Suppose you ask an AI to reply to a client: “We cannot finish this week. Explain the delay and sound sincere.”
It writes: “We sincerely apologize and will deliver by Monday, ensuring your plans are not affected.”
You never said Monday was possible.
The AI supplies the sincerity. You supply the overtime.
This is less conspicuous than an invented citation. The message sounds reasonable, so you skim it and send. The client does not read Monday as a stylistic flourish. They put it in their calendar.
Approval has more than one object
When an assistant can send messages, writing begins to create consequences outside the editor. Permission to send matters. Meta's introduction of Muse explicitly describes approval before sensitive actions such as sending email or making a purchase.
But an approval interface also needs to make the content of a decision visible. Did I authorize the action, and did I notice what the message commits me to?
A promise can hide in the third paragraph of an otherwise acceptable email. A generic “please review” banner does not guarantee I will spot it.
I would like an assistant to distinguish “you may send this message” from “these are the commitments this message contains.”
Sincerity does not require a fictional date
The same problem appears outside work. “I cannot visit this weekend” becomes “I will definitely come next weekend.” The conversation feels resolved. Someone may start planning dinner.
If the next visit is undecided, a more faithful message is: “I am sorry I cannot make it this weekend. I will let you know when I have the next date.” That leaves an open question open.
The reverse mistake matters too. If I have committed to Friday, replacing it with “we will confirm a date later” removes useful certainty. The goal is not maximum vagueness. It is to preserve the decision the person actually made.
An assistant that knows why a delay happened can help explain it. An assistant that knows next week is unresolved can avoid filling the gap with a convenient promise.
Highlight what changed in the world
Editors are good at showing character-level changes. The more consequential difference may be semantic:
- A delivery date was added.
- An undecided invitation became an acceptance.
- A price appeared that was not in the original instruction.
Those descriptions tell me where a decision is required. They are more useful than making every added adjective equally prominent.
Not every polite phrase deserves a warning. Replacing “received” with “thanks, received” is different from committing someone's Friday evening. If everything looks equally consequential, the important changes disappear into the noise.
The product challenge is to identify when editing has crossed into deciding.
Look for the future tense
A practical review habit is to scan dates, prices, participation arrangements, and words such as “will,” “definitely,” and “guarantee.” For each, ask: did I decide this, or did the assistant add it to make the sentence feel complete?
This complements ordinary fact-checking. A sentence may contain no incorrect name or number and still lead another person to believe a new arrangement has been settled.
The examples here are hypothetical, but the design question is concrete: what will someone reasonably do after reading this message?
As agents become more capable, I want fewer invisible commitments inside a convenient button. A line that occupies almost no space in an email may occupy an entire evening in someone else's life.