To automate tasks with AI in a small business, pick the tasks that save time, bring in money or keep money you already earn, then use AI for the steps that involve reading or writing: summarising, sorting, drafting and pulling details out of messages. Keep plain rules for everything that has a fixed answer, and put a person in front of anything that goes to a customer or changes a record.
This article is for owners and managers, not developers. Where it names a tool, the detail comes from the vendor’s own help pages as read in November 2026, so check their pages for current features and prices.
Start with the task, not the tool
The mistake most owners make is buying an AI tool and then looking for a use. Reverse it. Write down the tasks that repeat in your week, and score each against three questions.
- Does it save time? How many hours a week do you or your staff spend on it?
- Does it bring in more money? Does a slow or missed step lose enquiries, quotes or bookings?
- Does it keep money you already make? Does it chase late invoices, renewals or repeat customers?
A task that scores on none of the three can wait, however clever the AI looks. A task that scores on two is a strong candidate. Here is a worked example with made-up numbers. A four-person service business spends about 10 hours a week on admin that is mostly copying details from emails into a job list. At a staff cost of $30 an hour, that is $300 a week or roughly $15,600 a year. If automation removes even half, the saving covers a modest build. If the same business replied to leads faster and won one extra $2,000 job a month, that is $24,000 a year. Use your own figures, because the comparison is what matters.
How to automate tasks with AI: which steps suit AI
AI is good at tasks that involve language. It is not needed for tasks with a fixed rule, such as “send a reminder the day before”. Split each task into steps and pick the right tool for each.
| Step | Use | Example |
|---|---|---|
| Fixed rule, fixed answer | A plain automation | Send a reminder 24 hours before a booking |
| Reading messy text | AI | Pull the address, job type and urgency from an email |
| Sorting | AI, then a rule | Label an enquiry as quote, complaint or supplier |
| Summarising | AI | Summarise a long email thread into three lines for a record |
| Drafting | AI plus a person | Draft a reply that a staff member approves |
| Decisions with money or risk | A person | Approve a refund or a discount |
Simple AI steps an owner can try this week
You do not need to build anything to learn what works. Try these with the tools you already pay for.
- Summarise. Paste a long customer email thread into an AI assistant and ask for a three-line summary and the next action. Check it against the thread.
- Draft. Give the assistant three of your best past replies and ask for a draft to a new enquiry in the same voice. Edit before sending.
- Extract. Ask it to turn a voicemail transcript into a list: name, number, job, address, urgency.
- Sort. Give it ten recent enquiries and ask it to label them. See where its labels disagree with yours, and write that rule down.
Some tools build the automation from a plain-English description. Microsoft’s Power Automate says you can create a scheduled flow using natural language in Copilot, and its tutorial adds that a person should check AI-generated content for accuracy (Microsoft: run a cloud flow on a schedule, Microsoft: get started with cloud flows). HubSpot’s workflows can include AI actions that summarise records or fill in properties, and it says AI actions use HubSpot credits (HubSpot: workflows FAQ).
Where an AI step needs a human check
AI can be wrong in a confident tone, so decide in advance which steps get a check.
- Anything that goes to a customer. Draft, then approve. Once a type of message has been right for a month, you can let that type send itself.
- Anything that changes a record. HubSpot’s connector for Claude, for example, can create and update records, and HubSpot recommends setting its write tools to Needs Approval so it asks first. The same page says bulk changes are limited to 10 records at a time (HubSpot: the connector for Claude).
- Anything involving money, legal or medical advice. Keep a person in charge.
- Anything with private data. Check what the tool stores and who can see it before you paste customer details in.
A four-week plan to automate your first task
- Week one: choose. List ten repeating tasks, score them on the three questions, and pick one. Write down each step as it is done today, including who does it and how long it takes.
- Week two: do it by hand with AI. Run the task manually with an AI assistant doing the reading and drafting. Note where it gets things wrong, because those are the steps that need a check.
- Week three: connect the easy parts. Automate the fixed-rule steps first, such as creating the task or sending the reminder. Keep the AI steps as drafts for a person.
- Week four: measure. Compare the hours spent, the speed of replies and the mistakes with the numbers you wrote down in week one. Keep what saved time, drop what did not.
Most failed automations skip the second week. Owners connect the tools, discover that the process itself was unclear, and blame the AI. Doing it by hand first shows whether the task is well defined enough to automate at all.
Examples from admin, sales and service
Admin. An incoming supplier invoice is read, the amount and due date extracted, and a task created for the person who pays bills. A person approves payment.
Sales. A new enquiry arrives by email. AI labels it, drafts a reply that asks the one question you always ask, and creates a follow-up task for tomorrow. The owner approves the reply on a phone. See our guide to quote follow-up emails for wording.
Service. After a job, AI turns the technician’s voice note into a short job summary for the record and drafts a thank-you with a review request. A person checks the summary before the customer sees it.
When a custom build fits better
Try the simple steps first. They cost little and show you which tasks are worth more. A custom build becomes sensible when a task joins several systems, such as an email arriving, a record created in your CRM, a calendar booking and an invoice. CentriWeb builds these, usually in n8n, and sets the AI to answer from your own information and hand over to a person when it should. It looks after them afterwards, as part of its run with less admin service. We build custom work, so weigh our view with that in mind.
It does not suit a task you do twice a month, or a business that has not yet written down how the task is done. Fix the process first. Related reading: AI for small business: where to start, is AI worth it for a small business and our AI page.