Start with a complete piece of work and the outcome it should produce. Then ask where AI can improve it.
These notes are a starting point for a working session. They are intended to expose the questions a tool demonstration tends to skip.
1. Name the outcome
Describe the result from the perspective of the person receiving it.
For a customer support workflow, the outcome might be a resolved question. “An answer was generated” only describes an intermediate step.
Choose a boundary: where does the work begin, and what makes it complete?
2. Map the waiting
Write down each handoff, approval and search for missing information.
A task may take minutes while the complete workflow takes days. Faster generation does little for a process that spends most of its time waiting.
Find out why the waiting exists. Sometimes it protects a useful review. Sometimes nobody owns the next step.
3. Separate the kinds of work
Use three rough categories:
- Bounded actions: Clear inputs, clear rules and a result that can be checked.
- Recommendations: Useful proposals that need context and review.
- Judgment: Decisions where consequences, uncertainty or human context require accountable oversight.
These are design prompts, not permanent classifications. Review them when the workflow changes.
4. Specify the context
What information does the system need? Is it current? Can the organization legitimately use it for this purpose? Who can correct it?
More context is not automatically better. Relevant, understandable information is easier to maintain and inspect.
5. Design the exception
Before releasing a workflow, write down what happens when it cannot proceed confidently.
A useful exception path
The system identifies the uncertainty, preserves the relevant context and directs the work to a named owner. The owner can resolve the case and improve the underlying instructions.
An exception that disappears into a queue is unfinished work.
6. Measure the whole process
Compare the workflow before and after the change. Include elapsed time, correction effort, decision quality and the experience of the person receiving the output.
A faster draft is useful when it leads to a better finished result.
Start with a limited workflow and expand when the evidence supports it. Keep responsibility visible as the automation grows.
Continue with AI Should Change the Organization, Not Just the Tool Stack..