PRAPI Research · 2026-06-16
AI can draft your pitch but it can't send it: what gets a pitch through, what gets it deleted in three seconds, and why the polish is now the tell
We asked operators who pitch for a living, and who get pitched all day, a blunt question: when you use AI to write a pitch, cold email, or source-request answer, what actually earns a reply, and what gets ignored, filtered, or flagged? The answer was close to unanimous, and it inverts the promise of the tools. AI is a first-draft engine, not a send button. Raw AI output gets reply rates near zero, and the operators who win rewrite 60 to 70 percent of every draft before it goes out. Several tracked the gap precisely: one logged 58 percent reply rates on fully manual pitches against 22 percent on AI-first ones; another watched response rates fall from 40 percent to 5. The thing AI cannot supply is the only thing that earns a reply: a specific proof the recipient cannot get anywhere else, tied to their actual work. The sharper finding is on the receiving side. The people being pitched now spot AI in three to ten seconds, and the polish itself is the giveaway: flawless grammar with no voice, generic buzzwords, "catalog blindness" that proves the sender never read you. In an inbox full of machine-clean drafts, sounding like AI is now a negative signal. This report is the field guide both sides wrote: the tells that get you deleted, the specifics that get you through, and the hidden cost of letting the tool hit send.
24 contributors cited
Every operator who pitches now has the same tool open in another tab. So we asked the people who use it, and the people who get pitched by it, a plain question: what does an AI-written pitch actually do? What earns a reply, and what gets filtered, flagged, or deleted on sight?
The answer came back with unusual consensus, and it is not the one the tools promise. AI drafts the pitch. It cannot send it. Everything below is a force inside that one finding.
Force 1 - First draft yes, send button never
The universal workflow: AI for structure and speed, a human for the 60 to 70 percent that earns the reply. Raw drafts die.
Pavankumar Kamat, Co-Founder of Panto AI, put the division of labor exactly: "I use GPT to generate 6 to 8 angle variations from a short brief, then a human edits, typically 40 to 60 percent of the draft. The AI supplies structure and phrasing; we supply credibility. AI-assisted and humanized pitches that cite a clear metric or a named customer get responses at roughly 12 to 18 percent to journalists." Runbo Li, CEO of Magic Hour AI (Connectively), reduced it to a line: "Every pitch I send starts as an AI draft. Every single one. But the ones that land have something no model can generate: a detail so specific it proves I actually care about the person reading it. AI writes the bones, you write the blood." His numbers: 35 to 40 percent reply with the rewrite pass, 8 to 12 without it.
The operators who skipped the rewrite watched it cost them. Nam Dang of Cricket One ran the A/B: "When we tested lightly edited AI drafts against heavily rewritten ones, the lightly edited version saved time but got about 35 percent fewer replies. The message was clean, logical, and completely forgettable." Deepak Shukla of Pearl Lemon: "We tested completely AI-generated outreach against heavily edited AI outreach. The latter always wins because it includes opinions, concrete examples, and sometimes even errors that add personality." Supported by Nikita Baksheev/Ronas IT (Connectively, "every pitch needs one sentence AI couldn't know without our internal experience"), Eric Turney/The Monterey Company (Connectively, AI output is "structurally clean but contextually thin"), Tharun B., Firdaus Syazwani, and Arsen Misakyan.
Force 2 - The tells that get you deleted in three seconds
The receiving side is where the report turns. The people being pitched spot AI almost instantly, and the giveaway is no longer a typo. It is the polish.
Joe Spisak, CEO of Fulfill.com (Connectively), gets about 40 AI pitches a week: "I can spot them in three seconds. They all start with 'I came across your work on' followed by a generic compliment that proves they never looked at what we do. Last month someone pitched me warehouse automation software for our fulfillment operations. We're a marketplace. We don't operate warehouses. Dead giveaway." That is the dominant tell, and Ana Martinez, Head of Growth at Marqo, named it best: "the tell isn't vocabulary, it's catalog blindness. We train personalized models on retailer catalogs; anyone pitching us generic semantic-search solutions clearly ran our homepage through a summarizer." Damien Zouaoui of Oakwell Beer Spa (Connectively) corroborated from a different category: "the fastest tell is category-blindness. If a pitch calls us a spa or a brewery and never acknowledges we are neither, the sender didn't spend five minutes on our site. AI fails on category definitions; humans anchor to one true detail."
