AI Automation Examples in Nepal: What Ten Kinds of Business Actually Automate

The test for whether a process is worth automating is well understood by now. What people actually ask me is narrower: what would this look like for a business like mine?

So here are ten kinds of business I see in Nepal, the one process each of them automates first, and what it involves. The tools are the same across all ten: n8n, usually self-hosted, a language model API for the steps that need judgement, and whatever the business already uses for messaging and records.

The pattern behind all ten

Every automation below has the same shape:

1. A trigger — a message, a form, a file, a time of day.
2. A structured record — the messy input turned into fields.
3. An action — a reply, a document, an entry in another system.
4. A human checkpoint — one person who sees the output before it goes anywhere that matters.

The checkpoint is not a weakness in the design. It is the design. Automating the checkpoint itself is the version of this that goes wrong.

Ten businesses, ten first automations

The running cost column is deliberately boring. A small server for self-hosted n8n costs roughly USD 5 to 10 a month and runs thousands of workflow executions for that flat price. Language-model calls for a few hundred enquiries a month typically cost less than the server. SMS is the one line that scales with volume, because Nepali gateways charge per message.

A closer look at four of them

The retail quotation

This is the one I have built most often and described in detail elsewhere. A customer sends "20 bags of 50kg cement, delivery to Bhaktapur, price?" on WhatsApp. The workflow extracts the product, quantity and location, looks up the current price sheet, drafts a quotation, and puts it in front of the owner with one button: send.

Before automation, the owner did this by hand between other things and answered in an hour if they were free. After, the customer has a quote in two minutes and the owner has spent ten seconds. The checkpoint stays because prices change and the owner knows things the sheet does not.

The consultancy enquiry

Education consultancies in Kathmandu lose students in the gap between the first message and the first call. The workflow takes a WhatsApp or form enquiry, creates a record with the student's target country, intake and qualification, sends an acknowledgement with the document checklist for that country, and schedules a reminder for the counsellor.

The interesting part is the weekly review: the counsellor sees every enquiry that has had no human contact in three days. That list did not exist before, because nobody was keeping it.

The manufacturer's purchase orders

A manufacturer receiving purchase orders as PDFs by email was retyping each one into their order sheet. The workflow reads the attachment, extracts the line items, and adds them to the sheet with the PDF linked. Dispatch checks the quantities against the physical order before anything ships.

This one has a limit worth naming: it works on typed PDFs. Purchase orders that arrive as photographs of handwritten forms go to a person, and the workflow knows to route them there rather than guess.

The NGO's field reports

Twenty field officers submitting weekly reports as free text produced a pile nobody could read in aggregate. The workflow turns each report into structured fields (district, activity, beneficiaries, issues), and every Monday drafts a one-page summary for the programme manager, with the outliers flagged.

The summary is a draft. The manager reads it, corrects it, and sends it. The value is that reading twenty reports became reading one page and checking three things.

The three that are never worth it

I get asked about these regularly and the answer is always no.

Anything that runs fewer than twenty times a month. Building an automation costs more than doing the task by hand for a long time if the task is rare. The threshold I use is frequency first, then everything else.

Anything whose inputs are photographs of handwriting. Models are improving here, but a workflow that is right 85% of the time on handwritten forms creates more checking work than it removes. Fix the input first: a typed form, a template, a photo of a printed page.

Anything where the rules change when you ask. If two people in the business give different answers to "what happens when a customer asks for credit?", there is no process to automate yet. Write the rule down, run it by hand for a month, then automate it.

Where to start

If your business is not in the table, it almost certainly still has the first row: an enquiry arriving somewhere, waiting for a person, and turning into a record only if someone remembers. That is the one to build first, and the seven processes worth automating first covers what usually comes next.

If you would like me to look at your specific process, the AI automation service page explains how a scoping conversation works and what it costs. If the process needs a system of its own rather than a workflow on top of existing tools, that becomes a custom software conversation instead, and I will say which one it is.