What to automate first in a small business: a practical AI automation checklist

Start with the repeating task that costs your team the most hours and carries the least risk if it goes wrong. For most small businesses that is one of six: order intake, daily and weekly reports, invoices and payment reminders, routine customer questions, social media posting, or lead follow-ups. Pick one of them, and keep a person approving anything a customer will see.

Why most owners cannot name what to automate

Ask owners what is eating their week and most can describe the feeling but not the tasks. You hear “admin” or “everything”. Those are categories, and a machine can only take over a specific step.

Take orders. An order arrives by email or WhatsApp. Someone reads it, types it into a sheet, and later types it again into the billing system. Most people would not list that as a task. It is simply part of taking an order, so it rarely comes up when you ask what takes the time.

This kind of work is hard to see from inside. It comes in small pieces, a few minutes after lunch and a few before closing, and the person doing it stopped noticing long ago. Some of it lives only in one person’s head, so nobody has the whole picture, including you.

Watch one normal week

You do not need software to find the tasks. You need an ordinary week, not a holiday and not a rush, and a notebook. Watch four things.

• Who repeats what. Each time someone does the same job a second time in a day, write it down, along with roughly how long it took.
• Which tools are open. Email, the spreadsheet, the accounting software, the chat app. Switching between two screens usually means retyping.
• Where work waits or gets retyped. A job that sits until one person is back from lunch, or gets typed into two systems, belongs on the list.
• What lives only in one person’s head. The supplier who must be called on Thursdays, the customer with special terms. If it is not written down, nobody else can do the job, and a machine cannot either.

A job belongs on the list if it repeats and follows more or less the same steps each time. At the end of the week, turn each one into hours: minutes per time, multiplied by times per week. Say a job takes 10 minutes and happens 20 times a week. That is over 3 hours a week. Write the total next to each job.

The checklist: six usual first candidates, ranked

Judge each candidate with two questions. How many hours a week does it cost? And if the automation gets it wrong, who sees the mistake, you or your customer? My order leans on risk more than on hours, so work that stays inside your business comes first. Your own list may reorder it.

1. Order intake from email, WhatsApp or forms. Someone reads each message and retypes it into a sheet. Email and web forms are the easiest to connect. WhatsApp and Viber take more setup. The hours grow with every order, and the data stays inside your business, so the risk is low. For messy messages, a language model can pull out the item and quantity, and anything it cannot read goes to a person instead of being guessed.
2. Daily and weekly reports. The same numbers, copied into the same format, every evening or every Friday. A workflow can collect the figures and send the report on schedule. Each run is short, but the hours add up over a year. The risk is low because the report goes to you. Automation delivers wrong numbers faster too, so put the source and the date on every report.
3. Invoices and payment reminders. Invoices go out on schedule, and a reminder follows when a due date passes, so nobody reads back through old invoices to see who still owes you. The hours are real if you chase payments by hand. The risk is higher, because this touches customers and money, and a reminder sent to someone who already paid is embarrassing. So the workflow checks the payment status first, and a person approves the early batches.
4. Routine customer questions. Price, location, opening hours, delivery: the same few questions, again and again. A bot can answer from a short sheet of facts you wrote, and hand the chat to a person the moment a question is not on the sheet. The hours depend on how many chats you get. The risk is higher, since a wrong answer goes straight to a customer, so keep the scope narrow. At Appynox we worked on a Telegram bot for Linkify Social that does one job: send it the link to a social post and it replies with the product link. One kind of question is easy to check.
5. Social media posting. One photo becomes four captions for four platforms, by hand, several times a week. This is the one I know best. The AI system I designed and run handles social media for a Philadelphia hospitality entrepreneur: 10 brand accounts, 13 specialised sub-agents, each with one job, and 160+ posts in the first five weeks. The hours grow with how often you post. A mistake here is public, so a person approves every post.
6. Lead follow-ups. Who was called, who was missed, and who nobody remembers either way. A sequence can send follow-ups on schedule, stop the moment someone replies, and sort replies so your people only spend time on the live ones. The gain here is often leads you stop losing, which hours do not show. These messages reach prospects, so a person approves the wording first.

What not to automate

Judgement calls stay with a person. Whether to refund an upset customer, or to make an exception for someone who has bought from you for years: a machine can gather the facts, but the decision is yours. A language model can also sound sure when it is wrong, which is a bad trait in anything that decides.

Anything you cannot stop or check stays off the list too. If you cannot see what an automation did, or switch it off quickly, it should not send anything to a customer.

Two smaller cases. If two people do the same job in two different ways, agree on one way first, or you will only automate the confusion. And a job that happens twice a year rarely pays back the time it takes to build.

Keep a person approving anything a customer will see

I do not build automations that send to customers without a yes from a person. A sent message cannot be taken back. Machines are good at checks that are the same every time, like whether an invoice is already paid. People are better at deciding whether something sounds like you.

In that social media system, every post passes an automatic check first, then a reviewer that has no memory of how the post was made, and only then reaches a person. When that person corrects something, the lesson is written into the instructions the agents read next time, so the mistake is not repeated. The video system we use at Appynox, which made 85 social videos for SameShot, checks every render automatically before a person looks at it.

Every automation should also log what it did and send you an alert when something fails, such as a form that stops arriving, so you find out before a customer does.

Start with one workflow, and use plain tools

Take the top item on your list and build only that. Test it against your real work, not a demo, then let it run. Read what it reports, fix what it gets wrong, and only then pick the next one. A platform that does everything at once is harder to check.

A word on tools. n8n, Make and Zapier connect apps to each other: when something happens in one, like an email arriving, they do something in another, like adding a row to a sheet. I pick whichever fits the job and the volume. Your data usually lives in Google Workspace or Microsoft 365 already, so I connect to what you pay for and do not move you somewhere new. A language model goes in only where the job needs judgement, like reading a messy message or drafting a reply for a person to approve. Everywhere else, plain rules usually do the job faster, and they do not make things up. If a setting in a tool you already pay for solves it, I will say so.

Get the list written for you

If you would rather have someone do the watching and the arithmetic, that is where my AI automation service starts: a free automation map. It is one call, then a written list of your repeating tasks, ranked by the hours they cost. The list is yours to keep whether or not you hire me. I work from Kathmandu with owners in the US and in Nepal, and the calendar shows your own time zone. To start, book a 30-minute call and walk me through a normal week.