The vocabulary tells are real too: John DeMarchi of Social Czars ("dead on arrival if it contains 'I hope this email finds you well' or ChatGPT's favorite words like 'testament' and 'delve'"), Stephen Taormino of CC&A ("the pitch has structure but no judgment"). But the structural tell is the deeper one. Sahil Agrawal of Qubit Capital (Connectively): "what gives an AI draft away is that every sentence is equally polished and nothing is left hanging. A real person trails off, overstates one thing, forgets to explain another. You can sand a draft until it sounds like nobody." Natalia Lavrenenko of Smarfle (Connectively): "the mess is the signal. Real PR pros write less tidy paragraphs." And Runbo Li flagged the single phrase that now auto-filters: "I had an AI draft that used 'at the intersection of' and almost sent it. That phrase is the scarlet letter of AI content right now. The journalist later told me they auto-filter any pitch containing it."
Force 3 - What gets through: a proof only you have, tied to their work
If the tell is generic polish, the cure is a specific that a model could not invent.
Christopher Coussons of Visionary Marketing (Connectively) named the mechanism: "the line that gets a pitch picked up is a real number from a real account, a named mistake I made, an opinion an editor can quote. A model will not invent that I tanked a client's traffic by 34 percent one quarter, because it has no stake in being honest. I do, so I say it, and that line is usually the one that gets used." The proof points across the pool are concrete: Samuel Huang ("a pitch that named my $0.15 CPM campaign for Thai restaurants earned the reply"), Magee Clegg of Cleartail Marketing ("real outreach names a published outcome, like the 5,000 percent ROI on a Google Ads campaign, and ties it to a concrete next step"), Rob Dietz ("'check your robots.txt for GPTBot, ClaudeBot, OAI-SearchBot' beats 'AI is transforming search'"), and Mikhail Drozdov (Connectively, "AI can improve structure but cannot create genuine expertise; editors respond to firsthand experience and data that isn't already published everywhere"). The portable rule, from Nikita Baksheev: "every pitch needs one sentence AI couldn't know without our internal experience. If that sentence isn't there, the pitch isn't ready."
Force 4 - The hidden cost, and the war stories
Two costs recur: the editing that eats the time savings, and the placements a robotic draft burns outright.
Thomas Oldham of WebMotion Media lost a $10,000 automotive placement: "that editor told me it lacked genuine insight. He was right. I tested ChatGPT on one batch and my response rate dropped from 40 percent to 5. I spent hours editing every draft to sound like me. Those hours cost more than the tool saved." Ruth Cruz of Wolf King USA watched a $50,000 partnership vanish "because the CEO's pitch sounded like a ChatGPT template; the investor called it out in a 10-second email," and tracked the gap herself: 58 percent reply on fully manual, 22 on AI-first, across 210 pitches. The most quoted war story is Joe Spisak's: "I let an AI draft go out to a Forbes contributor with barely any editing. It referenced 'the evolving landscape of e-commerce logistics' in the first line. She replied 'This is clearly AI-written and I don't cover vendor pitches.' Burned that relationship completely." And the most ironic, from Kevin Lourd of Distribute.you (Connectively), who builds AI outbound infrastructure: "we cost ourselves a major PR placement because an AI drafted a follow-up so relentlessly polite and perfectly structured that the journalist assumed we were a bot and blocked the domain. We now run hard-coded scripts on our own servers to strip AI noise before it hits processing. The slightly unpolished drafts are the only ones that get a reply." The paradox belongs to Deepak Shukla: "the better they look, the less likely they'll grab my attention."
Force 5 - The operator playbook: triage, draft, heavy-edit, voice-check
AI earns its keep as a sorter and a structurer, gated by a human voice check.
