What you will learn
- How to break a process into input, steps and result
- Where a human check is needed
- How to handle errors and missing responses
- How to avoid sending or publishing twice
Key terms
- Process
- A sequence of steps from input data to a result.
- Trigger
- An event that starts a process: a new email, a file or a time on a schedule.
- Idempotency
- A property of a step: running it again does not create duplicates. Important for payments and publications.
- Log
- A record of every run: what came in, what was done, where the error is.
01Break down the process
Choose a task that repeats at least once a week. Write down the input (what comes in), the steps (what is done with it) and the result (what should come out). Mark where a human decision is needed and which data must not be passed to outside services.
Example. Input: a draft article. Steps: AI checks the spelling, an editor approves, the system saves a version. Result: the approved text in the publication folder.
02Start with a safe copy
Do the first runs only on test data and in a copy of the working environment. Keep the source materials and the result of every step: that way you can analyse a mistake and repeat the work.
Important. Never test automation on live customer data or real payments.
03Plan for errors
Decide in advance what to do if a step does not respond, returns an error or an empty result: retry in a minute, stop and tell a person, or skip. Record every such decision in the log.
04Prevent double actions
Sending an email, publishing and making a payment must happen exactly once. If the result of the previous attempt is unknown, for example because the connection dropped, first find out what happened, and only then retry.
- Every operation has a unique number
- Before a retry, check whether the operation was already done
- Doubtful cases go to a person
05Test the whole route
Run three scenarios: a successful one, an empty input and the failure of one step. Make sure the result can be found, the error can be understood and the process can be stopped. For external publishing and operations with money, keep an explicit human confirmation.
Before going live
- The input, steps and result are described
- The places where a human decides are marked
- The process has been tested on test data
- Every error has a clear action
- A retry does not create duplicates
- There is a run log and a way to stop the process
Common mistakes and what to do instead
Automating straight on live data
Test data and a copy of the environment first.
Retrying a send without finding out the result
First check whether the operation was done.
Working without a log
Without a log you cannot find an error or repeat the work.
Taking the human out of decisions about money
Payments and publications only with explicit confirmation.
Try it yourself
Describe the process of preparing a draft and run it on one test example without external publishing.
- Source text: where it comes from
- AI processing: what exactly the model does
- Human check: who approves and by which rules
- Saving: where and under what name
Result. A process that is described and tested on a sample, and can be automated.
Test yourself
Answer in your head first, then reveal the answer.
1The connection dropped while an email was being sent. What should you do before retrying?Answer
Find out whether the email went out. A retry without checking can create a duplicate.
2Where in a process is a human needed?Answer
In decisions about money, external publications and disputed cases.
3Why do you need a run log?Answer
To understand what happened, find the error and safely repeat the work.
In short
Describe the input, steps and result, test on a copy, decide in advance what to do on errors, protect sends and payments from repeats and leave decisions about money and publications to a person.