The 20-Second Checklist for Reviewing an AI Email Draft

The useful review takes seconds and follows a fixed order. When people say AI drafts are mostly fine and still get burned, the miss is almost always in the same place: a commitment the model invented, a name spelled wrong, a figure rounded, or the actual question buried under two paragraphs of preamble.

This is that routine. Six checks, in an order that matters, and one prompt you can paste into the model to make the checks faster still. It runs on any drafting tool because the failure modes are shared across them.

We build AI Emaily, an AI email client that keeps sends approval-gated by default and logs every action with undo. A few of the checks below are already lined up on our approval screen. Most are not tool-specific and work the same way in ChatGPT, Gemini in Gmail, or Copilot in Outlook.

The short answer#

Read the draft in one pass with six things in your head, in this order: facts, commitments, names and numbers, the single ask, register, attachments. Twenty seconds if the thread was routine, forty if it was long.

Then, before you send anything consequential, ask the model to list its own assumptions and mark each one either from the thread or inferred. Everything marked inferred is a place a human error usually lives.

Do this in the order given. Facts first, because a factual miss makes the register question moot. Attachments last, because you can only check them once the prose is settled.

Why the review order matters more than the checklist length#

A long checklist read in the wrong order does not catch anything the short one misses. You will start at the top, get bored, and skim. Order beats length: an ordered five is stronger than an unordered fifteen, because the top items are the ones that carry legal and commercial weight.

The two failures that actually cost you money are invented commitments and wrong numbers. Everything else — a stiff opening, a paragraph too long, a comma splice — is embarrassment, not exposure. Treat the two as different problems and check the exposure first.

NIST's AI Risk Management Framework names confabulation and misinformation as first-order generative-AI risks precisely because they hide behind fluency. The better the prose reads, the less carefully you check it. Ordering the review is how you fight that.

The six checks, in the order to run them#

Each check is one question and one place to look. Run them in this sequence on every draft, even the short ones. Especially the short ones — a short reply that agrees to the wrong date is the classic failure.

Check 1: facts, and only against the thread#

The rule is narrow. A claim is fine only if you can point to the message in this thread that supports it. Not something you remember from a call, not something the model probably read on the internet — a line in the mail.

This is where the model's default helpfulness turns into risk. Asked to reply about a proposal, it will summarise what the proposal seemed to say, in the tense of a decision already made. Delete anything you cannot underline in the thread.

Check 2: commitments get the veto#

Invented commitments are their own category. Treat any promise as suspect until you can point to the message where you agreed to it, and treat any deadline the model volunteered as suspect even if the recipient asked for one.

Fluency is the trap. A clumsy draft that says I will get back to you costs ninety seconds to edit. A polished draft that says I will have it to you by Friday costs a renegotiation, and it reads exactly like the nine you already approved, so you sign it.

Check 3: names and numbers, character by character#

Read proper nouns and figures as strings, not as words. The eye slides past a familiar name and a rounded figure. Compare the draft's version to the thread's version one character at a time.

Currency is the common miss. A number rounded to the nearest thousand for readability is a different sum from the one the client quoted, and it will be the version quoted back to you if there is a dispute.

Check 4: the single ask is the point of the reply#

Almost every incoming message has one thing it needs an answer to. The good reply answers it in the first two sentences. The AI reply often answers it in sentence four, after a paragraph that restates the situation for no one.

Reread the last incoming message and mark the ask before you read the draft. Then ask whether the draft addresses it directly and early, or whether it circles.

Check 5: register is a one-line test#

You do not need to reread the whole draft to check register. The opening line and the sign-off carry the tone signal, and they are where mismatch shows first.

The failure to watch for is over-brightness on hard threads. A reply to a complaint that opens with I hope you are doing well reads as tin ear even if the rest is careful. Fix the two lines and the middle usually holds.

Check 6: attachments last#

Attachments come last because they can only be checked once you know what the prose refers to. If the draft says the revised proposal is attached, open the tray and check the file is there, is named what the sentence claims, and is the current version.

This is the check that gets skipped the most because the prose is already read and the send button is one click away. Skip it and you send the old deck to the person you least want to send the old deck to.

