ISLE88 / AI IN PRACTICE / 4 MIN

Give the review step as much care as the prompt.

A repeatable AI workflow needs a way to notice when its output is wrong.

Choose a task you can inspect.

Start with a small task whose source and result can sit side by side. Turning a sanitized meeting note into a draft action list is easier to inspect than asking a system to run a project. Keep the person who knows the context in the loop.

Keep the source close.

Ask for a source reference next to each proposed action. An attractive summary is less useful when nobody can tell whether a deadline was stated, inferred, or invented. Mark uncertainty instead of filling the gaps.

Test the exceptions.

Try a note with no action, an ambiguous owner, a corrected date, and a quoted instruction that should remain quoted. Record what the tool did and what a person had to fix. A few well-chosen cases can expose the review burden before a team depends on the workflow.

Decide whether to continue.

Compare the time and effort needed to produce and check the draft with doing the task directly. Keep the workflow if the result helps. Change the task or stop if verification takes more effort than it saves. The useful outcome can be a clear reason not to automate.

Choose your subjects.

Occasional notes on design, development, AI, and learning. Confirm your email to subscribe.

Get Field Notes