Review · module 105 of 155
Review: from task to workflow and agent
Connect what you learned about workflows, agents, approval gates and measuring in one overview.
After this lesson
You design a small workflow with a clear stopping point, a human gate, a second route and a measuring point.
In view · Step by step
A workflow with an approval gate, a fallback and a measuring point
Input, AI step and result format are fixed; the agent knows its limits.
A human approves before anything is emailed, booked or paid.
A second route catches failures; you measure the total time, including corrections.
Part 1 of 3 · Fixed steps
Your preference for less movement is on. Choose each part yourself.
From one-off prompt to repeatable workflow
A workflow sets out what input is needed, which step AI carries out, in what format the result comes and where a person checks (lessons 9 and 65). An agent chooses some of its next steps itself and can use tools, such as searching or reading a calendar (lessons 29, 73 and 74). The more independence, the more important fixed limits become: which tools may it use, how many attempts does it get and when does it stop? An instruction on a web page it finds does not give it any new permission.
A proposal is not yet an action
Making a draft and carrying out an external action, such as emailing, booking or paying, are different steps (lessons 69 and 99). Set out who approves and what that person gets to see. In the event of a failure or time-out, first check the actual status before you repeat anything, so that you do not cause a duplicate action. Make sure there is a second route, for example manual handling, and test it in advance (lessons 90 and 91). Log errors without unnecessary personal data, so that you can see patterns (lessons 67 and 85).
Measure whether it really works better
When comparing, count all the steps: preparing, generating, checking and correcting, plus the one-off set-up and the costs (lessons 15, 66 and 94). A faster first draft with more correction work sometimes gains nothing. Test on the same, representative tasks and note your assumptions. Record your way of working on a workflow sheet with source, prompt, checks, limits and second route (lesson 100). That way you can repeat it, hand it over and improve it.
New practice example
From explanation to a usable result.
A workflow has fixed steps, a human approval gate before every real action, a fallback route for outages and a measurement point.
Fictional practice situationFictional: a bike repair shop wants to let customers know automatically that their bike is ready. The pick-up time and the price are in the work order software.
View a possible solution
This is how it can be done
Input: work order with status ready, price and pick-up times. Draft: AI writes a short message with only that data. Gate: for the first few weeks a staff member confirms every message. Action: send via the customer system. Check: delivery status. Fallback route: notify by phone. Measurement point: time per message and number of corrections.
Why this works
Each step has a clear input and outcome. The gate comes before the action and the fallback route ensures the customer gets a message even during an outage.
Make it slightly more difficult
After a month, no more corrections are needed. Which check do you keep anyway, and how would you decide to relax the gate?
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
Each step has a clear input and outcome. The gate comes before the action and the fallback route ensures the customer gets a message even during an outage.
A step further
After a month, no more corrections are needed. Which check do you keep anyway, and how would you decide to relax the gate?
Try & discover · no obligation
Put the steps in a workable order
Move the steps with the arrows. Check whether each step has its necessary input and control.
What does a step depend on?
- Set the goal and limits must precede Check the input and permission
- Check the input and permission must precede AI makes a proposal
- AI makes a proposal must precede A person approves
- A person approves must precede A separate function carries it out
- A separate function carries it out must precede Check the status and measure
Move a step and then check the sequence. The arrows also work with your keyboard.
This model checks dependencies, not duration, capacity, or actual execution. With multiple agents, the human decision remains a separate step.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
Using fictional data, design a workflow for answering general questions about opening hours. Use the steps check input, create draft, human approval, send and check outcome. For each step, give what is needed, who decides and when the workflow stops. Add a second route for a failure and one measuring point to see whether the workflow saves time. Do not carry out any action yourself.
Apply it
A workflow sends reminders automatically. After a malfunction, no one knows whether yesterday’s reminders went out.
Which approach best fits this situation?
- First check the sending status of each message and only resend what demonstrably did not go out.
- Send everything again to be sure.
- Ask the model whether the messages were sent and rely on its answer.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
An agent can send emails. Where does human approval belong in the workflow?
- After sending, to see whether the recipient is satisfied.
- Before the external action, based on the final text and recipients.
- Nowhere, if the prompt says the agent must be careful.
Extra practice
Draw the workflow from your prompt. Mark in colour the moment when a proposal becomes a real action, and note which measuring point you will look at after two weeks.
Input: only general questions, without personal data. Draft: AI writes an answer based on the approved opening hours. Gate: a staff member approves. Action: send via the ticket system. Check: check the sending status; in the event of a time-out, look at the status first. Second route: answer manually. Measuring point: total handling time per question, including corrections.
