AI automation in your company: a basic guide to your first automations and scheduled tasks

When the word automation comes up during AI training, most people picture either something huge and expensive or, on the contrary, a magic button that does the work for them. The reality I build in companies and for myself is more prosaic and more useful: an automation is a repeating activity you've described so precisely that a machine can run it – and you only check it. In this article we'll go through what that means in practice, which three automations make sense first, how scheduled tasks in today's AI assistants work, which rules never to break, and how to tell when a no-code tool is enough and when it's time for a custom application.

What automation is: trigger → action → check

Forget AI for a moment. An automation is the answer to the question "what should happen when…". It always has three parts:

1. Trigger. An event or a time. An enquiry email arrived. It's Monday 8:00. A customer filled in a form. A row was added to a spreadsheet.
2. Action. What should be done. It used to be only fixed steps (copy, send, save). AI added the step a person used to have to do: read, understand, sort, summarise, draft a reply.
3. Check. Who sees and approves the result. With AI automations this step matters most, because AI is sometimes wrong – and a mistake that's automatically sent to thirty people is a different mistake from one you catch in a draft.

Once you adopt this framework, you stop asking "what could AI possibly do" and start asking "which repeating activity has a clear trigger and can be checked". That's the right question.

Scheduled tasks: automation without a single extra tool

The simplest form of automation today needs neither Make nor a programmer. AI assistants have scheduled tasks: you tell them "every Monday at 8:00 go through this and send me a summary" and it happens. ChatGPT has a scheduled tasks feature (one-off and recurring, including monitoring for changes and notifications), Claude has scheduled tasks in Cowork that run on a schedule and deliver finished outputs – reports, overviews, summaries. Availability differs by plan and both are evolving quickly, so check the current help pages of the tool you use (links at the end of the article).

What it's good for right away:

• a Monday overview of what appeared in your field or at competitors during the week, with links,
• a Friday summary of the week from notes you drop into one document during the week,
• a reminder with a pre-prepared draft: "it's the end of the month, here's a draft email to clients asking for documents".

Limitations: a scheduled task in an assistant usually doesn't have full access to your company systems and doesn't have an output that automatically goes somewhere further. It's a great start and a test of whether the output helps you at all – not the final destination.

The first three automations for a small business

After years of training and building tools, I see that the same three always pay off first. They aren't the flashiest; they're the most useful and the safest.

1. Sorting and drafting replies to emails and enquiries – with human review

Trigger: a new email in the enquiries inbox or a submitted form from the website. Action: AI reads the message, assigns a category (enquiry, question about an order, invoice, spam), estimates priority, extracts key facts (what they want, by when, budget if stated) and prepares a draft reply in your tone. Check: the draft lands as a draft message or in an overview; a person reads it, edits it and sends it. Nothing goes out without a human click.

Why first: enquiries are where response speed directly affects revenue, and at the same time where an automatic mistake hurts most. A draft with human approval gives you speed without the risk. Once you have a few hundred approved drafts behind you and know where the AI goes wrong, you can consider automatic sending for the simplest categories – but not before.

2. A weekly report from data you already have

Trigger: Monday morning. Action: load numbers from a spreadsheet, accounting, analytics or CRM, compare with last week, write five sentences on "what happened and what follows", flag deviations. Check: the report goes to you and you decide what to do with it.

Why second: most small businesses have data nobody looks at regularly because preparing it takes an hour. An automated report turns "we'll look when there's time" into "I see it every Monday". Condition: the data must be in one place and readable – often the first step of the automation is actually tidying up the spreadsheet.

3. Preparing content and materials

Trigger: a new entry in the content plan, a new product, a new order. Action: AI prepares a first version – an article outline, variants of social posts, a product description from a template, meeting materials from available information. Check: a person edits and approves. Materials are a draft, not the final.

Why third: it saves time on activities where the first version costs the most effort and where a mistake has no immediate impact. How to do it so the results aren't generic is in How to teach your team to use AI; an overview of tools is in AI tools for entrepreneurs.

