You don't need a company name, a logo or a course to start an AI automation agency. You need one business with one repetitive task, a workflow you can build and test, and a price you can defend. This post is a method for getting to that first paid project in India, with the scoping and pricing arithmetic spelled out.
I'm not going to give you income figures, client results or a "make ₹1 lakh a month" timeline. I have no credible data on any of that, and anyone quoting it to you is usually selling a course. Every number below is an invented example, labelled as one, so you can swap in your own.
If you're still working out whether you can earn from AI skills at all, start with the freelancing with AI in India post. This one assumes you've decided to sell automation specifically.
What does an AI automation agency actually sell?
Not "AI." It sells a workflow that now runs without a person doing it by hand. A client doesn't care that a language model reads their enquiry emails. They care that the enquiry lands in the sheet, the right person is pinged on WhatsApp, and nothing sits unread over a weekend.
The unit of sale is a workflow with a start, an end and a way to check it worked. Examples of the right size:
- A form submission becomes a CRM row, a confirmation message and a task for the owner.
- Supplier invoice PDFs arrive by email and the extracted fields land in a sheet for review.
- A customer asks a common question on WhatsApp and gets an answer from an approved FAQ, or is passed to a human.
The wrong size is "automate my business." If you can't draw it on one page, it isn't a first project.
How do you pick a niche for your first client?
Use a three-part filter. A niche passes only if all three are true.
- You can name five real businesses of that type. Not categories, actual names you could message.
- Staff repeat a digital task weekly. Copying data between tools, answering the same questions, chasing the same reminders.
- A wrong output is easy to spot. A misfiled invoice is visible. A subtly wrong strategy memo isn't. Early on, choose work where mistakes are cheap to catch.
Then narrow by what you already know. A former clinic receptionist understands appointment chaos. Someone from a trading family knows what invoice season feels like. Your past is your edge because you already speak the client's language.
Write the offer as one line: "I automate [specific task] for [type of business] so [named result], set up within [time]." If your line needs the word "AI" to make sense, rewrite it around the task.
What should a first paid pilot look like?
A pilot is the smallest project that proves value and gives you a reference. Make it a real, paid piece of work with these properties:
- One workflow. Not three.
- A fixed price and a fixed finish line. You decide together what "working" means before you start.
- A real data sample from the client, handled under whatever confidentiality terms you agree. Don't invent test data and call it a pilot.
- A review date. A week or two after go-live, you look at the results together.
Why paid? A free pilot costs the client nothing, so nothing pushes them to send access, answer questions or approve a design. Paid pilots get faster replies. Price the pilot to at least cover your floor (calculated below), and say plainly that a good result can grow into the full project.
Where do you find the client? The same routes as any beginner: people who already know you, local businesses you can walk into, and a few well-matched platform jobs. The freelancing post covers these in order, so I won't repeat it. Bring a sample built from fake or public data, plus a one-page description of the workflow you'd automate for them.
How do you scope one workflow?
Scoping is where agencies lose money. Run a 45-minute call and fill in this template before you quote. Copy it.
WORKFLOW SCOPING SHEET
1. Name of the workflow:
2. Trigger (what starts it): [form / email / schedule / webhook / manual]
3. Inputs and where they live:
4. Steps today (who does what, in order):
5. Output and where it must end up:
6. How many runs per week, and the peak:
7. Time per run today, by hand (minutes):
8. Who is the human reviewer, and what do they check?
9. What must NEVER happen (e.g. send a message to the wrong customer)?
10. What happens when it fails or is unsure? (default: notify a named person)
11. Tools/accounts the client already owns:
12. Who owns the accounts and credentials after handover?
13. Out of scope (write at least three items):
14. Done means: [measurable acceptance test]
Questions 7 and 8 matter most. Question 7 gives you the value number for pricing. Question 8 tells you whether the workflow can be fully automatic or needs a review step. For anything that sends messages to customers or touches money, default to a human approval step at first, and remove it later if the client wants and the error rate supports it.
After the call, send back a one-page summary of the answers and ask the client to correct it. That reply is your scope. The proposal prompts post has prompts that attack a draft scope for holes, and the SOW template post turns the sheet into a contract-style document.
