Guide
AI SDR hallucinated emails: what goes wrong and how to catch it
A made-up fact in a cold email is worse than a boring one. Here are the ways it happens, and a check for each.
Short answer
Hallucination in a cold email means the email states something about the prospect that the writer never saw. It comes in a handful of types: invented facts, stale facts, the wrong company, the wrong person, invented numbers and customers, and praise that means nothing. The fix is the same for all: every claim about the prospect must be backed by a sentence on a page you actually fetched, and a person approves before it goes. Mailbox providers say messages should be accurate and not misleading [1].
Why it matters
A wrong claim does three kinds of damage. The recipient sees at once that nobody looked at them, so the email is deleted or reported. Google warns that "if messages from your domain are frequently reported as spam, future messages from you are more likely to be marked as spam" [2]. And a false statement under your name stays under your name: Google says message content should be "accurate" and not "misleading" [1], and in the US a subject line must be honest too: "The subject line must accurately reflect the content of the message." [3]
The catalogue
These examples are fictional. We built the list from how language models fail at this task and from the failures we fixed in our own writer; we cite no study, and nothing here is a statistic.
| Failure | Fictional example | Fix |
|---|---|---|
| 1. An invented fact | "I saw Harbour Lane Dental opened a second surgery in Leeds." Their site mentions one surgery. | Require a source sentence for every claim, and check that the sentence is on a page you fetched. |
| 2. A stale fact | "I saw you are running a summer open day." The page is two years old. | Prefer facts about how the business works (hours, services) over events; name recurring things, not dates. |
| 3. The wrong company | Two firms named Pennant. The email describes the other one's services. | Match on the domain you fetched, not the name; show the source page next to the email. |
| 4. The wrong person | "Hi Dr Alvarez" to an address that belongs to the receptionist. | Use a name only when the page ties it to that address; otherwise greet the team. |
| 5. Invented numbers and customers | "We helped 40 clinics cut no-shows by 30%." Nobody counted. | No figures in a first email unless they come from your own records. A first email does not need any. |
| 6. Meaningless praise | "I love what you are doing for families." The sender saw nothing. | Ban praise and hype words. State what you saw instead. |
| 7. The site's voice, not yours | "I saw we offer same-week appointments." "We" is the practice, not you. | Describe them as "they" or "you". Our check refuses an observation that speaks as the organisation. |
| 8. A misreading | The site says "we no longer offer evening clinics". The email builds on evening clinics. | Keep the claim close to the original sentence, and show that sentence to the person approving. |
| 9. An invented relationship | "As we discussed last week" or "your colleague suggested I write", when nobody did. | Allow a referral line only when a real reply from that person is on record. |
What a check can and cannot prove
A script can prove that a sentence appears on a page. It cannot prove the sentence is still true, that it is about the right entity, or that the email draws a fair conclusion from it. That is why the second line of defence is a person reading the evidence next to the email. See how to prevent sending without approval.
A checklist for any AI SDR tool
- Does every email show the page and sentence behind its main claim?
- Is a draft whose sentence cannot be found refused, rather than sent?
- Can you read sample emails before the campaign runs?
- Are praise, links, prices and numbers kept out of the first email?
- Can you stop the campaign at any time?
How Honest Intros does it
Honest Intros writes a first email in three parts: an observation that starts "I saw", who you are with one benefit, and an offer of a short 3-point outline. The observation must be backed by a sentence that really is on the prospect's own site, which is checked before the draft can reach you; drafts that fail are set aside with the reason. Other automatic refusals include hype words, links, prices, dates, exclamation marks, an observation phrased as a question, and an observation written in the site's own voice. Our check does not verify figures, so keep them out of a first email; the format has no place for them. You see sample emails next to their source sentences and approve the campaign before anything is sent.
Questions
What is an AI SDR hallucination?
A statement in a sales email that the AI wrote without having seen it, for example a fact about the prospect's business that is not on their website.
How do I stop an AI from inventing facts in cold email?
Require every claim about the prospect to come with a sentence from a page that was fetched, check the sentence is really there, and have a person read samples before sending.
Can a script prove an email is accurate?
No. A script can prove a sentence appears on a page. It cannot prove the sentence is current or well understood, so a person should still approve.
See what it would write for you
Paste your website. Honest Intros learns what you sell and shows who it would write to, and what it would say.
Free to start. Nothing is sent until you approve it.