AI-personalized cold email: the problem is not the writing

Models write cleaner cold email than most founders. They also invent the detail they personalize on, and a fluent email about something that never happened is worse than a generic one

Almost all advice about AI outreach is about tone. The giveaway phrases, the overused words, the opening line that every model reaches for. Real problems, fixable with an instruction, and less common with each release.

The failure that actually costs you something is not stylistic. Models are very good at writing and unreliable at knowing, and nothing in the output tells you which one you are looking at.

What that looks like in practice

We tested this on a real product: an ESG reporting tool selling to listed companies facing a mandatory climate disclosure deadline. Good conditions for personalized outreach, because the deadline is public and the affected companies are a knowable list.

The model returned three companies, each with a specific dated reason to get in touch, a source, and a drafted email. The companies were plausible. The founder's reaction was that the targeting looked broadly right.

Then we opened the sources.

• A large pharmaceutical group. The email referenced a Senior Sustainability Analyst role posted on a specific date, quoting the reporting standards from the job description. The link went to a real, well-known climate jobs board. The posting did not exist.
• A listed marketplace business. The email congratulated them on a sustainability report released that week. The link pointed at an investor relations address that does not resolve at all. Not a dead page: a domain that has never existed.
• An energy utility. The email cited an annual report with climate disclosures on specific page numbers. This one was closer to true, but the link went to the index of all annual reports rather than to the document, so the claim could not be checked without going hunting.

Three fluent, specific, confident emails. Every one of them cited something the model had not actually read.

How to spot a made-up citation

The three failures were not random, and once you have seen the pattern you can spot it in any tool's output.

• A link that does not resolve. Plenty of listed companies really do put investor material on an address shaped like the invented one. It is a familiar pattern applied to a company that does not use it.
• A real site with a page that is not there. The jobs board existed. The job did not, down to a convincing reference number. Every ingredient was true and only the combination was invented.
• The right site, the wrong page. A link to a permanent index rather than to the specific document. That is what a half-remembered link looks like: the page that is always there, instead of the one that appeared last week.

The uncomfortable part is that confidence rose with specificity. The most detailed claim, the one with page numbers, took longest to disprove.

Why this is worse than a generic email

A generic email costs you nothing but the send. The reader recognizes it in a second, deletes it, and forms no particular opinion of you.

A specific email about something that did not happen is a different transaction. It claims you did research, then proves you did not. In B2B that is often fatal, because the person reading it is an expert in exactly the subject you got wrong: a compliance lead knows their own reporting timeline, and a wrong date in the first line ends the conversation before the second.

The asymmetry is what makes this a real risk rather than an annoyance. The upside of a good cold email is a reply. The downside of a fabricated one is a person who now associates your company with carelessness, and there is no second attempt.

The fifteen second rule

Every claim in the email has a link, and you open the link before you send.

That is the whole practice, it takes about fifteen seconds per account, and it catches all three patterns above. It is also the step people skip, because the email looks finished and checking feels like a formality. It is not a formality. It is the only part of the process that distinguishes research from fluent guessing.

Three habits make it easier to keep:

• Prefer facts the company published on purpose. Job posts, announcements, filings. Easy to verify, and normal to reference in a first email.
• Drop rather than soften. If the source does not check out, remove the account. Rewriting the claim into something vaguer keeps the email and throws away the reason for sending it.
• Be suspicious of a full list. A tool that always returns exactly the number you asked for, however narrow your filter, is filling slots. Real research comes up short sometimes, and a tool that admits it is telling you something useful.

That last point is also the question worth asking a vendor. Not whether the tool uses AI or live data, which everything now claims, but what it does when it cannot verify something.

Where OctoLoops fits

We are writing this because we ran into it in our own product, on the first real run of our Outbound Loop, in front of the person we built it for.

What changed after that: the loop starts from the source rather than from the company. It looks for a recent, dated event, and the company is whoever that event names, so there is no step where a name gets picked first and a reason gets found afterwards. Every account arrives with the link the fact came from, and when it cannot find one it goes looking for a different company instead of filling the slot.

Sending stays with you, which is a design decision rather than a gap. Opening the source and sending from your own mailbox is the fifteen second check, and it is the only step that catches a bad citation before a prospect does.

If you want the method rather than the tool, what counts as a reason to reach out is the part that matters most, and it is free to read.