Brian Hansen of Rocket Pilots uses it as a filter first: "I have Claude separate what is obvious, what is useful, and what is actually quotable. Only the last category survives." Anthony May of NeedAnAttorney runs a five-step gate ending in the only rule that matters: "only send if the final response sounds like something I would say." Kruno Sulić of Cliprise (Connectively) has a hard test: "the first paragraph should already contain the answer. If it starts with compliments or a general statement about how important AI is, it is losing. If a sentence has strong polish but no real observation, cut it." And the meta-standout, Jordan Eaton of BOSS Assistants, bakes the tells into the tool: "every query goes through a Claude project with strict guidelines built in: no em dashes, no rhetorical triads, no constructions that contrast two ideas in back-to-back sentences." Callum Gracie of Otto Media (Connectively) summed the upside and the risk: "AI can make a good pitch faster, but it makes a bad pitch louder," crediting the disciplined version with moving his Featured account from 3,000-plus to the top 85 in under a year.
Force 6 - The dissent worth keeping
It is not unanimous, and the disagreements sharpen the finding. The honest version: AI helps response rates when you rewrite enough that it stops sounding like AI.
Anthony LoCascio of Marketing Baristas runs the other way on volume: "we use AI to hyper-personalize performance-based cold email, sent from separate outreach domains, and generate two to three times more responses than generic outreach." Bård Ionson, an artist who raised funds for an installation through long iterative threads with the model, pushed furthest: "if a pitch has been crafted iteratively with the model, where a human actively corrects its assumptions and injects real context, I am not sure I would be able to tell the difference." And the refusenik, Adam Gorham Films, a boutique wedding cinematographer: "referrals account for over 90 percent of my inquiries. When I reach out, I send handwritten notes or short video messages instead of templated emails. Cold email never fit what I do, so I've never drafted an AI pitch." The dissent does not break the consensus. It locates the line: AI is fine right up to the point where it sounds like AI.
It is not that AI cannot pitch. It is that AI supplies the half that no longer earns a reply, structure and polish, and the half that does, a specific proof the recipient cannot get elsewhere tied to their actual work, is the half only a human has. In an inbox full of machine-clean drafts, the AI tell is now the thing that gets you deleted. Use AI to start the draft and sort the noise. Do not let it hit send.
Contributors
Last quarter my team reviewed over 500 inbound pitches and logged every reply and rejection. Raw AI drafts get a 0% response rate. I spot AI pitches in 3 seconds: generic phrases like 'I admire your work' with zero specifics, no client name, no reference to our actual results. We flagged 70% of inbound AI pitches last month alone. The reply came from a pitch that showed real research.
In 2024 I tested ChatGPT for cold emails. My response rate dropped 40% in the first week. The raw AI output was all puff and no substance: 'expand your potential,' 'streamline your outreach.' Those went straight to trash. The ones that earn a reply show real research and name a specific result, like a $0.15-per-subscriber campaign for Thai restaurants.
The pitches that get ignored open with vague claims about 'boosting engagement' without referencing any of our actual client results. Real outreach that earns a reply always names one of our published outcomes, such as the 5,000% ROI on a Google Ads campaign, and ties it to a concrete next step.
I use GPT to generate 6 to 8 angle variations from a short brief, then a human edits, typically 40 to 60 percent of the draft. The AI supplies structure and phrasing; we supply credibility. AI-assisted and humanized pitches that cite a clear metric or a named customer get responses at roughly 12 to 18 percent to journalists.
When we tested lightly edited AI drafts against heavily rewritten ones in B2B outreach, the lightly edited version saved time but got about 35 percent fewer replies. The message was clean, logical, and completely forgettable. My best workflow is AI for structure, then rewrite roughly 50 to 60 percent before sending.
I watched a $50,000 partnership vanish because the CEO's pitch sounded like a ChatGPT template; the investor called it out in a 10-second email. I tracked 210 pitches over 9 months: fully manual emails got a 58% reply rate, AI-first drafts got 22%. Tailoring a pitch to the journalist's recent work takes 15 minutes, and it works because it shows I read their work.
My workflow: use AI to triage the source request and decide if I'm a fit, draft a first version from the writer's exact question, then heavily edit so it includes my real experience and a specific point of view. Cut anything that sounds like a press release or generic thought leadership. Only send if the final response sounds like something I would say.
I lost a $10,000 placement with an automotive editor because my pitch sounded like a machine wrote it. He told me it lacked genuine insight. I tested ChatGPT on one outreach batch and my response rate dropped from 40% to 5%. I spent hours editing every draft to sound like me. Those hours cost more than the tool saved.
