Editing an AI-Drafted LinkedIn Opener: A Checklist Before You Send

Why the unedited draft is costing you replies

I'm Hlib Storchak. I build outbound systems for B2B founders and sales teams, and I've booked 2000+ meetings for B2B clients doing it, which means I've read a lot of AI-drafted LinkedIn openers before they went out, and a lot more after they'd already been sent and quietly underperformed. The pattern is almost always the same: someone turns on an AI drafting tool, the first few messages look fine on a screen, and three weeks later acceptance has drifted down and nobody can point to the exact message that caused it. There usually isn't one. It's the accumulated effect of sending a draft instead of a message.

That's not a hunch. Expandi's State of LinkedIn Outreach H2 2026 report ran the comparison properly, controlling for the fact that stronger operators tend to be the ones who adopt AI tools in the first place, and the corrected number says human-written messages beat AI-generated ones on acceptance by about 12% within the same accounts. I went through why that gap exists and what the report's naive comparison got wrong in the case against AI-personalized LinkedIn outreach. This piece is the other half of that argument: not why the gap exists, but the actual edit that closes it, step by step, on the message sitting in your drafting tool right now.

The 12% gap, and what's actually inside it

The study behind this is specific enough to take seriously: 118 campaigns from 73 accounts, October 2025 through June 2026, each account's own AI-generated campaigns measured against that same account's own human-written campaigns, so the operator, the targeting, and the list quality all stay constant and only the writer changes. Human-written copy won acceptance by roughly 12%. Reply rate came out effectively identical either way, which tells you the gap is at the top of the funnel, in whether someone accepts the connection at all, not in whether they eventually reply once they're in your network.

Expandi's own read on why is worth quoting directly, because it's the whole thesis of this article in one line: "AI-generated copy may read as personalized to the sender but feel formulaic to recipients who've encountered the pattern before." The report's own recommendation is not to drop AI drafting, it's to "use AI to generate a starting point, then rewrite with your own knowledge of the prospect's situation, industry, and pain points." That's the checklist below, just expanded into something you can actually run against a message.

What to keep from the AI draft

Before the checklist, it's worth being precise about what isn't broken. A drafting tool is genuinely fast at surfacing a prospect's recent post, a job change, a company announcement, or a shared connection, the research legwork that used to take a rep two or three minutes per lead. LinkedIn's own Message Assist, still in public beta as I covered in LinkedIn's AI Sales Assistant, still in beta, is explicitly framed by LinkedIn itself as a co-pilot, not a finished send, which matches what the data above says. Keep the research. Keep the first pass as a starting point. The edit below is about the sentence that actually gets sent, not about throwing out the tool that got you to a first draft in fifteen seconds instead of two minutes.

The checklist: seven edits before you hit send

Run these in order. Most drafts only need two or three of them to stop reading as AI-generated, but checking all seven takes under two minutes once you've done it a handful of times.

1. Cut the compliment-then-pivot

The single most recognizable shape in an AI-drafted opener is a compliment about the prospect's profile, post, or company, followed immediately by a pivot into the pitch: "Loved your recent post on X. That got me thinking about Y, which is exactly what we help with." It's not wrong information, it's the structure that's the problem. A prospect who's active on LinkedIn has seen this exact shape dozens of times this year alone, from dozens of different AI tools drafting for dozens of different reps. Reorder it, or drop the compliment and open with the observation itself stated plainly, without the "I noticed" or "I saw" framing that announces you're following a script.

2. Kill the tell phrases and formulas

Some specific phrasing has become recognizable enough that it now actively works against the sender. This is better documented for LinkedIn posts than for cold DMs, but the underlying pattern-matching a reader does doesn't care which format it shows up in. A 2026 analysis of AI writing patterns on LinkedIn found the long em dash mid-sentence jumped from under 2% of posts to 15.6% as AI drafting spread, and that four specific structural moves, a "stop X, start Y" advice frame, a "here's how" opener, a manufactured "reveal" mid-message, and the "it's not X, it's Y" contrast formula, each carry a measurable reach or trust penalty once a reader has seen the pattern enough times. None of that is unique to public posts. If your draft leans on any of those four moves, or on stock openers like "I hope this finds you well" or "I wanted to reach out because," cut them. They cost you nothing to remove and they're doing active damage left in.

