9horsemenStart building
Essays

Is AI Automation Worth It? When Human Hands Are Cheaper Than GPUs

Published on 21 Sept 2026 by Shri Krishnan

An essay from 9horsemen's founder. For practical help running a business, see the blog. More thinking like this on our podcast, Why Your Software Sucks.

A candid, cluttered Vietnamese logistics office in Ho Chi Minh City with office workers seated at cramped desks typing manual data, showcasing the low-cost human labour market.

TL;DR:

  • Silicon Valley assumes global AI adoption is inevitable, but they are ignoring basic economics
  • In many markets, a room full of data entry clerks is simply cheaper than a bloated enterprise SaaS subscription.
  • Big tech ignores these localised, messy bottlenecks because they don’t scale fast enough for venture capital.
  • To actually win, software needs to stop acting like a shiny oracle and start acting like a digital tractor built for the mud.

Is AI automation worth it? Not always: if the people doing a job cost less than the software that would replace them, automating it loses money. Here’s the maths from a freight office in Vietnam, and what kind of software actually pays off.

Does every business need more AI?

No: much of the world’s business is held back by messy paperwork, not by a lack of clever AI.

Spend five minutes reading tech commentary and you will walk away believing the entire global economy is starved for cognitive compute power. The narrative is always the same: we just need more Nvidia GPUs, larger data centres, and models that can reason through complex logical puzzles.

The implicit assumption is that every business on Earth is just waiting for a smart enough “AI Agent” to finally automate their operations. We are told the future belongs to software that can pass the bar exam, write python scripts from scratch, and generate photorealistic videos on demand. It is a beautiful, trillion-dollar vision. And if you step outside of a major Western tech hub, it is completely irrelevant.

When is AI automation not worth it?

When it costs more than the people it replaces: a $2,500-a-month AI tool can’t compete with clerks paid $400 a month each.

Walk into a mid-sized freight forwarding office in Ho Chi Minh City. Forget the sleek dashboards of Silicon Valley. Under flickering fluorescent lights, you will find five clerks huddled over dual monitors. Their entire job consists of taking crumpled, coffee-stained PDFs of shipping manifests, squinting at hand-written supplier notes, and manually typing that data into a sluggish, twenty-year-old Oracle database.

Why hasn’t an aggressive, well-funded Asian startup automated this with AI?

Because the startup walks in pitching a slick, cloud-based document parsing tool that costs $2,500 a month. The freight forwarder looks at the maths. He pays those five clerks about $400 a month each.

Human hands are currently $500 cheaper than the GPU cycles required to replace them.

The software might be faster. It might even make fewer typos. But it doesn’t matter. The AI startup loses the deal because they are trying to sell a premium digital brain to an operator who just needs cheap, brute-force labour. The friction here isn’t a lack of technological capability; it is the ruthless reality of localised economics.

Why don’t big AI companies build cheaper local tools?

Because their model depends on one product sold to millions, not on handling each country’s forms, languages and old systems.

You might ask why the massive US software giants don’t just lower their prices or build something tailored for that freight forwarder.

They won’t. They can’t.

Venture capital models require software with near-zero marginal cost of distribution and 80% gross margins.

They build generalised hammers. The global supply chain is built out of stripped, rusted, highly specific screws.

What kind of automation is worth paying for?

Simple, reliable software aimed at one repetitive job, which often costs a fraction of a cent to run instead of a GPU bill.

We need to stop pretending every commercial business requires a multi-billion-dollar GPU cluster just to parse a PDF or draft marketing platitudes. Real economic leverage is never about generating more synthetic digital noise; it is about ruthlessly destroying administrative friction.

True operational progress isn’t an omniscient conversational chatbot that charges you per token while hallucinating basic maths. It is a digital tractor: lightweight, deterministic software engineered specifically to plough through the messy, repetitive reality of daily business operations. When a four-line Python script or a structured database trigger can solve a problem for a fraction of a cent, burning GPU cycles is financial malpractice.

That’s how we build at 9horsemen: use the expensive model only where it earns its keep, and plain, tested code everywhere else.

Our first product, Gallop, works that way for local service businesses. AI reads a description of the business; tested code builds its website, online booking and business app. It’s free to build, then one plan priced for your country (see pricing).

If you’re building software of your own, I write more about the economics in the other essays and talk about it on the podcast Why Your Software Sucks.