AI Workflows Need Review Before Autonomy
2026-06-09 · 4 min read · Janaina Maia
Apple’s most useful AI announcement today was not another assistant promise. It was the quieter idea that people should be able to describe a workflow in plain language and have the system build the automation for them.
TechCrunch reported that Apple is adding AI-powered workflow creation to Shortcuts, its visual scripting app for automating tasks across apps and system actions. Instead of manually choosing actions, variables, triggers, and app steps, a user can describe what they want. Apple’s example was simple: when I leave work, tell my partner I am on the way and include my estimated arrival time.
That sounds small, but it is a real product shift. The hard part of automation has never only been whether the machine can execute steps. The hard part is whether the human can express intent clearly enough, inspect what the system understood, and trust the result before it starts acting on their behalf.
Natural language is a better entry point, not a full governance model.
Prompting makes workflow creation more approachable because it removes the blank-canvas problem. A person does not need to know which app action to pick or how variables work. They can start with the outcome they want.
But in enterprise environments, “describe it and let AI build it” is not enough. Workflows touch documents, customer records, approvals, messages, calendars, payments, engineering models, and audit-sensitive decisions. A generated workflow needs to show its interpretation back to the user before it runs.
The design question becomes: how do we turn a sentence into a safe operating plan?
The review surface matters more than the prompt box.
If an AI builds a workflow, users need to see the trigger, the data sources, the actions, the permission level, the failure mode, and the moment where a human stays in control. The interface should make the automation legible, not magical.
For product teams, this means the real design work sits between intention and execution. The system should ask: “Is this what you meant?” It should show: “Here is what will happen.” It should make it easy to edit, pause, test, undo, and narrow the scope.
That is especially important for agentic products, where the AI is not only answering a question but coordinating tasks. The more capable the agent becomes, the more visible its boundaries need to be.
Design implications.
- Translate intent into steps: show the generated workflow as a readable plan, not only as hidden automation logic.
- Separate draft from execution: creating a workflow, approving it, and letting it run should be distinct moments.
- Expose permissions early: users should know which systems, files, people, or records the workflow can touch.
- Design for correction: the user should be able to edit the AI’s interpretation in plain language and in structured controls.
- Make failure visible: every useful workflow needs a clear fallback when data is missing, the action is risky, or confidence is low.
My take: the next wave of AI UX will not be won by the biggest prompt box. It will be won by products that can turn messy human intent into visible, reviewable, accountable action.
Autonomy is not the first screen. Review is.