AI email writing tools: how to check what they produce

AI email writing tools have solved the wrong half of the problem. Producing a well structured, correctly punctuated, appropriately brief sales email is now trivial and free. Producing one that is true about the person receiving it is exactly as hard as it ever was, and the tools have made it much easier to skip checking.

That is not a complaint about the models. It is a description of what they do. A language model is trained to produce the most plausible next sentence, and a plausible sentence about a company you have partial information on is, quite often, a sentence about a company that does not exist. The failure is not bad prose. It is a confident, well written, specific claim that nobody at your company can back up, sitting in a buyer's inbox with your name at the bottom.

Here is a four pass review that takes about ninety seconds on a draft, and can be run by anybody. The fourth pass is the one almost nobody does.

Fluent is not the same as true

Start by naming the specific failure, because "the AI hallucinates" is too vague to act on. In sales email there are four distinct versions, and they need different fixes.

• The invented event. A funding round, an acquisition, an office opening, a product launch that did not happen, or happened to a company with a similar name.
• The invented attribute. Headcount, revenue, tooling, "your team of forty engineers", "since you moved to Snowflake". Usually inferred from the industry rather than from anything about this account.
• The invented consensus. "Most companies your size are losing a double digit share of..." A statistic shaped sentence with no statistic behind it. This is the most common one by a wide margin and the one people are least likely to flag, because it does not feel like a claim about the prospect.
• The invented relationship. "Following up on our conversation", "as promised", "great to connect last week". Sometimes a template artefact, sometimes the model filling a slot.

Pass one: underline every assertion

Read the draft once and physically mark every sentence that asserts something checkable about the world. Not opinions, not questions, not offers of a meeting. Assertions.

On a normal four paragraph cold email you will mark between two and five sentences. That number is the surface area of your risk, and seeing it as a number is most of the value of this pass. A draft with seven marked sentences is not seven times better researched than a draft with two. It is seven opportunities to be wrong in writing.

Do this before you read for tone. If you read for tone first you will start editing, and once you are editing you will stop noticing claims.

Pass two: find the source for each one

For each underlined sentence, answer one question: where did this come from? Only three answers are acceptable.

1. A document, page or record you can open right now.
2. A note from a colleague who spoke to them, with a date on it.
3. Something the prospect themselves said or wrote.

"It is probably true" is not one of the three. Neither is "it is true of companies like them", which is the invented consensus wearing a coat.

Anything without a source gets one of three treatments, and choosing between them is the actual skill in outbound writing:

• Source it. Two minutes on their site is often enough, and the sentence gets stronger because you can now be specific rather than approximate.
• Turn it into a question. "Most fleets your size lose a double digit share of backhaul" becomes "How much of your backhaul currently goes out empty?" You lose the appearance of insight and gain a reason for them to reply, which is what you wanted anyway.
• Cut it. If it cannot be sourced and does not work as a question, it was decoration.

Pass three: read it as the recipient

Read the draft once as the person receiving it, at eight in the morning, on a phone, with forty other messages behind it. Three questions:

• Is it obvious in the first line why this arrived, and why now?
• Is there exactly one thing being asked for?
• Does any sentence claim to know something about me that would make me suspicious about how it was obtained?

That third question is underrated. There is a threshold where personalisation stops reading as attention and starts reading as surveillance, and models cross it happily because more specific looks more personalised. A reference to something the person published is attention. A reference to something you inferred about their internal processes is not.

Pass four: the mechanical checks

These are boring, they are checkable by a machine, and skipping them is what turns an ordinary email into a compliance problem or a spam complaint.

• A physical postal address is present.
• There is a way to opt out that requires nothing but an email address.
• The subject line describes what is actually inside, and implies no relationship that does not exist.
• The from name and reply address are real and monitored.
• Length is under roughly 125 words for a first touch, and the reading grade is somewhere around 6 to 9. Not because short is virtuous, but because long first emails get skimmed and skimming is where the reply dies.
• One link, at most, and no image-only content.

The pass nobody runs

It is pass two, and specifically the part where you check the sentence that sounds like a statistic.

People check the specific claims. If a draft says "you opened a warehouse in Reno in March", somebody will check that, because it is obviously a fact about the prospect. Almost nobody checks "most operators at your scale are leaving margin on the table in returns handling", because it is not a claim about the prospect, it is a claim about the world, and it feels like framing rather than assertion.

It is an assertion. It is the one a well informed buyer is most likely to push back on, because they know their own market better than you do, and being wrong about their market in the second sentence ends the conversation more reliably than being wrong about their headcount. If you check nothing else, check the sentence with the word "most" in it.

Six patterns that should always be checked

1. Any sentence containing "most", "typically", "usually" or "companies like yours".
2. Any number that did not come from a document you can open.
3. Any reference to a tool the prospect supposedly uses.
4. Any reference to a recent event, especially funding, hiring or expansion.
5. Any sentence implying prior contact.
6. Any comparison to a named competitor of theirs.

Those six cover, in our experience of reading generated drafts, nearly every claim that turns out to be unsupportable. They are also mechanically detectable, which is the point of the next section.

Doing it in ten seconds instead of ninety

Ninety seconds a draft is fine when you send four a day. It is not fine at forty, and at forty the review quietly stops happening, which is the moment the tool starts costing you more than it saves.

The fix is not to review faster. It is to make the tool do pass one and most of pass two itself, and to make the result visible rather than optional. That means the drafting step classifies its own sentences, ties each assertion to a source it actually read, and marks the ones it could not source. What is left for a person is the judgement call: source it, ask it, or cut it.

That is precisely what our ai email writer does, and it is why the demo shows a claim ledger next to the draft rather than the draft on its own. The mechanical checks in pass four run at the same time, which is the part described in detail on the spam score checker page.

Whether you use our tool or a text box and a highlighter, run the four passes. The draft that goes out with your name on it is your claim, not the model's.