A journalist from TechRound told me, 'Your last three pitches sounded like they were written by the same robot.' That stung because I'd placed with them seven times before. The AI drafts had a weird rhythm: perfect grammar, no contractions, always three paragraphs, always ending with 'happy to discuss further.' Nobody talks like that.
AI creates the first drafts of my HARO responses and cold emails, but around 60 to 70 percent of the final draft is me. The main thing not to do is send the initial AI draft. The pitches that resulted in coverage always had a contrarian thought or a personal anecdote. The robotic ones got ignored. In some ways, the better they look, the less likely they grab my attention.
What earns replies is specificity: 'check your robots.txt for GPTBot, ClaudeBot, OAI-SearchBot, and PerplexityBot' beats 'AI is transforming search.' I edit AI-assisted replies heavily enough that the final version would pass as something I typed quickly between client calls. Fully manual still wins when the ask needs a personal story.
I use AI to pull search-trend data, then rewrite every line with our actual affiliate commission structures and client-specific product details before sending. Manual versions consistently secured placements. AI-only drafts get ignored when they skip the data-backed incentive angle entirely.
I use ChatGPT to structure the first draft, then rewrite about 60 to 70 percent myself. The version that performs best has one specific story, one clear opinion, and one line that sounds like something only I would say. A nearly AI-written pitch full of 'leveraging innovation' and 'driving growth' got no reply; the rewritten version landed.
Referrals account for over 90 percent of my inquiries. Cold outreach has never fit what I do, so I've never drafted AI pitches. When I reach out to collaborators, I send handwritten notes or short video messages instead of templated emails. Those feel human, and they work.
We use AI to draft initial templates but mandate a human-in-the-loop review for all outbound. The pitches that earn replies lead with objective technical audits, naming a specific mobile-responsiveness error or accessibility issue. Inbound pitches get instantly deleted when they offer web services without running a basic scan on our site.
I use AI for pitch triage, not final voice. What earns replies is specificity: the audience, the trigger, the offer, and why now. When I screen inbound pitches, the giveaway is that the pitch has structure but no judgment. It sounds polished, but it can't reason about lead stage, pain point, or trust barrier.
I use AI to map foundational positioning but never to write the final output, and I manually rewrite 100 percent of the draft to maintain a human tone. Fully automated AI drafts lack contextual depth and cost you the placement because they read like an instruction manual.
I use AI as a first-draft assistant, not a send button: pull the journalist's exact angle into the prompt, get a rough draft, then rewrite the opening, examples, and strongest opinion by hand. If the draft still reads like it could have been sent to 200 other people, it is not ready.
Every query I respond to goes through a Claude project with strict guidelines built in from the start: no em dashes, no rhetorical triads, no constructions that contrast two ideas in back-to-back sentences. The part most people skip is the rules. That is where AI quietly kills the pitch before it gets read.
I use AI as a sorting tool before it becomes a writing tool. Journalists ask broad questions, so the first step is having Claude separate what is obvious, what is useful, and what is actually quotable. Only the last category survives. From there the answer is rebuilt around a specific moment, metric, or mistake.
On the receiving side, the tell isn't vocabulary, it's catalog blindness. We train personalized models on retailer catalogs; anyone pitching us generic 'semantic search solutions' clearly ran our homepage through a summarizer and missed what we do. On the sending side, the difference between a reply and silence is whether the answer contains a number a journalist can't get anywhere else.
Inbound AI pitches are dead on arrival if they contain dead giveaways like 'I hope this email finds you well' or ChatGPT's favorite words like 'testament' and 'delve.' In high-end reputation management, sounding robotic instantly kills trust. We draft response frameworks with Claude but heavily edit every word before sending.
You need to do some editing, but there is no single right amount. For a journalist's work the editing needed is minimal but not zero. You need to change some phrasing and plug the holes in the story, go through a couple of passes, then send. Fully automated bot responses range from 10 to 40 percent success, and quality complaints make many users stop.
We use AI to hyper-personalize performance-based cold email and generate two to three times more responses than generic outreach. We never send raw AI drafts, always heavily editing to inject our voice and local market context, and we send from separate outreach domains to protect our reputation.