AI will write you a press release in nine seconds. It will also be the same press release everybody else got, stuffed with phrases no human has ever said out loud, containing a statistic it invented and a quote your CEO never approved.
That doesn't mean don't use it. It means use it for the parts it's actually good at. Here's the split, a prompt sequence that works, and the specific tells that make an editor delete your email.
What AI is genuinely good at
Structure. Press releases have a rigid, boring shape: headline, dateline, lede, supporting paragraph, quote, details, boilerplate, contact. AI has read a million of them. Ask it to arrange your facts into that shape and it will do it correctly every time, which saves you the twenty minutes of staring at a blank page. If you want to see the shape first, our free templates are already built to it.
First drafts from raw notes. Dump your ugly bullet list (what happened, when, numbers, who's involved) and ask for a draft. The draft will be mediocre. Mediocre-and-existing beats perfect-and-imagined. You edit from there.
Headline variants. This is the single best use. Ask for fifteen headlines and you'll get twelve bad ones and three that are better than what you'd have written alone. Generating options is cheap for a model and expensive for a human brain at 4pm.
Compression. "Cut this to 450 words without losing the numbers" is a task AI does well and people do badly. Since most wires price by word count, this literally saves money.
Tone audits. Paste your draft and ask: "Which sentences here are claims without evidence? Which adjectives could be deleted with no loss of meaning?" It's a decent critic even when it's a weak writer.
What AI is bad at, and will not tell you
Quotes. AI-generated quotes are the loudest tell in the document. They all sound the same: "We are thrilled to announce this milestone, which represents our continued commitment to delivering value for our customers." No executive has ever said that. Nobody has ever said that. A real quote has an opinion, a specific, or a concession in it. Get a human to say something on a call, transcribe it, tidy the grammar, and read it back to them. More on this in our guide to press release quotes.
Facts and numbers. Models fill gaps confidently. If you don't give it your funding amount, it may produce a plausible one. If you mention a market, it may attach a market-size statistic with a research firm's name on it that does not exist. Every number in your final draft must come from you, not from the model. Verify each one against a source you can point to.
News judgment. Ask AI whether your announcement is newsworthy and it will encourage you, because it is built to be helpful. It cannot tell you that a feature update isn't news, that your rebrand interests nobody outside your building, or that you are three weeks too late on the story. That judgment is yours, and getting it wrong is more expensive than any writing error, because you'll pay for distribution on something nobody reads.
Your voice. Models regress to the mean of everything they've read. The mean of press releases is sludge.
A prompt sequence that actually works
Don't ask for "a press release about our launch." Run it in stages, and give it facts at every step.
Step 1: Facts in, structure out.
Here are raw facts about an announcement: [dump everything, including dates, numbers, names, titles, what changed, who it affects]. Do not add any facts I have not given you. Arrange these into a standard press release structure: headline, dateline, lede paragraph answering who/what/when/where/why, two supporting paragraphs, a placeholder marked QUOTE NEEDED, and a boilerplate. Mark with [MISSING] anything the structure needs that I have not provided.
The [MISSING] instruction is the important part. It turns the model from an inventor into an auditor.
Step 2: Fix the lede.
Rewrite the lede three ways. Version A leads with the number. Version B leads with what changes for the customer. Version C leads with the thing a skeptical reporter would find most surprising. Keep each under 35 words.
Step 3: Headlines.
Give me 15 headlines under 12 words. No colons. No words like revolutionary, innovative, game-changing, leading, or seamless. Five should be plainly factual, five should lead with the number, five should lead with the consequence.
Step 4: The quote, handled honestly.
Here is a transcript of what our CEO actually said on this topic: [paste]. Pull the two sentences that are most specific and most quotable. Lightly fix grammar. Do not add ideas she did not say.
If you have no transcript, you have no quote. Go get one.
Step 5: The hostile read.
You are a business editor with 200 emails in your inbox. List every reason you would delete this. Then list every unsupported claim.
Step 6: Cut.
Cut to 500 words. Preserve every number, name, and date. Cut adjectives first.
The tells editors spot
Editors don't run detection software. They pattern-match in about ten seconds:
- Triads everywhere. "Faster, smarter, and more reliable." Models love lists of three. Humans mix it up.
- "Not just X, but Y." An AI tic that has quietly become a tell.
- Empty intensifiers. Thrilled, excited, proud, delighted. Nobody is sad in a press release, so the word carries no information.
- Claim without number. "Significantly reduces costs" instead of "cuts average onboarding time from 11 days to 3."
- Uniform sentence length. Real writing has a rhythm: a long sentence, then a short one. AI drafts run flat at 18–22 words for paragraphs at a stretch.
- Boilerplate adjectives. "A leading provider of innovative solutions." Every company says it, so it identifies nobody.
- Quote that could be swapped into any other release. The test: if you replace the company name and the quote still works, it's worthless. Delete it or get a real one.
The honest framing
AI moves you from blank page to competent draft in fifteen minutes instead of two hours. That's real value. It does not make your news newsworthy, and no tool (AI, wire, or agency) makes a journalist care about an announcement that has nothing in it. Distribution puts your release in front of people. What's in it decides whether anyone reads past the headline.
Use the model for structure, options, and compression. Use your own judgment for facts, quotes, and whether to send it at all. Then check it against a proper writing checklist before it goes anywhere.
Skip the blank page entirely: [our guided builder](/how-it-works) walks you through the same structure and asks for the facts AI would otherwise invent.