Short answer
AI writing tools make sales messages smoother by turning facts into stronger claims: "they are hiring" becomes "you are struggling with capacity". To stop it, keep the fact, the guess and the unknown apart, and put the source next to every claim. Turn every guess into a question, and count the time it takes to check a draft as part of its cost.
4 ways AI upgrades a fact
Watch what happens to a fact on its way into an AI-written email. Each sentence gets smoother. Each one also gets less true.
| What was true | What the AI wrote | What to write instead |
|---|---|---|
| They are hiring 3 engineers. | You are struggling with capacity. | I saw you are hiring 3 engineers. Is that for a new project, or to keep up with today's work? |
| They plan to open a second warehouse. | You have launched a second warehouse. | I read you plan to open a second warehouse. How will you plan stock across both? |
| They downloaded a guide. | You are evaluating a replacement. | Thanks for reading the guide. Was it useful, or is this research for later? |
| We discussed pricing. | We agreed on pricing. | We discussed pricing. Here are the two options we talked about. |
The buyer notices, because they know what they said and what they did. One upgraded fact tells them the message was written about someone like them, not about them.
Why it happens
- Language models are good at fluent, confident text. A hedge like "might" or "plans to" reads as weak writing, so the model smooths it away.
- Summaries compress, and compression drops conditions: "if", "next quarter", "someone raised a concern".
- The sender was in the meeting, so "agreed" reads as close enough to "discussed". The buyer remembers it differently.
- Reviewers skim for tone, not for truth. A guess is easy to spot. A promise nobody made reads like a record of the meeting.
4 habits that stop it
These come from building AI agents for pharma quality teams at Leucine. There, every claim had to trace back to a source record, or nobody could use it. Selling deserves the same discipline.
- Keep the fact, the guess and the unknown in separate places. A hiring post can justify a question. It can't justify a diagnosis.
- Put the source next to every claim, so the person approving the message can check it in 30 seconds.
- Turn every guess into a question. The doubt has to survive into the sentence the buyer reads, not only into a note for the reviewer.
- Count the checking time. If the draft took 10 seconds and the review took 15 minutes, you didn't save anything.
The best messages ask a narrow question the buyer can answer honestly, including with "no, that's not us".
A 30-second review
Before you approve an AI-written message, read each claim and ask:
- Is there a source for it?
- Does the source say exactly this, or something weaker?
- Is any verb stronger than the source? Watch for agreed, launched, struggling, evaluating and decided.
- Is a guess written as a statement?
- Could the buyer reply "that's not true"?
If any answer is wrong, cut the claim or turn it into a question.
How Mach handles it
Mach shows a source for every fact it writes with, so you can open the page and read the words behind a claim. A fixed formula sets fit and intent, and the brief gives the reasons against. When there's too little to go on, Mach says so. And you approve each message before it goes out, by default. See how research works and what to look for in an AI SDR.