article · cross business
Choose your first AI workflow
Choose a bounded first workflow using a completed synthetic brief, a rejected broader candidate, and explicit checks for ownership, data, and fallback.
The direct answer
Choose one repeated task with a clear output, approved sample inputs, a reviewer, and a manual fallback. Start with a draft or read-only result. Compare the proposal with a checklist or ordinary automation before choosing AI.
A completed workflow brief
Synthetic planning example: A service team considers preparing support-email drafts. Every entry below is an invented teaching assumption, not a report about a client.
| Brief field | Filled example |
|---|---|
| Task | Categorize guide-access questions and prepare a source-backed reply for review. |
| Trigger | A synthetic message enters a dedicated test mailbox. |
| Inputs | Stable message ID, message body, approved FAQ F01 revision 1. |
| Current steps | Read, check FAQ, draft, review, send manually. |
| Volume | Unknown; count eligible messages during the chosen baseline period. |
| Owner | Process owner and backup must be assigned before testing. |
| Desired output | One pending row with message ID, category, source ID, draft, and escalation reason. |
| Simpler alternative | A saved reply plus a manual category selection. |
| Permissions | Read test messages and write the review sheet; no sending access. |
| Baseline | Measure reading, lookup, drafting, checking, and correction minutes on comparable work. |
| Success criteria | No extra rows for duplicate fixtures; unsupported questions escalate; every draft reviewed against its source. |
| Stopping rule | Pause on unexpected access, duplicate writes, or a backlog the reviewer cannot clear. Resume the manual inbox process. |
This brief is complete as a planning example even though volume and named ownership remain unresolved. Those gaps block a live trial; they are not invitations to guess.
A rejected candidate
Synthetic rejected proposal: “Handle all customer requests, issue refunds, update account access, and reply automatically.” The inputs span conflicting policies, the actions have different consequences, and no single reviewer can check the whole process.
The alternative is to separate tasks and keep refunds and access decisions manual. The guide-question draft is a smaller candidate because its evidence and output can be inspected together. Smaller scope does not prove it works; it makes a meaningful test possible.
Compare candidates consistently
For each candidate, record frequency, output clarity, source readiness, review burden, action consequence, and recovery method. Do not simply total points. Missing permission or an absent owner blocks a trial even if volume looks attractive.
Observe the current task before estimating savings. Include awkward cases and interruptions, and use the same definition of eligible work when comparing manual and assisted handling. A workflow that moves work from drafting to correction has not eliminated that work.
Limits and non-example
“Improve sales” is a department ambition, not a workflow. “Prepare a draft follow-up from approved notes without inventing dates” names an input and a reviewable result. Avoid selecting a product first and expanding the task to use its features.
Make the decision
Complete a brief for two candidates. Choose the one whose inputs and correct outcomes you can demonstrate most clearly. Write why the other candidate was deferred and what would make it suitable. If neither has approved data, an owner, and a fallback, spend the next work session resolving those conditions.
Sources and scope
This is original editorial guidance with two synthetic planning examples. No implementation, time saving, or return is claimed.
Next step
Fill in your own workflow brief, then compare it with the support triage playbook. Suggest a correction.
Change notes
Foundation draft prepared October 20, 2026. No public publication or successful live-tool test is claimed.
Next learning step: One-page workflow brief