# First automations worth setting up in a small team

> Which tasks a small business should automate first, per official guidance in the US, Canada, and Australia, five starter workflows, and what trials found.

The **first automations** worth setting up in a small business are repetitive, low-risk tasks that you can check: sorting email, sending form responses to a spreadsheet, invoice reminders, and meeting follow-ups. That is what the US Small Business Administration, Australia's business.gov.au, and Canada's BDC suggest. As of November 11, 2026, no official study had measured the time such automations save in a small business.

**Key takeaways**

1. The UK's National Cyber Security Centre wrote in December 2025 that there is a good chance prompt injection will never be properly mitigated, and advises limiting what an AI tool that reads outside content can do.
2. In a UK government trial with a comparison group, Microsoft 365 Copilot users at the Department for Work and Pensions saved an estimated 19 minutes a day, based on self-reported measures.
3. Zapier's invoice reminder template pauses for one hour before sending, so that someone can hold or correct a reminder.


## Choosing your first automations

The best first automation is boring: a task you do every week, the same way each time, that would not hurt anyone if it went wrong once. This page is part of [automating work with AI](/guides/ai-at-work/automation-and-agents/) and draws on official small business guidance and the companies' own templates, read on November 11, 2026. The US Small Business Administration's [page on AI for small business](https://www.sba.gov/business-guide/manage-your-business/ai-small-business) starts with exactly that kind of task: sorting your email into inboxes by task.

Australia's [business.gov.au](https://business.gov.au/online-and-digital/artificial-intelligence) adds two pieces of advice worth following: look first at the AI features of tools you already pay for, such as your accounting software or CRM, and try AI in one or two areas of the business before anything else. Canada's BDC suggests a pilot of six to twelve weeks on a clearly defined task, measuring the time saved and the errors avoided.

### What official guidance suggests automating first

| Source | Tasks it suggests |
|---|---|
| SBA, US | Sorting email by task, deadline reminders, to-do lists, restocking, recording and summarizing meetings |
| business.gov.au, Australia | Data entry, quotes and invoices, spreadsheet formulas, appointment booking, task tracking |
| BDC, Canada | Data entry, email sorting, meeting summaries, filing and tracking invoices, drafting replies |

We found no equivalent UK government guide. A survey published by the UK government in January 2026 found that 16% of businesses used AI, mostly for content and administration; 84% of them said they checked the outputs at least somewhat.

### Five starter workflows, each with a human check

These recipes are our recommendation, built on templates published in the official galleries of Zapier, Make, n8n, and Power Automate. Each keeps a person at the point where a mistake would matter.

1. **Form responses to a spreadsheet.** A new form entry adds a row and notifies you. Zapier, Make, and Power Automate all offer this as a ready template. Check: a weekly look at the sheet is enough.
2. **Labeling incoming email.** An AI step reads each new email and applies a label such as Support, Sales, or Invoices. Check: the automation only labels; it never sends, archives, or deletes.
3. **Invoice reminders.** A new invoice in your accounting software triggers a filter, a pause, and a reminder email. Zapier's template waits one hour so that someone can hold or correct a reminder before it goes out.
4. **Draft replies for review.** An AI step drafts a reply in the email thread. One n8n template states the principle well: the draft stays available for human review and manual sending.
5. **Meeting follow-up.** A recording becomes a transcript, a list of action items, and draft emails to attendees. Check: drafts, not sent messages; and recording a meeting has consent rules of its own, covered in another guide of this series.

Skip what does not repeat often enough. A consultant quoted by the BDC puts it bluntly: with only 20 or 30 invoices a year, automating them is not really worth it.

### Three rules before you switch anything on

1. **Give the least access possible.** Australia's small business cybersecurity handbook, dated October 7, 2026, says to give an AI agent only the minimum access its task needs and to review permissions regularly.
2. **Make approval a setting, not a sentence.** The UK's National Cyber Security Centre says not to rely on prompting alone. Zapier's “Require approval before running,” off by default, applies however the prompt is worded; telling an AI “wait for my OK” does not.
3. **Keep the dangerous tools away from untrusted text.** The NCSC wrote that there is a good chance [prompt injection will never be properly mitigated](https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection) the way SQL injection was, and cites a simple rule: an AI that processes email from strangers should not hold tools that can send, pay, or delete. In tests the US National Institute of Standards and Technology published in January 2025, adapted attacks raised the success rate of hijacking one AI agent from 11% to 81%.

### How much time it saves, honestly

| Study | Method | Result |
|---|---|---|
| UK, cross-government trial, late 2024 | 20,000 civil servants; self-reported, end-of-trial survey | 26 minutes a day |
| UK, Department for Work and Pensions, published January 2026 | Users compared with non-users; self-reported | 19 minutes a day |
| UK, HM Revenue and Customs, published July 2026 | Licenses assigned at random; estimates reduced by about 20% | About an hour a week |
| UK, Department for Business and Trade, 2025 | Diaries, interviews, observed tasks | No evidence of improved productivity |
| Australia, whole-of-government trial, 2024 | Self-reported, by task | Up to an hour a day on some tasks; up to 7% lost time checking |

All five measured an AI assistant built into office software, in large government departments. None measured an automation in a small business; we found no official study that did. Treat any vendor's promise of hours saved as a claim, and measure your own pilot.

### Picking the tool

Once you know the task, see [how Zapier, Make, n8n, and Power Automate compare](/guides/ai-at-work/automation-and-agents/zapier-make-n8n/), including a worked cost example.

### When a task needs judgment

Some jobs do not fit fixed steps. See [what the desktop agents can do](/guides/ai-at-work/automation-and-agents/desktop-agents/), and what they ask before acting.