Back to Learn

What AI Automation Actually Costs in India

Real numbers, what makes one workflow cost five times another, and the running costs that only show up in month three. Written because most pricing pages in this industry say "it depends" and stop there.

Most agencies in this space will not put a number on a page. The reason given is that every project is different, which is true and also convenient. The result is that buyers cannot tell whether a quote is reasonable until they have collected three of them.

So here are the numbers, including ours, along with what actually moves them.

The published ranges

For workflow automation work in India, the figures that appear consistently across the market:

  • A basic workflow connecting three to five applications: roughly Rs 25,000 to Rs 60,000.
  • A complex workflow with custom API integrations, AI steps, error handling and multi-branch logic: roughly Rs 80,000 to Rs 2,50,000.
  • Hourly engagement with Indian agencies: around $25 to $60 an hour, which is where longer or less defined work tends to land.
  • Ongoing retainers for maintenance and iteration: from about Rs 25,000 a month at the low end, rising with volume and response-time commitments.

Those ranges are wide for a reason, and the reason is worth more to you than the numbers.

What makes one workflow cost five times another

How many decisions it has to make

A workflow that moves data from A to B on a fixed rule is cheap. A workflow that has to read something and decide what it is costs considerably more, because the decision has to be specified, tested against real examples, and given a sensible fallback for the cases it gets wrong.

"When a form is submitted, create a CRM record" sits at the bottom of the range. "When an email arrives, work out whether it is a new enquiry, a reply, a complaint or noise, and route it accordingly" sits near the top. The trigger looks similar. The work is not comparable.

How messy the input is

Structured input is cheap. A webhook delivering clean JSON needs no interpretation. Unstructured input is where the money goes: scanned PDFs at varying quality, emails with the real content in an attachment, WhatsApp messages written in a mix of English and Telugu, handwriting.

Our SmartDoc OCR work sits in this category, and document extraction is consistently the most underestimated line in any quote. The demo works on the clean sample. Production has to cope with a photograph taken at an angle in poor light.

How many systems have to agree

Each additional system adds more than its own integration. It adds a way for the workflow to be half-finished. If a lead has to land in the CRM, trigger a WhatsApp message and update a sheet, and the second step fails, you now need to decide what happens to the first. Getting that wrong produces duplicate messages to customers, which is worse than no automation.

Whether anyone has written down how it works today

This is the cost driver nobody expects. If the process lives in one person's head and has fourteen exceptions they have never articulated, a meaningful part of the project is discovering those exceptions. Teams with a documented process routinely pay less for the same build.

The costs that appear in month three

Build cost is the number everyone compares. Running cost is the one that surprises people.

Platform. n8n cloud starts around $20 to $24 a month for small teams. The community edition is free if you self-host, which trades a subscription for a small server and someone to keep it patched. Some competing platforms charge per task, which means a good month costs more than a slow one. At low volumes this is irrelevant. At ten thousand operations a month it is the largest line on the bill.

Model usage. If a step calls a language model, you pay per document processed. It is usually small per item and entirely predictable once you know your volume, but it is not zero and it scales with success.

Maintenance. The one people forget. APIs change, a vendor deprecates an endpoint, someone renames a column in the sheet the workflow writes to. Something breaks every few months in any system with external dependencies. Either you have someone who can fix it or you have a retainer.

A useful rule: assume running costs of ten to twenty percent of build cost per year, and be pleasantly surprised if it is less.

How to tell whether a quote is reasonable

Ask what happens when it fails. A quote that includes error handling, alerting and a retry strategy is describing a different product from one that does not, even if both say "workflow automation". The cheaper one is genuinely cheaper because it is less.

Ask who owns the workflow afterwards. If it is built inside an agency's own account, you are renting. If it is built in your infrastructure and you have the credentials, you own it. Both are legitimate. Only one lets you change provider without rebuilding.

Ask what it does not handle. Anyone who has built these knows the edges they chose not to cover. A provider who says it handles everything has either not thought about it or is not telling you.

When the honest answer is not to spend

Three cases where we say wait, and mean it.

Volume is too low. If a task takes ten minutes a week, automation will not pay back before the surrounding systems change. Frequency matters more than duration. Five minutes daily is a better candidate than two hours quarterly.

The process is still moving. If how you do this changed twice in the last quarter, building it now means rebuilding it next quarter. Let it settle.

Nobody owns the outcome. If no single person is accountable for the process working, there will be nobody to notice when it stops and nobody to feed corrections back in. That project fails regardless of how well it is built.

What we do

We scope fixed-fee where the work is clearly defined, because open-ended hourly billing on a well-understood problem mostly transfers risk to the buyer. Where the requirement is genuinely unclear, we would rather run a short paid discovery and quote properly afterwards than produce a number with a large margin of error built in.

The first conversation is free and takes about half an hour. You get a range and the reasoning behind it, including the cases where the answer is that the numbers do not work yet.

Our AI automation and AI agent services page covers the six areas we work in, and every project on our case studies page is real and shipped. If you want a number for something specific, book a scoping call and describe the task.

Back to all articles