A scoring table for the review, not the tool#

If you want to know whether the checklist is paying off, run it for a working week and score each draft on the six checks as you review. Not to grade the model — to see which checks catch things and which are noise for your inbox.

The weights say what to fix first when time is short. A score below 2 on either triple-weighted check means rewrite, not edit. A score below 2 on the doubles means change and reread. A score below 2 on the singles means send and note.

The self-audit prompt that shrinks the check#

Most drafting tools will tell on themselves if you ask. Paste this prompt into the same chat you generated the draft in, before you send:

You are not asking the model to fix itself — that produces a second draft with a second set of inferred lines. You are asking it to expose its own workings so your review is a check against a labelled list, not a hunt through prose.

Everything labelled inferred goes through checks 1, 2 and 3 first. Everything from thread you can trust as far as the quoted line supports it.

A worked example#

A supplier writes with two questions and a soft ask for a discount. The AI drafts a friendly, capable reply. Here is the twenty seconds of review, in order.

Four fixes in about forty seconds. Two of them — the discount that was never offered and the wrong month on the last order — are the kind of miss that ends up in a written record you cannot walk back. The other four are grade-and-polish edits you would make on your own writing too.

Red flags that mean stop, not edit#

Some patterns mean the draft is not salvageable inside a review. Recognising them saves the ten minutes you would spend polishing a reply that was going to embarrass you anyway.

• The draft agrees to something the thread did not settle. Rewrite the paragraph, do not delete the sentence.
• The draft uses a figure or date that is not in the thread at all, and you cannot source it in ten seconds. Do not send.
• The draft answers a question that was not asked. This usually means the model summarised the thread wrong, and the whole reply is off-target.
• The draft opens with an apology you did not intend to make. Apologies are commitments in disguise.
• The draft addresses the wrong person. In a thread with two contacts, a model will sometimes swap them.
• The draft references an attachment that is not in the tray. Either attach it or rewrite. Never send with a hanging reference.

What we'd pick, and where the honesty is#

This is a decision-guide about a review routine, not a tool ranking. The checklist above works in any drafting tool. The pick is what workflow you build around it.

We build AI Emaily. In Copilot mode, sends are approval-gated by default, and every action lands in an audit log with undo. The approval screen is where the six checks happen, so the routine and the product are the same shape — you review before send because there is no other way to send. Voice is shaped by a Personal Context you write plus per-client profiles you set, which is why our drafts are less likely to invent a tone but require you to write the Context in the first place. AI Emaily is a Pro subscription with a 7-day free trial (card required, cancel by day seven for no charge) — the pricing page has the current numbers.

If your workflow is copy-and-paste with ChatGPT and it fits your day, the checklist is exactly as useful without us. You are already doing the review outside any client, which is fine — the six checks do not care where they run.

Where we'd point you elsewhere: if you live entirely inside Outlook and the review you want is one that Microsoft has folded into the compose window, Copilot in Outlook does that natively and the compose experience is coherent. Microsoft's own Copilot in Outlook FAQ tells users to review generated content before sending, which is the same instruction as this page — the difference is only where the review sits.

The point of the placement: use the checklist regardless. The tool is the container. The routine is the safety.

How to build the routine into a habit#

A checklist read once from a blog post is a checklist you have forgotten by Thursday. Two small moves make it stick.

1. 1

   Pin the six words above your compose window

   Facts, commitments, names, ask, register, attachments. A sticky note, a text expander, a browser sidebar. You are training the eye to run the checks without deciding to.

2. 2

   Use the self-audit prompt only when it counts

   Not on every draft — the ceremony wears out. Use it on any reply that involves money, dates, scope, or a decision you have not made yet. Those are the drafts where an inferred line becomes a bill.

3. 3

   Score for one week, then stop

   Run the weighted table for five working days. You will find one or two checks that catch nothing in your inbox and one that catches something every day. Adjust the weights to fit your work.

4. 4

   Keep the failures

   When a draft slips through and causes a problem, save it in a folder called Missed. Rereading the folder in a month is the fastest way to tune the routine, and it beats guessing at where AI drafts go wrong in general.