Connecting tools and automating routine work
Learn how automations connect your apps, where to put ChatGPT inside them, what to automate and how to build in safety checks.
In this lesson you will
- Describe the trigger, action pattern
- Choose tasks worth automating
- Add an AI classifier with validation
- Design safe, reviewable automations
About 18 minutes
Your tools talk to each other, if you introduce them
Mina’s day: a customer fills in an order form, she copies it into a spreadsheet, writes a thank-you email, adds a task to her to-do app and pastes the details into a delivery note. Five tools, five manual copy-and-paste jobs, thirty orders a week. Each step is small, and together they steal hours. Automation means connecting tools so that one event triggers the next steps without you. Adding ChatGPT in the middle turns rigid rules into flexible ones: the assistant can read a messy message, decide what it is and write the reply. This lesson maps the ground: how connections work, what is worth automating, and how to build with safety.
Three ways to connect ChatGPT to your other tools
| Method | How it works | Skill needed |
|---|---|---|
| Built-in connections | Apps you link inside ChatGPT, so it can read files, calendar or documents with your permission (availability varies by plan and region) | None |
| No-code automation platforms | Visual tools where you choose a trigger and actions, including an AI step | Low |
| Your own code with the API | A script calls the model and other services directly (lesson 26) | Medium to high |
The trigger, action pattern
Every automation has the same skeleton:
- Trigger. Something happens: a form is submitted, an email arrives, a file is added, a time is reached.
- Steps. Data is read, checked, transformed.
- AI step (optional). The model classifies, extracts, summarises or drafts.
- Action. A row is added, a message is drafted, a task is created.
- Review. A human approves anything consequential.
Example: new order to tidy records
| Stage | What happens |
|---|---|
| Trigger | A new form response arrives |
| Step | Read name, product, quantity, address |
| AI step | Extract the delivery city and flag any unclear request as NEEDS REVIEW |
| Action 1 | Add a row to the orders sheet |
| Action 2 | Create a draft (not sent) confirmation email |
| Review | You read the drafts and send them |
Notice that the email is a draft. A good first automation prepares work for you rather than acting on its own.
What to automate, and what not to
| Automate | Keep human |
|---|---|
| Repeated, rule-based copy-and-paste | Decisions that affect money, health, law, or people’s jobs |
| Sorting and labelling incoming messages | Sensitive or emotional customer conversations |
| Drafting replies and summaries for review | Sending anything unreviewed to the public |
| Collecting data into tables | Anything you cannot undo |
| Reminders and reports | Tasks you still do not understand |
A useful rule: do the task by hand about ten times first. Then you will know the steps, the exceptions and the quality bar. Automating a process you do not understand just makes mistakes faster.
Designing a safe automation
- Start small. One trigger, two actions, one human check.
- Use drafts and logs. Save what the AI produced and what it received.
- Limit permissions. Use accounts and access that can only do what the automation needs.
- Handle failure. If the AI step returns something odd, route to a person, not to the customer.
- Protect data. Do not send personal or confidential data to services you have not approved. Check each platform’s privacy terms.
- Watch the cost. Automations can run thousands of times. Check usage limits and charges.
- Test with fake data before connecting real accounts.
Turning messy input into clean data
The classic AI step is an extractor and classifier. Here is the kind of prompt you place inside an automation, with a strict output so the next step can read it.
You will receive one customer message. Reply with JSON only, in this shape:
{"type": "order" | "question" | "complaint" | "other", "summary": string (max 15 words), "needs_human": true | false}
Set needs_human to true for complaints, refund requests, anything legal, or when unsure. Do not invent information.
Message: [the incoming message]
Remember lesson 22: validate the JSON before the next step uses it, and send anything invalid to a person.
Simulating a rule-plus-AI routing step
import json
replies = [
'{"type": "order", "summary": "Wants two lavender candles", "needs_human": false}',
'{"type": "complaint", "summary": "Parcel arrived broken", "needs_human": true}',
'Sure! Here is the JSON: {"type": "question"}',
]
valid_types = {"order", "question", "complaint", "other"}
for r in replies:
try:
d = json.loads(r)
assert d["type"] in valid_types and isinstance(d["needs_human"], bool)
route = "PERSON" if d["needs_human"] else "AUTO-DRAFT"
except (json.JSONDecodeError, KeyError, AssertionError):
route = "PERSON (invalid format)"
print(route)
AUTO-DRAFT
PERSON
PERSON (invalid format)
Three replies, three outcomes: the clean order is auto-drafted, the complaint goes to a person, and the reply with a chatty introduction and a missing field is rejected and sent to a person. This small safety net is the difference between an automation you can trust and one that embarrasses you.
Choosing a platform
- Write the workflow on paper first: trigger, steps, AI step, action, review.
- List the apps involved and check the platform connects to them.
- Check the privacy terms and where data is processed.
- Try the free tier with test data.
- Compare cost per run if you expect volume.
- Keep a record of what is automated, who owns it and how to switch it off.
Connecting a live email account with permission to send and letting the AI reply to customers automatically. One misunderstood complaint and you have a public problem. Keep sending manual until you have months of evidence.
Forgetting that automations keep running while you are on holiday. Set alerts for errors, and review the logs weekly.
Time saved per run x runs per month x your hourly value, minus the cost of the tools and the review time. If the number is small, a simple template may serve you better than an automation.
Check what you learned
Five quick questions. Pick an answer and you will see straight away whether it is right and why.
What is the basic automation pattern?
- Question, answer, delete
- Install, update, restart
- Trigger, steps, action, review
- Copy, paste, forget
Something happens, then actions follow, with a human check.
A good first automation should…
- Prepare drafts for human review
- Send everything automatically
- Handle refunds alone
- Run without logs
Drafts let you check before consequences.
What should happen if the AI returns invalid output?
- Send it to the customer
- Route it to a person
- Ignore it
- Retry forever
A safety net protects customers.
Why do the task by hand several times first?
- To make it slower
- To avoid tools
- To impress someone
- To understand steps, exceptions and quality
You cannot automate what you do not understand.
Which data should you avoid sending to unapproved services?
- A made-up test row
- Personal and confidential data
- A public tagline
- A product list with no private data
Check privacy terms first.
Summary
- Automations connect a trigger to actions, with an optional AI step.
- Start with drafts and a human check.
- Validate AI output and route failures to a person.
- Protect data and watch cost and logs.
Questions
Do I need to code to automate?
No. No-code platforms handle many workflows. Coding gives more control and is covered in the next lesson.
Is it safe to connect my email?
Only with care. Use the minimum permissions, keep sending manual at first, and read the platform's privacy terms.
What if the platform changes?
Services change often. Keep your workflow documented so you can rebuild it elsewhere.