3. Swap generic praise for one verifiable detail

"Impressive growth at your company" could be sent to anyone at any company. "Your Series A close in Q2 and the two sales hires you posted about since" could only be sent to this one person. The fix isn't to write more, it's to replace one vague compliment with one specific, checkable fact the AI tool actually pulled from the prospect's activity, then delete the rest of the flattery around it. This is the single highest-leverage edit on the list because specificity is the one thing a model genuinely can't fake without you feeding it something real to work with, and it's the fastest tell a reader uses to separate a template from an actual message about them.

4. Vary the sentence shapes

Read the draft's sentences by length alone, ignoring the words. AI drafts tend to default to a narrow band of medium-length sentences with the same subject-verb-object shape, one after another, which is part of why a fluent draft can still feel flat. Break that up deliberately: one short sentence, one longer one, maybe a fragment. It reads more like how a person actually types a quick message, because that's exactly what it is once you've done this.

5. Add a stake or a genuine question, not a compliment

The mistake underneath most of the tells above is treating the opener as a place to flatter rather than a place to ask something real. Replace the closing compliment with a specific, answerable question tied to the detail you kept in edit three, something like asking whether the new hires are focused on outbound or an existing book of business. A real question signals you actually want an answer, not just a click on "accept." Flattery gets a polite ignore. A specific question is what gets a reply once someone's already said yes to connecting.

6. Read it out loud before you send it

This catches what a silent read misses almost every time. Phrasing that looks grammatically fine on screen often sounds stilted the moment you say it, because written AI fluency and spoken naturalness aren't the same test. If a sentence makes you stumble or sounds like something you'd never actually say to this person on a call, rewrite it. This step takes fifteen seconds and it's the one people skip most often, usually right before the message that reads the most obviously templated.

7. Time-box the edit, don't skip it

Here's the actual math, with the assumptions stated so you can swap in your own numbers. Assume a drafting tool produces a first pass in about 15 seconds, and this checklist takes roughly 90 seconds once you're used to it, call it under two minutes total per message. Assume you're sending 40 connection requests a day and the 12% acceptance gap holds. At a baseline 25% acceptance rate, unedited AI drafts land you roughly 10 acceptances a day; edited ones land closer to 11 to 12. Over a month of sending days, that's the difference between roughly 200 and roughly 220 to 240 accepted connections, for a cost of about an hour of total extra editing time across the whole month. Run your own baseline acceptance rate and volume through the same formula. The point isn't the exact number, it's that skipping a two-minute edit to save time is close to the worst trade available on this list.

Before and after, side by side

A concrete example makes the checklist easier to apply than a list of rules on its own.

The split I run for clients

The mistake I see most often when I take over a LinkedIn seat that's been running an AI-drafting tool for a while isn't a disaster, it's a slow fade: acceptance drifts down gradually over a few weeks and nobody can point to a single bad message that caused it. The fix that actually works, and the one I run for clients now, is splitting the job in two: AI does the research and the first pass, a person runs the checklist above on every message before it goes out, not just the first batch when the tool is new and everyone's paying attention. The same discipline applies whether the draft came from a LinkedIn tool or from an email sequencer's AI writer, Salesforge's own draft-assist included, which is the stack I default to for cold email clients. Treat any AI draft as version zero. Never treat it as the version you send.

When I skip the AI draft entirely

For a short list of named, high-value accounts, fewer than twenty or so, I usually skip the AI draft altogether and write the opener from scratch, because the research time you save on a small list is smaller than the risk of a templated line landing on the one account that matters most this quarter. AI drafting earns its keep on volume, a hundred-plus messages a week where the research time savings compound and a person still edits every single one before it sends. It's the wrong tool for a target list small enough that each message is effectively a one-off.

Proving the edit actually worked

Don't take the checklist on faith any more than you should take Expandi's number on faith. Split a real segment of your own list, run the same targeting through an unedited AI draft on one half and the checklist-edited version on the other, and let both run for two to three weeks before reading the result, since a fresh sequence swings a lot in its first few days. I go into the fuller version of that testing setup, plus what else actually moves acceptance and reply on LinkedIn beyond the writer, in LinkedIn outreach metrics that actually matter. If you're already running Expandi or a comparable tool to pull your own before-and-after numbers, I've laid out how it actually compares against HeyReach for LinkedIn outreach if you're deciding between the two.

FAQ