← Blog

Sell ·

How to Review an AI-Written Sales Message Before It Goes Out

An eight-check review for any AI-written sales draft, with weak and better examples, a send, edit or drop rule, and a prompt that flags invented claims.

By , Founder, Baljia AI

The short answer

Whether it came from an AI SDR, an AI sales agent or a chat window, review the draft in this order: is it true, is it for this person, is it about them, does it sound like you, is the ask small, would you stand behind it, does it fit the sequence, and is it allowed. Truth comes first, because a false claim goes out under your name and no amount of polish fixes it.

For a short message, the whole review takes about two minutes. If the first two checks fail, stop reading and drop the message or the person.

Every example below is invented. The seller is a made-up payroll service for roofing contractors, and the buyer is a made-up owner named Dana. No result here is real. The weak and better lines are fragments, not finished emails. Sent as a cold email, any of them would still need check 8's ad line, postal address and opt-out.

The checklist

Keep this next to your approval queue.

  • True. Every name, number, customer and claim can be shown.
  • This person. There is a real reason it went to them, this month.
  • About them. It talks about their problem more than your company.
  • Your voice. You would say this out loud to a person.
  • One small ask. One question, answerable in a line.
  • Defensible. You would be fine seeing it in a screenshot.
  • In sequence. It still makes sense after the last message, and after a reply.
  • Allowed. They are not on your do-not-contact list, and cold email says it is an ad, with your postal address and an opt-out.

Check 1: Is every fact true?

AI writes confident sentences whether or not it has facts. Highlight every name, number, customer, date and feature in the draft. Each one needs a source you can open, or it goes.

Weak (invented):

We helped over 200 roofing companies cut payroll errors by 43%.

Better (invented):

We run payroll for roofing crews, including overtime across job sites.

The weak line may sound stronger, but nobody can back it up. The better line is plain and true. If you have a real result, you can use it, as long as you can show it and the customer agreed. One customer's result is also not a promise to the next one.

Watch for these inventions in particular:

  • A past that never happened, such as "following up on our chat".
  • A wrong job title or an old employer.
  • Customers you do not have.
  • Features on your roadmap, stated as live.
  • Percentages that came from nowhere.

Check 2: Why this person, why now?

Finish this sentence: "I am writing to you because ____." If the only answer is their job title, you have a list problem, not a writing problem.

Weak (invented):

As a leader in the roofing industry, you know payroll can be a challenge.

Better (invented):

You posted last week that two crews came up short on overtime after the storm jobs.

The weak line fits every roofer in America. The better line could only go to Dana, and it points at a public fact she can recognize. Check the link yourself, because a model can also invent the post.

If no real reason exists, drop the person. Editing the words will not help, and a compliment about their company is not a reason. Our one-day exercise shows how to find real ones.

Check 3: Is it about them?

Count the words "we", "our" and "us". Then count "you" and "your". If the first number wins, rewrite.

Weak (invented):

We are a full-service payroll provider. Our team has years of experience, and our platform automates tax filing.

Better (invented):

Adding up overtime across two job sites by hand is slow. Is that what you are doing now?

The reader does not care about your company yet. They care about their week.

Check 4: Would you say it out loud?

Read the draft aloud, once. Wherever you stumble, or feel a little silly, rewrite that part. Models lean on a stock of polite, empty phrases.

Phrase to cutWhat to do instead
"Hope you're having a great week"Delete it and start with the reason
"I noticed your impressive work"Say what you actually saw
"I wanted to reach out"You did. Say why
"tailored solutions"Name the thing you do
"optimize your workflow"Name the task and who does it
"Would love to connect"Ask a real question

Weak (invented):

We offer tailored, end-to-end payroll solutions for your whole back office.

Better (invented):

We do the payroll, file the taxes and text you when it is done.

Concrete nouns also make a message shorter. As a house rule, we keep a first email under 80 words and a LinkedIn note under 50. Those limits are our preference, not research.

Check 5: Is there one small ask?

A stranger owes you nothing. Asking for half an hour is a big request, and three questions at once means none get answered.

Weak (invented):

Do you have 30 minutes for a demo this week? Also, who handles payroll there? And can I send over a deck?

Better (invented):

Is overtime the part that goes wrong, or is it something else?

Dana can answer that in three words while standing on a ladder. If she does, you have a conversation, and a call is an easy next step.

Check 6: Would you stand behind it?

Picture the message in a screenshot, posted with your name on it. Would you be comfortable?

This check catches messages that are true but rude, too familiar, or built on guesses about the reader.

Weak (invented):

Your bookkeeper is clearly costing you money.

Better (invented):

If your bookkeeper already has this covered, ignore me.

The weak line accuses someone you have never met. The better line gives Dana an easy way out, and it may be the reason she keeps reading. Also check for anything a model might guess about a person's health, age, family or finances. Leave all of that out.

