A useful automation needs an owner, a reviewable result and a route for exceptions, as well as a working demo.
Choose a task with a visible outcome
Start with something people can describe and review, such as turning an enquiry into a callback brief. Record how often it happens, where time is spent and who checks the result. A vague goal like ‘use more AI’ is difficult to evaluate.
- Name the trigger, inputs and desired output.
- Record the manual starting point before automating.
- Choose an owner who can decide what counts as useful.
Separate assistance from authority
Drafting a response and sending it to a customer are different actions. Extracting an amount and approving a payment are different decisions. Decide where AI can suggest, where deterministic rules apply and where a person must approve.
- Write the actions the workflow may perform.
- Keep high-impact actions behind an explicit review step.
- Design a pause switch and a manual fallback.
Test the awkward cases
Use examples with missing fields, duplicate requests, ambiguous wording and unavailable tools. Check whether the system can say it is uncertain and give the owner enough context to finish the task.
- Define acceptance criteria before judging the demo.
- Check quality and operating cost, not only speed.
- Retest important cases after changing prompts, tools or models.
Give the release an operating plan
A workflow needs somewhere to show failures, retry limits and a person who receives an issue. Store the minimum useful information, restrict access and agree how long it is kept. A small controlled release gives you better evidence than a broad unattended rollout.
Illustrative example
An enquiry assistant that prepares a callback brief
A demo summarises incoming requests well, but sometimes omits the requested callback time. Staff still need to check the original message before acting.
- The decision
- Keep the assistant in draft mode. Mark missing details explicitly and route every brief to the office owner before any customer commitment.
- The check
- Try a complete request, one without a callback time and a duplicate. A reviewer should see the original context, missing details and whether the request was already handled.
Put it to work
Leave the conversation with a useful first step.
Write a one-page workflow brief: trigger, inputs, output, owner, approval point, three failure cases and one baseline measure. If these are hard to name, start with consulting rather than a build.
A one-page workflow brief covering the trigger, inputs, permitted actions, reviewer and manual fallback.
What would make this workflow ready for a controlled pilot?
Free editable notes for your team. No email needed.
Prefer to prepare it yourself?
Automation readiness workbook
You have several automation ideas and need to choose one. Review an illustrative example and the requirements before choosing this starting point.
Explore the toolkit preview ↗When you need support
What should your AI prototype do reliably?
Share the task, what already works and where you need help moving beyond the demo. Keep credentials and confidential data out of the enquiry.
Review ai automation builders ↗