Proving AI pipeline economics with numbers anyone can recompute

●1 ●7
calendar_today ago • schedule3 min read

I'm one developer. I built a working AI content engine at one company in the wellness field, for the US and other English-speaking markets. Agents wrote the software that edits and publishes the videos.

The hardest part was not the code. It was proving the economics with numbers someone else can recompute. This post is about that part.

What the engine does

Scripts, text, avatars and voice are made by AI. Editing and publishing are done by software. That software was written by Claude agents.

I don't name the company, its channels or its platforms here. The numbers are what I want to talk about. One person built the engine, and that person is me.

What it produced in 101 days

In the first 101 days the engine made 1,497 short videos. They went out on 67 channels.

Those videos collected 8,250,024 organic views. The channels gained 1,126,322 subscribers.

I like to start with output counts because they are easy to check. A video count and a channel count are plain numbers. Anyone can hold them next to the view count and see the scale.

What it cost

I counted the AI service subscriptions over the same 101 days. They came to $4,313.76.

From that I get two unit costs. It is $0.52 per 1,000 views and $2.88 per video.

I use the same 101 days for output, views and cost. If the periods differ, the unit costs mean nothing. A developer can check this in a minute.

The human-studio figure is not a saving

The report also asks what the same volume would cost if people made it. By the report's model, that needs 42 to 78 employees and $179,000 to $1.21 million a month. That is 138 to 935 times the subscriptions.

Buying the same reach from influencers would cost 3.8 to 57 times more.

I want to be careful with the human-studio number. It is the price of the same volume made by people. I don't call it a saving. It is a reference point, and I give it as a range.

From a report to a calculator

I published the method as a report with a DOI. Then I turned the cost method into a free open-source calculator called reachburn. It uses the MIT license and is on PyPI.

Install it and look at the options:

pip install reachburn
reachburn --help

The calculator computes the price per 1,000 views and per video. It computes the human-studio equivalent, which is roles x headcount x BLS median wages + shoot days. It also computes the price of the same reach through ads and influencers.

Results come as ranges, not single points. Every rate has a source and a date. The data does not leave your computer.

The goal is simple. Someone who doubts my numbers can run the same method with their own inputs.

What to measure before you call a pipeline cheap

I'd measure these things first:

  • Output: videos and channels over a fixed number of days.
  • Reach: views, and whether they are organic.
  • Cost: what you paid for AI services over the same days.
  • Unit costs: cost per 1,000 views and cost per video.
  • Comparison: the same volume made by people, and the same reach bought through ads or influencers, as ranges.

Write down the source and the date for every rate you use. A rate without a source cannot be checked. A rate without a date may be out of date.

Where this leaves me

My numbers are one case. One engine, one company, 101 days. I'm not claiming they fit your project.

I'm saying you can test your own project with the same method. If you build with agents, run your numbers through the calculator and compare them with mine. The report has the details.

The report: AI Content Engine Economics: The Cost of Short-Form Video Production by AI Agents and by People, DOI 10.5281/zenodo.22802228. Course page: Wikiversity: AI content engine economics.

Evgenii Arsentev, MD, PhD - creator of Arsminal; founder of ArsentevAI and Dad4Kids. ORCID 0000-0002-9120-7298 · Wikidata Q141398931

Want more #python & #ai-agents deep dives?

Join 5,986+ developers sharing real benchmarks, code breakdowns, and projects. Free forever.

More Posts

Dashboard Operasional Armada Rental Mobil dengan Python + FastAPI

Masbadar - Mar 12

Meet kalbee: State Estimation Without the Boilerplate

Vincente - Jul 18

Can Rust Make Unsafe AI Agent Actions Unrepresentable?

Ken W. Algerverified - Sep 6

Your Service Desk Data Is Smarter Than You Think. AI Is Finally Proving It.

Tom Smithverified - Jun 10

I Wrote a Script to Fix Audible's Unreadable PDF Filenames

snapsynapseverified - Apr 20
chevron_left
186 Points • 8 Badges
3Posts
0Comments
Evgenii Arsentev, MD, PhD (Евгений Арсентьев) — creator of Arsminal; Founder & CEO of Dad4Kids and A... Show more

Related Jobs

View all jobs →

Commenters (This Week)

2 comments
1 comment
1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!