How do you price an automation project?
Two numbers set a range, and your fixed price sits between them.
Floor: build hours (including testing, documentation and handover) multiplied by your minimum hourly rate. This is what you must charge to avoid working for less than you need.
Ceiling: the value to the client, which is time saved multiplied by what that time costs them, over a period they'd accept for payback.
Here's the arithmetic with invented numbers. This is an illustrative example, not a market benchmark.
| Item | Illustrative figure | Working |
|---|---|---|
| Runs per week | 30 | from question 6 |
| Minutes saved per run | 12 | from question 7 |
| Hours saved per week | 6 | 30 x 12 / 60 |
| Cost of staff time per hour | ₹300 | client's own estimate |
| Value per week | ₹1,800 | 6 x 300 |
| Value per year | ₹93,600 | 1,800 x 52 |
| Your build estimate | 25 hours | design, build, test, handover |
| Your minimum rate | ₹1,200 per hour | see the freelancing post |
| Floor | ₹30,000 | 25 x 1,200 |
The floor is ₹30,000 and the yearly value is ₹93,600. A fixed price of ₹45,000 would sit between them. At the illustrative value of ₹1,800 a week, that price pays back in 25 weeks (45,000 / 1,800), which is easy to explain to a client. If your sums show the value is below your floor, the project isn't worth doing for that client, however nice the tech is.
Three habits keep pricing honest:
- Use the client's numbers. Ask them what the task costs. Don't assume.
- Add buffer on the build estimate. Integrations with someone else's tools nearly always take longer than the diagram suggests. Whatever buffer you pick, pick it deliberately.
- Price care separately. Automations break when a form changes, an API changes or a token expires. Offer a monthly care plan as its own line (details in the handover runbook post). Don't hide ongoing work inside the setup price.
Don't discount because AI made you faster. Faster delivery means you can take more projects. The client is paying for the result.
What do you do about GST, contracts and getting paid?
Check with a chartered accountant before your first invoice. Whether you need GST registration, how payments from foreign clients are treated, what paperwork your bank wants for foreign receipts and how to record income are all questions with answers that depend on your situation and change over time. I haven't verified any of those rules against a primary government page for this post, so I'm giving none of them.
For contracts, don't rely on a template from a blog, including mine, for payment terms, liability and ownership of the work. Use a template as a starting point and have a lawyer review it once. Meanwhile, practical habits cost nothing: take part-payment upfront, keep a written record of what the client approved, and invoice promptly on delivery.
What can go wrong in your first project?
- Credentials in your name. If you build in your own accounts, you've created a hostage situation for both sides. Set up client-owned accounts from day one (see the runbook post).
- Silent failures. A workflow that errors with nobody watching is worse than none. Add an alert to a named person for every failure path.
- Customer data in the wrong place. Be explicit about which tools see which data. Don't paste client data into tools their contract or yours doesn't allow.
- Scope creep. "Can it also do X?" gets a written change request with a price, not a yes in a chat.
- Overpromising. Promise "30 test runs reviewed together by Friday." Don't promise "saves 90 percent of your time."
- Model changes. If a language model step is inside the workflow, its behaviour can shift when the provider updates it. Agree on a monthly spot check, and put it in the care plan.
A 21-day plan to a first pilot
| Days | Do | Done when |
|---|---|---|
| 1 to 3 | Choose a niche using the three-part filter. Write the one-line offer. | You can name five businesses |
| 4 to 8 | Build a sample workflow on fake data. Write the one-page workflow summary. | A demo you can show in five minutes |
| 9 to 15 | Message 20 businesses or contacts. Book scoping calls. | At least two calls held |
| 16 to 21 | Send a scoping summary and a fixed-price pilot quote with the floor and value arithmetic done. | One quote out and a date to decide |
If twenty messages produce nothing, change the offer before you send more. I can't tell you how long a first client takes. It varies a lot, and nobody who gives you a date has your inbox.
If you're building the workflows in n8n, the n8n with the Claude API tutorial is a concrete build to practise on before your first client call.
