humans (at the moment) is just doing editing. While AI can generate scenes, it's still not great at editing i think. but AI is great at making shorts/super simple slideshows. but i think the lesson is to clearly push the limits and see what is still human defined rather than try to make everything AI only or human only. it makes sense to try to push it to the super extreme to see where humans can drive value.

let's say the task of selling spreadsheets:

there are several stages

i think info products make the best AI driven business at the moment because there are several stages that go into any business:

the only drawback with info product is that there is no real retention, besides possibly selling a retention offer as a post purchase upsell.

I think it makes sense to just assume that it will cost $30 to acquire a customer and work backwards from there then the recurring upsell is just gravy!

1) Finding a niche

- AI: With enough data, I think AI can find a proven market

- Human: AI can give the recommendation but I think human is still best to have final call since it's high enough leverage and everything else is downstream of the market + offer in this steps
How to decrease variables:  just find something that is proven to be working, then do it in a different market.

2) Acquiring traffic

- The traffic that makes most sense for AI is probably just Meta, the issue with a lot of the other traffic sources is that it adds a ton more complexity (warming up TikTok/IG accs)

- With Meta, it's simply "pay to win"

How to decrease variables:  just stiching/clipping organic content

3) Testing

it's probably worth it to test the waters with content on TikTok/IG and see if it gets any reaction

**4) Product creation (99% AI) **

Pretty straight forward with Claude Fable + subagents

**5) Support (99% AI) **

Using Codex/Claude to manage all of support

so it can spit out infinite ideas

but you need some focus on what to actually spend time and money on

I've only had one real corporate job (at Amazon) but from my personal experience the amount of bureaucracy would slow a project down by 1-2 months with all the approvals.

6) Maintence

i think with the newest Claude models, maintence should take <1 hour

7) Product improvement

This is tricky because at some point, there is lower returns in trying to improve the product than to just creating a new product (esp if it is not taking off fast)

 

one person company

let's imagine where are not allowed to hire any humans, what would a startup in that case look like?

 

in the past, a startup may invest a ton a large % of a startup costs may have towards marketing, product development, infra. now most of the costs may be towards paying meta for traffic, anthropic for intelligence, and the rest towards infra.

a 1m arr business may look more like.

50% margins