Chapter BAI for emailPage 3 of 8

AI for email

Rewrite and clarify without distortion

A rewrite should change expression while preserving facts, intent, uncertainty, and responsibility.

~14 minEditing practice

Before you start

Why this matters

Rewrite this sentence in fewer words without changing its certainty: “We expect, but have not yet confirmed, that the replacement will arrive on Tuesday.” Now compare your version word by word. Did “expect” become “will”? Did “not yet confirmed” disappear? This small exercise shows why rewriting is an accuracy task as well as a style task.

1Learn the idea

Read

Rewriting is not merely polishing

Email assistance is often most useful after you have a rough draft. The model does not need to invent the message; it can help make your meaning easier to read. Yet every rewrite is also a transformation. Shortening can delete a condition. Softening can hide urgency. Making text “more confident” can turn an estimate into a promise. Making it “friendlier” can invent gratitude or familiarity.

Before rewriting, identify the invariants: details that must not change. Common invariants include names, numbers, dates, quoted language, ownership, requested action, uncertainty, and the difference between confirmed and proposed plans.

Then identify the editable layer: sentence length, ordering, headings, repetition, jargon, transitions, and tone. This separation gives the model freedom where freedom is helpful and limits it where accuracy matters.

Read

Use a preservation contract

A preservation contract is a short instruction that defines what the rewrite may and may not alter:

Rewrite for clarity and concision. Preserve all names, dates, numbers, conditions, quoted text, and open questions. Do not add facts, explanations, promises, apologies, or deadlines. If a sentence is ambiguous, flag it after the rewrite rather than guessing.

For a long or consequential email, ask for two outputs:

  1. the revised email;
  2. a change log listing deleted claims, combined ideas, unresolved ambiguity, and any wording that may have changed meaning.

The change log is not proof of correctness—the model can miss its own error—but it focuses your review.

Read

Clarify structure before wording

Many “writing problems” are actually structure problems. A useful email often follows this reading order:

  • purpose or decision needed;
  • essential context;
  • key facts or options;
  • requested action and deadline;
  • next step or contact.

Ask the model to reorganize around the reader’s task. For example: “Move the approval request to the opening, group the three risks under bullets, and end with the exact deadline. Preserve every risk and do not introduce a recommendation.”

This is stronger than “make it clearer” because clarity has observable criteria. The recipient should know why they received the message, what matters, and what to do next without rereading.

Read

Edit one dimension at a time

A rewrite is easier to control when each pass has one named purpose. Jargon, length, tone, and translation create different risks, so review them separately rather than combining them under “make this better.”

Reduce jargon without losing precision

AI can expand acronyms, replace abstract nouns, and split dense sentences. But simple language is not the same as vague language. “The system experienced an availability degradation” may become “The service was unavailable for 18 minutes,” if the 18-minute fact is supplied. Without that fact, the model should not invent specificity.

Use layered explanation when a technical term must remain:

Keep the term “data retention period,” then explain it in one short clause for a non-technical reader. Do not replace the policy term in the quoted requirement.

You can also request a jargon audit before rewriting. Ask the model to list terms that an outside reader might not understand, explain why each could confuse, and wait for your choice. This keeps you in control of domain language.

Shorten with priorities

“Cut this in half” encourages deletion without judgment. Tell the model what earns space:

  1. preserve the action, deadline, and decision criteria;
  2. preserve facts needed to evaluate the request;
  3. compress background the reader already knows;
  4. remove repetition and ceremonial openings;
  5. move optional detail below a divider or into an attachment.

Word limits can improve discipline, but the target should fit the task. A 50-word limit is inappropriate when a customer must understand a material service impact. If the draft cannot be shortened safely, the right result may be a brief email plus a linked document.

Soften tone without weakening the message

Tone repair is useful when a rushed draft sounds blaming or abrupt. Ask for neutral observations and concrete requests:

  • “You failed to send the file again” becomes “I have not received the file due Tuesday.”
  • “Obviously this won’t work” becomes “This option does not meet the stated capacity requirement.”
  • “ASAP” becomes an exact, justified deadline.

But do not erase accountability. “There was a mistake” can hide who must correct it. Neutral wording can still say, “Our team used the wrong rate, and we are recalculating the invoice.”

Request a tone diagnosis separately: “Identify phrases that may sound accusatory, dismissive, passive-aggressive, overly familiar, or uncertain. Explain the likely reading. Do not rewrite yet.” Diagnosis lets you decide which edges are intentional.

Translate carefully

Email translation combines language transfer with cultural and organizational context. Tell the model which locale, level of formality, and terms must remain unchanged. Ask it to flag idioms, humor, legal wording, and ambiguous pronouns rather than improvising.

For important communication, back-check the translated version with a qualified speaker. A fluent translation can still choose the wrong honorific, convert a date format ambiguously, or alter the force of “must,” “should,” and “may.”

Never assume translation makes sensitive data safer. The same privacy rules apply in every language.

Read

Compare meaning, not just style

After a rewrite, perform a sentence-level comparison:

  • Does every new factual claim exist in the source?
  • Did any condition, exception, or caveat disappear?
  • Is the requested action identical?
  • Are responsibility and agency still accurate?
  • Did an estimate become a commitment?
  • Did “may” become “will,” or “proposed” become “confirmed”?
  • Were names, dates, amounts, and links copied exactly?

For a high-risk message, ask the model to create a claim table pairing each revised claim with source wording, then verify it yourself. The model is an aid to comparison, not an independent witness.

Read

A practical rewrite prompt

Rewrite the email below for a busy external partner. Put the request first, use paragraphs under three sentences, and remove repeated background. Preserve the exact amount, dates, owners, uncertainty, and quoted contract sentence. Do not add an apology or imply approval. After the rewrite, list (a) anything removed, (b) ambiguities you did not resolve, and (c) sentences that deserve human verification.

This prompt defines audience, edits, invariants, prohibitions, and review output. It makes drift easier to see.

Checking tutor…

Continue learning · glossary & guides
  • What belongs in a preservation contract?
  • Why can shortening change truth, not just style?
  • How would you soften blame while retaining ownership?
  • Which modal-word changes should trigger verification?
  • Glossary: faithfulness · Cheatsheet: prompt recipe