One proposed action, one named button

Task 3 of the Mail2Flow test opened an urgent client request with two AI proposals next to each other: a reply to the client with two meeting slots, and a message asking a colleague to register the client’s questions. Under both sat three buttons, Approve, Consult partner and Mark done, with a line of grey help text. Five of seven got through it. One participant executed both proposals because she thought both were required, one read them as two questions from the client, one took the left proposal for the original e-mail, and one got stuck on the buttons. Most did not read the help text, and one person was afraid that clicking the draft would send it.

Task 5 showed one proposal, a junior’s prepared VAT workbook to review, and all seven completed it: read the summary, look at the work, approve or edit. The report’s recommendation is the pattern. One proposal at a time, or proposals that expand one by one, and one button per proposal that says what it does, Send this reply, Assign to Alex, instead of three abstract buttons at the bottom of the page.

Two cautions. The comparison is an observation, not a controlled one: task 5 came after task 3 for everyone, so people had learned the screen. And the pattern is only complete with a confirmation. One participant pressed Approve, saw nothing change, and did not know whether the reply had gone out.