AI writing tells kill B2B reply rates

Outbound that sounds like every other model dies in the inbox.

On Oct 1, 2026, TechCrunch covered Graphite research on Opus 5.5 overusing phrases like "this matters," plus other AI writing tells. Labs already trained models to cut em dashes. New tells showed up anyway. Buyers can smell the pattern even when they cannot name the model.

A note that sounds like every other model does not get a reply.

What are AI writing tells in outbound?

Patterns that show up across vendors because the same models draft the same polish.

Common tells in B2B notes right now:

1. "This matters" / "Here's why it matters" as filler with no evidence.
2. Em dash addiction in older model output (labs are killing it; watch for the replacement filler).
3. Symmetric three-beat lists that say nothing specific about the account.
4. Hollow intensifiers. Seamless, robust, innovative, game-changing. Buyers skim right past them.
5. Generic empathy openers. "I know you're busy." Everyone knows. Nobody cares.
6. Company-name mad libs. Swap the logo and the note still works. That is the failure.

I do not need a classifier to spot this. If the note could have gone to anyone in the ICP, it will go to the archive.

Why do polished AI notes underperform?

Because cold email that books a meeting is specific, not pretty.

Buyers skim for a reason that applies to them. Signal-based outbound gives you that reason: a pricing-page visit, a G2 compare, a job post, a champion change. The note has to name the signal in plain words. Intent-based outbound messaging is the same rule with a sharper score.

A model that only knows your ICP description will write ICP mush. A model that gets the signal, the proof point, and the ask will still need a human pass for voice. Skip the pass and you ship tells at scale.

How do you rewrite an AI tell into a signal-matched note?

Start from the signal. Cut the polish. End with one ask.

Bad (tell-heavy): "Hope you're doing well. I know you're busy building an innovative GTM motion. This matters because teams like yours need a seamless way to unlock pipeline. Would love to share how we help."

Better (signal-matched): "Saw Acme on the Vector pricing page twice this week. We help teams turn those de-anon visits into a weekly meeting list, not another dashboard. Open to a 15-minute teardown of your last ten visits?"

Same length. One named signal. One mechanism. One ask. No "this matters." No hollow intensifiers. That is the shape of a first-touch note that matches the signal.

What should GTM teams do this week?

Audit the last 50 sends for tells. Then change the draft stack.

1. Highlight every sentence that would still work with a competitor's name. Delete or rewrite.
2. Ban the Graphite-style fillers in your prompt and style guide. Add "this matters," "hope you're doing well," and empty three-beat lists to the kill list.
3. Require a signal field in every draft. No signal, no send.
4. Human-pass the first line and the ask. Models can draft the middle. You own the claim and the close.
5. Score reply and meeting rate by template family. If a polished template underperforms a shorter signal note, kill the polish.

Adapt or fail. Specific beats pretty. Pretty without a signal is spam with better grammar.

Start Signals, Convert, Grow

FAQ

What did the Graphite study find?

Per TechCrunch on Oct 1, 2026, Graphite research flagged Opus 5.5 overusing phrases such as "this matters," among other AI writing tells. Treat that as a warning that buyers can detect the pattern, even when the model changes.

Are em dashes still an AI tell?

They were a loud one. Labs are suppressing them. New tells replace old ones. Do not use em dashes in your outbound either. Periods and commas are enough.

Can I still use AI to draft outbound?

Yes. Feed it the signal, the proof, and your voice rules. Then edit. AI without a signal and a human pass is how tells scale.

How short should the note be?

Short enough that the signal and the ask fit above the fold on a phone. If you need three paragraphs of context, you do not have a signal. You have a brochure.

Does this apply to LinkedIn and email the same way?

Yes. Same tells, same fix. Name the signal. Cut the filler. One ask.