- This video shows how to test a workflow with an example, and what a test does and does not do.
Workflows
Test a workflow
Run your draft with an example, and know what a test does and does not do.
Email received
Sort the request: damaged item, delivery question, return or refund.
Draft a reply that follows the Returns policy.
- A test runs your draft with one example event. Open the menu beside Run workflow, and choose Change event.
- Pick an event that looks like real work, such as a customer email about a damaged order.
Choose an example event
Email received
Sort the request: damaged item, delivery question, return or refund.
Route by category.
Draft a reply that follows the Returns policy.
Sam approves the reply before it is sent.
Send the approved reply to the customer.
Customer requestsDraft
Go live
Discard
You're in draft mode, only read permissions are on by default
PermissionsChoose what this run is allowed to do for real.
Change event
Read
Nothing here reads from your connected integrations.
Write
Gmail · Create Draft
Gmail · Send Email
Slack · Send Message
Gmail · Create Draft
Gmail · Send Email
Slack · Send Message
Publish Events
Each step publishes an event that other jobs and workflows can listen for.
Auto-approved
A test run never waits for a person
Approve the reply
Always runs
This job executes on every run, including test runs
Sort the requestcalls GPT-5.4 Mini
Sort the requestcalls GPT-5.4 Mini
Select Test Event
Choose an event to use for testing this job configuration.
Cracked lamp shade
Maya Chen · order 4821
Where is my order
Jonas Berg · order 4790
Invoice question
Lena Ortiz · invoice 0932
Press Enter or double-click to selectCancelSelect Event
- Select Run workflow. Each job runs in turn, and its card shows how it went.
- Then open Logs, and select the run, to see what came in and what came out of each job.
- Check that the request is sorted correctly, and that the reply follows your policy.
Run the test
Read it in Logs
Check the result
Email received
Sort the request: damaged item, delivery question, return or refund.
Route by category.
Draft a reply that follows the Returns policy.
Sam approves the reply before it is sent.
Send the approved reply to the customer.
Not sent in a test
Customer requestsDraft
Go live
Discard
LogsLive Feed
StartedStatusSummaryDurationCost
26 Oct 2026, 10:22:41Test runSucceeded
All steps completedOutput event emitted
6.204s€0.0021
26 Oct 2026, 10:04:12Test runSucceeded
All steps completedOutput event emitted
5.718s€0.0019
Logs
Test resultsSucceeded26 Oct 2026, 10:22:41
Steps
Sort the request10:22:42Succeeded
Draft a reply10:22:44Succeeded
ExecutionInput & outputCost
Input eventRequest sorted · category damaged itemSorted correctly
Output eventHi Maya, I am sorry your lamp arrived damaged. A new shade for order 4821 is on its way.Follows the policy
- A test is safe by default, but it is not a pretend run.
- Read is on, so steps that look things up in your tools get real data. This workflow looks nothing up, so its list is empty.
- Write is off. The draft is not created, the reply is not sent, and nothing is posted to team chat.
- Write also covers webhook requests, database writes and changes to knowledge documents. With Write off, none of them happen for real.
- Publish Events is off, so events from the test stay inside the test.
- Approval steps do not wait for a person. A test run approves them automatically.
- Model calls always run. Here a model sorts the request. OAIZ calls this an LLM call, and it uses your model in every test.
- Turn on Write only when you are sure what the workflow will send.
Safe by default, but not pretend
You're in draft mode, only read permissions are on by default
PermissionsChoose what this run is allowed to do for real.
Change event
Read
Nothing here reads from your connected integrations.
Write
Gmail · Create Draft
Gmail · Send Email
Slack · Send Message
Gmail · Create Draft
Gmail · Send Email
Slack · Send Message
Publish Events
Each step publishes an event that other jobs and workflows can listen for.
Auto-approved
A test run never waits for a person
Approve the reply
Always runs
This job executes on every run, including test runs
Sort the requestcalls GPT-5.4 Mini
Sort the requestcalls GPT-5.4 Mini
Also under Write
Webhook requests
Database writes
Knowledge document changes
Only when you are sure what it will send
On
ReadReal data from your tools
Off
WriteMessages, webhooks, database writes, knowledge documents
Publish EventsEvents stay inside the test
Always runs
LLM callsThey use your model
Auto-approved
Approval stepsNobody is asked
- One good test is not enough. Try an ordinary case, then an unusual one. A message with no order number. A message in another language. An upset customer.
- When a result is wrong, change the instructions or the knowledge, and test again with the same event, so you can compare.
Try different cases
Cracked lamp shade
Ordinary · Maya Chen
No order number
Nora Lind
Message in German
Klara Vogel
Upset customer
Omar Haddad
Change the instructions or knowledge
Test again, same event
Compare
Same event
Before
Your new shade for order 4821 is on its way.
After
Could you send your order number?