Writing

Honest unit economics of a small AI automation studio

·4 min read

AI automation studios look deceptively simple from the outside.

A client pays ₹50,000 for an automation. You connect a few APIs, build an agent, deploy it, and move on to the next project.

But revenue is not profit.

When you're running a small studio, the real unit economics are determined by much more than API costs.

The real cost of an AI automation project

For every project, we think about five layers of cost.

1. Acquisition

Finding the client takes time. Cold outreach, calls, networking, demos, proposals, follow-ups, and sometimes multiple conversations happen before a project is signed.

If it takes 20 conversations to win one ₹50,000 project, those 20 conversations are part of the acquisition cost, even if no money was directly spent on advertising.

2. Engineering

This is usually the largest hidden cost. A seemingly simple workflow can involve:

  • API integrations
  • Authentication
  • Database design
  • Background jobs
  • LLM orchestration
  • Prompt engineering
  • Error handling
  • Monitoring
  • Deployment
  • Testing
  • Client-specific edge cases

The first version may take two days. Making it reliable can take two weeks.

3. AI and infrastructure

LLM calls are only one component. There can also be costs for:

  • Model APIs
  • Web search
  • Voice APIs
  • Hosting
  • Databases
  • Vector storage
  • Queues and Redis
  • Email, SMS, WhatsApp
  • Monitoring and logging

The mistake is optimising only the token bill while ignoring the infrastructure around the model.

4. Support and maintenance

A project doesn't necessarily end when it goes live. Clients will discover new edge cases. APIs will change. Credentials will expire. A workflow will occasionally fail, and someone has to investigate it.

If maintenance isn't priced into the business model, the founder eventually becomes the unpaid support team.

5. Founder time

This is the cost most small studios ignore. When the founder does sales, solution design, engineering, deployment, customer support, and account management, a ₹50,000 project can look highly profitable.

But if the project consumes 60 to 80 hours of founder time, the economics look very different.

Revenue per project isn't the metric

A better way to think about the business is:

Contribution margin = project revenue - variable delivery costs

Then separately track:

Founder-adjusted profit = contribution margin - cost of founder time

This gives a much more honest picture.

For example, imagine a ₹60,000 automation project. If direct software and infrastructure costs are ₹8,000, it initially looks like:

₹60,000 - ₹8,000 = ₹52,000

But suppose delivery requires 50 hours. At an internal founder cost of ₹1,000/hour:

50 × ₹1,000 = ₹50,000

The apparent ₹52,000 margin becomes only ₹2,000 of economic profit.

That doesn't mean the project was bad. It means the studio needs to either charge more, deliver faster, reuse more components, reduce customisation, or increase the lifetime value of the client.

This changes how we build

The goal isn't to build every client project from scratch. We are increasingly thinking in terms of reusable infrastructure:

reusable agent patterns -> reusable integrations -> reusable deployment -> reusable monitoring -> reusable workflows

Every project should make the next project cheaper and faster to deliver. That's where a small AI studio starts becoming a real business rather than a collection of freelance projects.

The metric we care about

The question isn't:

"How much did we charge?"

It's:

"How much value can we deliver per unit of founder time?"

If we can move from ₹30K projects taking 40 hours to ₹1L projects taking 20 hours, the business changes dramatically. The AI model is only one part of that equation.

The real moat of a small automation studio is operational leverage.