Automation and AI solve different parts of a task
Imagine a team receiving routine enquiries. A rule can check whether a required field exists or route a request by a known category. An AI tool might help summarise free text or propose a reply, but its suggestion is not the same as a verified fact or an authorised action.
The first question is whether the current process is clear enough to improve. Mapping who receives a request, which information is needed and who owns the decision often reveals an unnecessary step before any new tool is chosen. In administrative or operational work, a simple reliable rule can be preferable to an impressive demonstration nobody can explain.
Automate a reply without losing oversight.
A small business receives repeated order-update events. A workflow could draft a reply, but lookups can fail and the same event can arrive twice.
Your role: Add checks before a message could be sent, then test normal, failed and repeated events.
What you’ll learn: Separate useful automation from uncontrolled side effects.
No account needed. Nothing is sent or saved. You can restart at any point.
Read the full guided story without the interactive controls
You can follow the same lesson here without JavaScript.
See the repetitive task
Run an order lookup to see what information a reply needs.
What happens and why: The event identifies an order. A lookup retrieves the details; a reply should not invent missing information. The model sends no real messages.
Put a person before the send action
Add human approval between preparing the reply and sending it, then test a normal event.
What happens and why: The workflow prepares a draft and waits. An approval gate requires a person’s review before the side effect of sending.
Give failures and repeats somewhere to go
Add a review queue for failed lookups and a record of processed event IDs. Test both a failed lookup and a repeated event.
What happens and why: Failures go to a visible review queue. Repeated event IDs are skipped rather than producing a second reply. Neither failure nor duplication should silently become a customer message.
Run the complete workflow
Process a fresh valid event with the safeguards in place.
What happens and why: The normal event reaches a draft awaiting approval; failure reaches review and a repeat is skipped. This deterministic rule-based model does not use AI inference or send mail.
Takeaway: Useful automation needs a normal route, a human approval boundary and explicit failure/retry behaviour. Automating a task does not remove responsibility for its effects.
Where this is useful: Useful when designing order updates, document approvals and administrative workflows.
A fictional educational example, not a confirmed TWA assignment, assessment or learner outcome. Simplified to explore one idea; real situations need context and authorised procedures.
Optional deeper exploration: open the original technical workspace
Want to explore the technical model?
The guided story above introduces the idea. This original workspace lets you try actions in different orders. It has its own reset and does not inherit your story choices.
- Event
- A signal that something happened, such as an order changing.
- Workflow
- An ordered set of actions and decisions.
- Approval gate
- A required review before an action proceeds.
- Deduplication
- Recognising repeated event IDs so an action is not performed twice.
Approval is part of the workflow.
Learning objective: Route side effects and exceptions deliberately.
A fictional workflow drafts a customer reply after an order event. Duplicate events and failed lookups can occur.
Evidence desk
- Event
- Order update received
- Side effect
- Send customer reply
- Exception
- Order lookup fails or event is duplicated
Your workspace
Worked explanation & takeaway — available without JavaScript
Require human approval before sending, route lookup failures to a visible review queue, and deduplicate event IDs. An exception must not silently become a sent reply. This model simulates routing only; no email or external automation runs.
Fictional educational model. No live system, personal assessment or confirmed TWA assignment. All actions run locally in your browser.
A draft is not a decision
Choose how the fictional workflow should handle two kinds of request.
- Receive
Collect a sample request with no sensitive information.
- Suggest
Prepare a category or draft using approved reference material.
- Review
A person checks facts, permissions and tone.
- Act
Only the approved action proceeds; exceptions stay visible.
Prepare a draft and show the source to the reviewer. The reviewer decides whether it is accurate before sending.
Read both scenarios and explanations
Routine question with an approved answer
Prepare a draft and show the source to the reviewer. The reviewer decides whether it is accurate before sending.
Request for an unconfirmed fee
Do not invent a price. Flag the missing approved information and route the request to TWA for confirmation.
An educational illustration using fictional situations. It is not a confirmed TWA assignment, live system, assessment or promise of a particular tool.
Describe the workflow in plain language
Start with a trigger, the information it supplies, a decision and an action. Microsoft’s cloud-flow documentation uses triggers and actions to explain automation. Add the parts a diagram can easily omit: permission to access information, what happens when a step fails and who reviews a consequential output.
TWA’s existing direction covers AI literacy, prompting, checking outputs, workflow mapping, tool connections, human review and responsible use. A particular automation platform, paid account or integration is not confirmed. Practise describing an input and an expected output before trying to connect real client systems.
Keep approval where the consequences matter
For a fictional training-enquiry workflow, an AI draft could suggest a learning area. A person should check it against the approved programme information before a reply is sent. A request about an unknown fee should be escalated, not answered by an invented figure. A failed connection should leave a visible item for follow-up rather than quietly discard the enquiry.
Use sample data to test normal, ambiguous and missing-information cases. Write down where a person can override a recommendation and how to avoid sending the same reply twice. NIST’s AI risk guidance provides industry context for evaluating AI systems; it does not certify this example or guarantee that a workflow is safe.
A reviewed enquiry-assistance workflow
Map a fictional enquiry process, prototype a draft step with sample inputs and keep sending under human control.
What you could document
- A current-process and proposed-process map
- An input/output specification with review points
- A small test log including ambiguous and failed cases
- An exception-handling and handover note
What makes the explanation useful?
The work should explain what is automated, what remains a human decision and how an error becomes visible. A time-saving claim needs evidence rather than an assumed percentage.
The learning plan determines actual activities and assessment. These deliverables are examples, not contractual curriculum outcomes.
Progress from a useful task to an understandable system
A foundation route can focus on asking clearer questions and recognising unreliable outputs. Deeper workflow work can consider exception handling, access, monitoring and maintainable documentation. Building model-backed applications is a related but different direction covered by AI Engineering.
Bring one routine process and its current pain point to TWA. Explain what you are permitted to automate, what information should remain private and who needs to approve the result. This helps define a realistic learning discussion without promising a ready-made business integration or a fixed three-level syllabus.