Rules not to break

• Human in the loop. Everything that goes out of the company (an email to a customer, a post, an invoice) is approved by a person. Only things pointing inward (a report, sorting, a draft) run automatically without approval.
• Sensitive data doesn't belong in public tools. Customers' personal data, pricing, contracts, health or financial information – either a business plan with clear data-processing terms, or a custom solution, or not at all. Check in advance who the processor is under GDPR. I summarise the legal framework including AI obligations in the article on the AI Act for Czech companies.
• Log everything. Every run of an automation leaves a record: when, what it loaded, what it created, who approved. When something goes wrong, you want to know where – and without a log you won't.
• Start where a mistake doesn't hurt. Internal reports and drafts before anything that goes to customers.
• Measure. How many hours a week the automation saved, how many drafts had to be rewritten. Without that you can't tell whether it pays off.

When no-code is enough and when you need a custom application

No-code tools – Make, Zapier, n8n (which can also run on your own server) – connect ready-made services: email, spreadsheets, CRM, AI models, chat. They're enough when:

• you're connecting common tools that have ready-made integrations,
• the rules are simple (if–then, a few branches),
• the volume is hundreds, not tens of thousands of runs per month,
• a few hours of downtime won't kill anyone.

You need a custom application when:

• the automation works with your own data and needs to store, look up and link it (customer history, orders, inventory),
• the rules are complex or change often and the no-code tool has grown an unreadable "spaghetti" scenario nobody understands,
• the operation depends on the result – you need monitoring, retries on error, permissions and an audit trail,
• the per-run costs of the no-code tool have outgrown the price of a custom solution, or the data must not leave your infrastructure.

The typical path I recommend: start with a scheduled task in an assistant, second step Make/n8n, and build a custom application only for a process that has proven itself and hit a ceiling. Most companies find out this way that they need custom development for one or two processes, not for everything.

How to start – a plan for the first month

1. Write down the team's repeating activities for a week. Not ideas, real activities with the time they take.
2. Pick one with a clear trigger, where a mistake doesn't hurt and where you'll save at least an hour a week.
3. Describe it on one page: trigger, steps, what a good result is, who checks.
4. Build a pilot in the simplest tool that's enough – a scheduled task in an assistant is fine.
5. Watch the log and the rewrite rate for two weeks. Adjust the instructions.
6. Only then the next process. A company with three working automations is further ahead than a company with twenty half-finished ones.

This is exactly the process I take teams through in corporate AI training – we finish with one automation built on the company's real process, not with a presentation. Anyone who wants to try it themselves first starts with the open course. And when a process hits the ceiling of no-code tools, I build a custom application – always checking first whether you really need custom code.

Summary

• Automation = trigger → action → check. Ask which repeating activity has a clear trigger and can be checked.
• Scheduled tasks in AI assistants (ChatGPT, Claude) are the simplest start with no extra tools – suitable for overviews, summaries and reminders with a draft.
• The first three automations: sorting and drafting replies to enquiries with approval, a weekly report from data, preparing content and materials.
• Rules: human in the loop for everything going out; no sensitive data in public tools; log every run; measure the benefit.
• No-code (Make, Zapier, n8n) is enough for common connections and simple rules; a custom application comes into play with your own data, complex rules and processes the operation depends on.

Frequently asked questions

How many automations can a small business introduce at once?

One. Introduce it, watch it for two weeks, fine-tune it – and only then the next. Three working automations the team understands and trusts are worth more than twenty half-finished ones.

Can I let AI reply to customers automatically?

Not at the start. Let AI prepare drafts and a person approve them. Only after hundreds of approved drafts, when you know where the AI goes wrong, can you consider automatic sending for the simplest categories – with a clear log and the ability to switch it off immediately.

Do I need a programmer?

Not for the first automations – scheduled tasks in an assistant and tools like Make or n8n can be handled by a trained person on the team. A programmer or a custom application comes into play when the automation works with your own data, has complex rules or the operation depends on it.

What data may I send to AI tools?

Without hesitation, only public or internal data without personal information. Customers' personal data, contracts, pricing and sensitive information only into tools where you have data processing contractually covered, or into a custom solution. Check in advance who the processor is under GDPR.

Sources and links

• Scheduled Tasks in ChatGPT – OpenAI Help Center – one-off and recurring scheduled tasks in ChatGPT
• Schedule recurring tasks in Claude Cowork – Claude Help Center – scheduled tasks in Claude
• Make, Zapier, n8n – no-code tools for connecting services
• GDPR Regulation (EU) 2016/679 – processing of personal data