AI NEWSROOMNews · guides · analysis
Search
← Home
AI at Work

The sentence you have to delete shows when AI should not draft first

Look at the sentence you have to delete from an email draft — a wrong name, an invented date, a promise nobody agreed — to see which mail AI should not draft first.

Beginner4 min read1 source

A reply to a customer is already sitting there as a draft. Before you send it, you have to delete a person’s name, a delivery date, and a sentence that promises a price. The example below was constructed by the editors to show how to read an edit trail. It is not a real email, and it is not a result that was run on a specific tool.

The draft says:

Hi Minh, thank you for waiting. We will deliver the design on Friday and keep the price we already agreed. Please take a look and confirm.

You change it before sending, because the person in the thread is Lan, nobody has agreed Friday, and the price is still being checked internally:

Hi Lan, thank you for waiting. I am sending the design back for you to look at. I will confirm the delivery date and the price once the team has agreed.

The three deletions are not spelling mistakes. Each one is a fact the draft did not get from the open thread: the wrong recipient, an added date, and an added commitment about price.

Read the deleted sentence, do not add up minutes

For work email you still have to answer for before you send it, keep the edit trail instead of building a table that compares minutes of writing it yourself with minutes of letting AI write it. Each message needs only three lines:

  • the sentence or phrase you deleted or rewrote;
  • what kind of error it was: wrong person, wrong date or time, an added commitment, a wrong figure, or only a change in wording;
  • where the correct fact lives: an older thread, a calendar, a quote, or only in your head so far.

A template that is enough:

Message: reply to a customer about progress
Deleted: "deliver on Friday" — added a date that is not in the thread
Deleted: "keep the price" — added a commitment that has not been agreed
Kept: the thank-you, because only the wording was off

After a few messages, look at the kind of error, not the number of minutes. If most of the edits only adjust how you address someone, the draft is close to what you meant to say. If the tool keeps adding a date, a price, or a promise that was not in the material you gave it, typing speed is no longer good news: your time is going into removing sentences that should not be sent.

You do not need ten messages to see the pattern. Three messages of the same kind, with the same kind of deletion, are enough to stop letting AI draft first for that group.

An OpenAI article also treats the edit as the signal

On 14 September 2026, OpenAI published an article about Fyxer, a tool that drafts and replies to email. OpenAI writes that when a user edits a draft before sending, the difference between the original and the edited version shows which sentences they wanted. OpenAI also cites a figure supplied by Fyxer: 53% of AI-generated drafts are accepted unchanged. That is a vendor figure inside an OpenAI article, not an independent measurement for every inbox. What you can take for your own check is narrower: whether a draft is kept as-is or edited is a signal; the number of drafts produced is not.

You do not need a training system like Fyxer. The three-line edit trail above already shows which error is repeating.

The deleted sentence decides the next time

Take the most recent message you had to rewrite before sending. If the rewrite only made the sentence easier to read, you can keep letting the tool draft that kind of message, and you still read it before you send. If the rewrite removes a name, a date, an amount, or a promise, next time write those facts yourself first, and ask the tool to adjust the wording only after the facts are yours. If the thread does not contain that fact, do not ask the tool to guess it.

Sources