Why generic AI-written outreach fails

You tried a tool. The message came back fluent, well structured, and completely interchangeable. Either nobody replied, or somebody replied to point out that you clearly had not looked at their website.

The writing quality was not the problem. The problem is that the message claims to know something about the reader that it does not know.

Our argument is narrow: generic AI outreach fails when it creates the appearance of relevance without doing the work required to understand the prospect.

The tell

Almost every failed AI outreach message contains a sentence performing familiarity it has not earned. An illustration, written for this article rather than taken from a real message:

I came across your company and was really impressed by the innovative work you are doing in the recruitment space. I had a specific thought rather than a generic pitch.

Read it as the recipient. What has actually been said about their business? Nothing. The sentence has the shape of research without any research in it. "A specific thought rather than a generic pitch" is doing the heaviest lifting in the paragraph, and it is a claim about the message rather than a fact about the reader.

Now a version that says only what is genuinely known:

Your site lists contract and permanent placement across two offices. I work with agencies of about that size on the admin between a vacancy landing and a shortlist going out.

Shorter. Less flattering. Checkable.

That last property is the one that matters. If the second version has the details wrong, the recipient can correct you, and a correction is a conversation. The first version cannot be wrong, because it never said anything. A message that cannot be contradicted also cannot be engaged with.

Why models produce it

This is our interpretation rather than a research finding, but it is a consistent one.

Ask a model to write a personal message and give it nothing personal, and it will still produce something. It cannot hand back an empty draft. So it supplies the form of personalisation: the appreciative opener, the vague industry compliment, the assurance that this is not a template. Those patterns are heavily represented in the text these systems learned from, because they are heavily represented in the outreach people already send.

The model is not lying. It is filling a shape you asked for with the only material it has.

The rule that fixes most of it

One rule from our own product's writing guide does most of the work:

Never pretend to know something Scout does not actually know about the prospect.

In practice: if there is no real detail, say less rather than invent more.

Most people reach for AI at the wrong moment. They ask it to make a message sound personal. The useful question is what you actually know about this company that is worth mentioning, and that is answered by research, not by generation.

The specific tells

These are the patterns we check every draft against. They work as a manual checklist too.

Generic openings. Fixed phrases that could precede any message to any company. "I hope this email finds you well." "I came across your company." They cost you the first line, which is the line most likely to be read at all.

Empty personalisation. A company name dropped into a sentence that would be identical without it. Merging a variable is not personalisation. Personalisation is a sentence that could not have been written about a different company.

Clichés. Circle back. Touch base. Leverage. Synergy. Reach out. At your earliest convenience. Low-hanging fruit. Move the needle. Each one signals a production line.

Stiff phrasing. Kindly. Please do not hesitate. To whom it may concern. Pursuant to. This reads as a form letter, which is the opposite of a founder writing to someone.

Repetition. The same run of four or five words appearing twice in one short message. It happens constantly in generated text and it reads as padding, because it is.

Overlong paragraphs. We flag anything past roughly 320 characters, about two to three short sentences. Cold outreach is read on a phone, between other things.

What a machine can fix, and what it cannot

This is the part that matters more than the checklist.

When we built automated checking into our own product, the rules split into two groups, and the split turned out to be the interesting result.

Some things are safe to fix mechanically. An em dash becomes a comma. A known banned phrase is an exact string and can be removed. A duplicated signature block after the sign-off can only be duplication, so it can be cut. These are fixed patterns, and correcting them cannot change what the message means.

Some things can only be flagged. Take a cliché sitting mid-sentence. Delete the word and you usually break the grammar around it. Fixing it properly means understanding what the sentence was trying to say and writing it again, which is judgement. So the check flags it, shows the exact phrase, and leaves it to a person.

That boundary is not a shortcoming we are apologising for. It is the real line between work a machine can take off your hands and work that merely looks mechanical. A machine can remove an exact pattern. It cannot tell you whether a claim is true. Which is why a person approves every message before it goes anywhere.

The honest limit

We would rather state this than have you find it.

The mechanical phrase removal is best-effort string surgery, not a language model. On text that already reads naturally it never needs to trigger. On heavily edited or adversarial text, stripping a banned phrase can occasionally leave a short grammatical fragment behind.

It is a defence net, not a guarantee of publish-ready prose. The human review step is what actually catches this, by design rather than by accident.

If you are writing by hand

None of this needs a tool.

1. Before writing, list what you know about this company and where each fact came from.
2. Write only from that list. If the list is thin, the message is short.
3. Delete any sentence that would be equally true about a different company.
4. Read the first line alone. If it could open a message to anyone, rewrite it.
5. Check that nothing claims familiarity you have not earned.

That is roughly what a good outreach check performs, and it takes two minutes.

How we use this in Scout

Scout drafts outreach for founder-led B2B service firms and runs every draft against the rules above before saving it. Fixed patterns are corrected, judgement calls are flagged with the exact phrase for you to rewrite, and the finished draft lands in Gmail as a draft. Scout does not send anything.

One limitation worth naming. Scout writes from a defined founder voice, not from yours. It does not study your previous emails or adapt to your personal writing style, so expect to edit for voice as well as for facts. We would rather set that expectation now than let "writes in your voice" sit unchallenged in your head.

Scout is in early access, and this article is the method it is built on.

Explore Scout

None of the above is an argument that AI cannot write good outreach, or that all AI-written outreach fails. What we have described is a misuse, not a property of the technology. Used after the research rather than instead of it, these tools are genuinely useful. We have offered no response-rate or conversion figures here, because we have none that would survive scrutiny, and a piece about manufactured credibility is a poor place to start manufacturing some.