Writing Press Releases with AI: What to Keep Human

1. A press release is a news document, not a marketing page

The first mistake AI tools make with press releases is treating them like ads. Ask a model to write one and you’ll get something that reads like a landing page: “We are thrilled to announce our revolutionary, game-changing solution…” Journalists delete that sentence before they finish reading it, because “thrilled” and “revolutionary” are the two words that tell them there’s no actual news here.

A press release has one job: give a busy stranger the facts they need to decide, in under a minute, whether your story is worth their readers’ time. That means the inverted pyramid — the most important fact first, then the supporting facts, then the background. No buildup, no tease, no clever opening. The AI can learn this structure, but you have to demand it. Tell it: “No adjectives in the first paragraph. Lead with what happened, when, and why it matters.”

If the first paragraph of your draft could belong to any company in your industry, it isn’t news yet. Rewrite it until it could only be yours.

2. Let the AI build the skeleton

Here’s what AI is genuinely good at in a press release: the boring, structural work. Give it your facts — the date, the product, the price, the quote you already wrote, the boilerplate — and ask it to assemble a clean draft in standard press release format. Headline, subhead, dateline, first paragraph, quote, supporting details, boilerplate, media contact. It will do this in thirty seconds, and the structure will be correct.

I treat the AI draft the way a carpenter treats a frame: useful, necessary, and not the finished house. The frame saves me the fiddly formatting work — getting the dateline right, keeping the boilerplate consistent, making sure the contact block is complete. That’s real time saved. But nobody lives in a frame, and nobody publishes a skeleton.

One practical tip: feed the AI your previous press releases as examples, not just a prompt. “Write it in the style of these two releases” produces far better results than “write a professional press release”, because the model stops guessing what a press release sounds like and starts copying your actual house style.

3. The news judgment is yours — the “why now” test

Every press release needs an answer to one question: why is this news now? A product launch has one. A funding round has one. A new hire usually doesn’t, and neither does a “we’re excited to share” update about a feature your competitors shipped last year.

AI cannot answer this question for you, because it doesn’t know what your industry’s journalists care about this week. It will happily write a confident release about a non-event. Your job is to be the editor who says no — or who finds the real angle.

Before I let a release go out, I run the why-now test on it. Would a journalist who covers my field find this worth a story, or would they file it under “company sent a press release about itself”? If it’s the second one, the fix isn’t better writing. It’s a better reason to send it — a real number, a real customer, a real change in the market. Sometimes the honest answer is “don’t send this one yet”, and that’s the most valuable edit of all.

4. Quotes are the human fingerprint

This is the part you must never delegate. AI-generated quotes are the easiest thing in the world to spot: they’re smooth, vague, and say nothing a real person would actually say. “We are delighted to bring this innovative solution to our customers and look forward to continuing to deliver value.” No human being has ever said that sentence out loud. A journalist knows it instantly, and it costs you credibility in the exact paragraph where credibility matters most.

A real quote has a pulse. It has an opinion, a specific detail, or a sentence structure that sounds like one particular person. “We built this because our support inbox was full of the same complaint for six months” is a quote. “We’re thrilled to announce” is a placeholder wearing a suit.

Write your quotes yourself. Better still, get them from the actual person — a two-minute voice note, a sentence in an email, anything in their own words. Then clean up the grammar and leave the voice alone. If your CEO’s quote sounds like it could have come from any CEO, rewrite it until it sounds like your CEO.

5. Facts, numbers, and names — verify every one

AI tools invent facts. This is not a flaw in your prompt; it’s how the models work. They will produce plausible-sounding statistics, confident dates, and quotes attributed to people who never said them. In a blog post, a wrong number is embarrassing. In a press release, it’s a correction notice — or a journalist who never opens your emails again.

Every number in your release needs a source you can point to: the sales figure from your actual dashboard, the survey result from the actual report, the date from the actual calendar. Every name needs its spelling checked against the person’s own email signature. Every claim about being “the first” or “the largest” needs proof, because a journalist will check it, and “the AI wrote it” is not a defense.

My rule is simple: I read the final draft with a pen and circle every fact, then check each circle against its source. It takes ten minutes. Skipping it has cost other people their reputations.

6. The subject line and the first paragraph do the real work

Most press releases are never opened, and the ones that are opened are mostly skimmed. The subject line of your email and the first paragraph of the release are the whole game. AI is decent at generating ten subject line options — let it. Then pick the one that sounds like news, not like marketing. “Local bakery launches gluten-free line” beats “Exciting news from our kitchen!” every single time, because one of them tells the journalist what the story is and the other one wastes their time.

The first paragraph should answer what, who, when, and why it matters — in two sentences. If your AI draft opens with background or throat-clearing (“In today’s fast-moving market…”), cut it. The news goes first. Everything else is support.

The honest truth

AI has made press releases faster to produce, and that’s genuinely useful — the formatting, the structure, the first pass at a subject line. But a press release is a trust document. You’re asking a journalist to stake their reputation on your facts, your quotes, and your judgment about what’s newsworthy. No model can lend you that trust. It has to come from you: the real quote, the verified number, the honest answer to “why now”.

Use the AI for the skeleton. Keep the news judgment, the quotes, and the facts for yourself. That’s the division of labour that gets releases opened — and, more importantly, believed.

Frequently Asked Questions