Check 7: Does it fit the sequence?

AI drafts each follow-up as if it were new. Read them in order, as the recipient would.

Weak (invented):

Just bumping this to the top of your inbox!

Better (invented):

One thing I left out: we also handle crews who work in two states. Does that apply to you?

A follow-up has to bring something new, or it is only a nag. Then ask three more questions. Does it repeat the first message? Does it contradict it? And has Dana replied since the last one? If she has, it should not go at all. The tool should stop that step, and you should test that it does.

Check 8: Is it allowed?

This takes ten seconds and it is the check people skip.

  • The person is not on your do-not-contact list.
  • Commercial email has your physical postal address and a clear opt-out.
  • Email is identified as an ad, unless the recipient gave prior affirmative consent.
  • The price and offer match what you actually sell today.

Our safety post covers the rest of the risks, including LinkedIn's terms.

Send, edit or drop

Most reviews end in one of three decisions.

What you foundDecision
Passes all eightSend
A wrong fact you can correctEdit, then recheck it
The false fact was the whole pointDrop the message
No real reason for this personDrop the person
Stale, or they already repliedDrop the message
Wordy, or more than one askEdit
The same flaw across the batchStop and fix the instructions
On the do-not-contact listDrop, and find out how it got in

Pay attention to that "same flaw" row. If you keep fixing the same thing by hand, the problem is upstream, in the instructions or the facts the tool was given.

A full before and after

Both versions are invented.

Before:

Subject: Quick question

Hi Dana! Hope you're having a great week. I noticed your impressive work at Ridge Line Roofing and wanted to reach out. We are a full-service payroll provider that has helped over 200 roofing companies cut errors by 43%. Our tailored solutions cover your whole back office. Do you have 30 minutes this week for a demo?

After:

Subject: Overtime across job sites

Hi Dana, you posted last week that two crews came up short on overtime after the storm jobs. We run payroll for roofing crews, and that overtime math is most of what we do. Is that the part that goes wrong, or is it something else?

[Your name]

This is a sales email from [business name]. [Postal address]. Reply "stop" and I will not write again.

What changed: the invented number went, along with the flattery. The reason moved to the first line. Two asks became one question. The email now says it is a sales email and carries your address and an opt-out.

A prompt that finds invented claims

Use a second AI pass to check facts instead of rewriting them. Paste this into any assistant:

Below are my facts and a draft message.

FACTS: [what is true about us, and about this person, with links]
DRAFT: [the draft]

Do not rewrite the draft. List, in order:
1. Every claim in the draft not supported by FACTS.
2. Every detail about the recipient with no link.
3. Every sentence that would fit a different recipient.
4. Every question the draft asks.

If you are unsure about something, write "unknown". Do not add new facts.

This helps, but it does not replace reading. The checker is also AI, so it can miss things. It is most useful when you have a long queue.

How much to read when volume grows

You cannot read everything forever, so use a rule. Ours, and it is a judgment call: read every draft in the first batch. After that, keep reading every first message to a new type of buyer, and sample at least one in five of the rest. Go back to reading everything when the list, the offer or the instructions change.

Before you let any type of message run alone, see the mode-by-message table in our send modes guide.

Where Baljia fits

Baljia prepares email and LinkedIn outreach and holds what you ask to review, so this checklist is what we would have you run on its drafts too. It can get facts wrong like any AI, and reading before you approve is the protection. You can approve each message or switch on what may run by itself. LinkedIn campaigns come with Growth and Scale. More detail is on the AI sales outreach page.

Frequently asked questions

How do I check an AI-written email before sending it?

Check facts first, then the reason you are writing to this person, then the wording. Highlight every claim and confirm it. Read the message aloud. Keep one small question at the end, and confirm the email says it is an ad and has your address and an opt-out.

Should I tell people that AI wrote my message?

Nothing in the CAN-SPAM basics we read turns on who drafted an email. You are responsible for what it says either way. If someone asks, tell the truth, and never claim personal research that nobody did.

How long should a cold sales message be?

Short enough to read on a phone in one go. Our house limits are 80 words for a first email and 50 for a LinkedIn note. A clear reason and one question matter more than the exact count.

How do I make AI messages sound human?

Give it real facts and a real reason to write, then read the result aloud and cut anything you would not say. Most robotic writing comes from having nothing specific to say, so better inputs help more than a tone instruction.

What if the AI keeps making the same mistake?

Fix the instructions, not the drafts. Add a short list of what is true and what must never be claimed, and give examples of your own writing. Then read the next full batch before you trust the fix.

See the drafts yourself

To run this checklist on drafts for your own business, try Baljia with a 7-day free trial, no credit card, at https://baljia.ai/. See pricing for what each plan includes. The trial does not include LinkedIn outreach.

Sources