top of page

How to Write an AI Press Release in 4 Minutes

Shift Quality Editorial
11 minutes ago
10 min read

A satirical template, with notes for the people reading them.



Every AI vendor announcement in 2026 sounds the same. That is not an accident. There is a template, and once you can see the template, you can see the announcement underneath it — which is usually a lot smaller than the announcement implies.


This is the template. The translations are for the people who have to read these things and make budget decisions.



1. The Lede


Today, [STARTUP_NAME], the [SUPERLATIVE] leader in [BROAD CATEGORY] AI, announced the launch of [PRODUCT_NAME], a [ADJECTIVE] platform that uses agentic [NOUN] to [VERB] the way [INDUSTRY] teams [BUSINESS_PROCESS].

Translation tips:


  • "The leader in" — There is no category yet. The startup is creating one in the press release. The leader of a category of one is a phrase, not a position.

  • "Agentic" — Often means "we put a for loop around an LLM call." Ask what the agent is actually allowed to do without a human in the loop. The honest answer for most products shipping right now is "nothing that costs money or touches production."

  • "Transform the way [industry] teams [do thing]" — Translation: this is software that does the same thing existing software does, but with a chat box in front. Sometimes that is genuinely better. Sometimes it is a chat box in front of existing software.


2. The Problem Statement


Today's [INDUSTRY] teams are drowning in [VOLUME NOUN], struggling to [ASPIRATIONAL VERB] at the speed of business. Legacy tools were not built for the AI era.

Translation tips:


  • "Drowning in [data / tickets / requests]" — The problem is real. The solution is rarely as proprietary as the press release implies.

  • "Speed of business" — Means nothing. If a vendor uses this phrase, mentally replace it with "faster" and notice that the sentence is now empty.

  • "Legacy tools were not built for the AI era" — Most "legacy tools" have been adding AI features for two years. The honest version of this sentence is "our tool has AI features and a competitor's tool also has AI features and we are pretending that is not the case."


3. The Solution


[PRODUCT_NAME] combines state-of-the-art foundation models with proprietary [DOMAIN] expertise to deliver enterprise-grade [OUTCOME] out of the box. The platform offers native integration with the tools your team already uses, including [3 LOGOS].

Translation tips:


  • "State-of-the-art foundation models" — They are calling OpenAI, Anthropic, or Google. The state of the art belongs to those labs. The startup's contribution is the prompt and the UI. Sometimes the prompt and the UI are very good. Sometimes they are a wrapper, and the wrapper is the entire product. Worth knowing which.

  • "Proprietary [domain] expertise" — Usually means "we have a system prompt" or "we fine-tuned on a small dataset." Both can be valuable. Neither is a moat.

  • "Enterprise-grade" — Has no fixed meaning. Ask for the specific evidence, and notice that the usual list mixes different kinds of thing. SOC 2 Type II is not a certificate; it is an auditor's report on whether controls operated over a period, so ask for the report, its period and which systems it covered. ISO/IEC 27001 is a certification, and the scope statement on the certificate tells you whether the product you are buying is inside it. HIPAA and GDPR are laws, not badges: the useful questions are whether the vendor will sign a business associate agreement, and what its data processing agreement says about where your data lives and which sub-processors see it. A vendor that has done this work will hand over the documents. A vendor that says "enterprise-grade" and stops there is asking you to assume. (Your AI Vendor's Compliance Doesn't Make You Compliant covers what stays your responsibility even when the documents are in order.)

  • "Native integration" — Could mean a deep, supported, two-way sync. Could also mean a webhook that fires when something happens. Ask: what data flows, in which direction, on what trigger, and what happens when the integration breaks?


4. The Customer Quote


"[PRODUCT_NAME] has been transformational for our team," said [TITLE] [NAME] of [CUSTOMER]. "We've seen a [ROUND PERCENTAGE]% increase in [METRIC] and a [ROUND PERCENTAGE]% reduction in [OTHER METRIC], all without [PAIN POINT]."

Translation tips:


  • "Transformational" — Listen for a quieter word. "Useful," "promising," and "saves us time" mean something. "Transformational" means the customer was asked to provide a quote and the vendor's PR person wrote a first draft.

  • "[Round percentage]% increase" — Round numbers without a baseline are not measurements. They are estimates someone gave a salesperson. Ask: increase from what to what, measured how, over what period, by whom?

  • "Without [pain point]" — The pain point named here is usually the one the competitor causes. Notice which competitor is implied.


5. The Analyst Quote


"[STARTUP_NAME] is doing exciting work in a space that is ripe for disruption," said [ANALYST NAME], [VAGUE TITLE] at [ANALYST FIRM]. "Their approach to [TECHNICAL CONCEPT] is differentiated."

Translation tips:


  • Analyst quotes in press releases are arranged, not overheard. The big firms control how their names appear: Gartner, for example, requires its Content Compliance team to approve any Gartner mention or quote in a third-party press release. That process exists to stop misquotation, and it also means the quote is a negotiated artifact, not an independent recommendation. Read it as a signal that the vendor cares about appearing in analyst coverage, which is itself useful information about who they are selling to.

  • "Differentiated" — A word analysts use when they have not yet decided whether something is good.

  • "Ripe for disruption" — Means the analyst could not name the existing solutions that already work.


6. The Vague Customer Logos


[PRODUCT_NAME] is trusted by [3 LOGOS YOU HALF-RECOGNIZE], including teams at [FORTUNE 500 COMPANY].

Translation tips:


  • "Trusted by" — In B2B SaaS, "trusted by" can mean anything from "they paid us seven figures and renewed twice" to "an engineer at this company signed up for the free trial once." Both look the same on the website.

  • "Teams at [Fortune 500]" — Note the phrasing. Not "[Fortune 500] uses." That would be a procurement decision. "Teams at" means one team, possibly two people, possibly on a credit card.

  • If a logo matters to your decision, ask the vendor for a referral to a named contact at that company. A real customer relationship survives that ask. A logo lift does not.


7. The Pricing Section (Often Missing Entirely)


[PRODUCT_NAME] is available today with flexible pricing to meet the needs of teams of all sizes. Enterprise pricing is available upon request.

Translation tips:


  • "Flexible pricing" — Means "we will charge you what we think you can pay."

  • "Available upon request" — Means "we want to qualify you before showing you the number, because the number is high and we want to be in a conversation when you see it."

  • Ask for the price list at the first call. A vendor that cannot give you a starting number is telling you that the number depends on what you tell them about your budget. Plan accordingly.


8. The Boilerplate


About [STARTUP_NAME]: Founded in [YEAR] by [FOUNDERS], [STARTUP_NAME] is on a mission to [MISSION STATEMENT THAT IS IDENTICAL TO COMPETITORS]. The company is backed by [TIER 1 VC] and has raised $[AMOUNT] to date.

Translation tips:


  • "On a mission to" — Notice if the mission could be cut and pasted onto a competitor's About page without losing meaning. Most can. That is not necessarily a bad sign, but it is worth seeing.

  • "Backed by [Tier 1 VC]" — A signal of momentum, not of product quality. Some of the worst enterprise software ever shipped was backed by Tier 1 VCs. Some of the best was bootstrapped.

  • "Raised $[amount]" — Bigger is not better for a buyer. A heavily funded startup needs to deliver outsized returns to investors, which means pricing pressure and feature pressure. You may be a beneficiary or a casualty of that depending on the year.



One release, read properly


Composite example, assembled from common patterns. The vendor, product and buyer are invented.


The head of IT operations at a mid-sized logistics company receives a release for an "agentic incident co-pilot" that "reduces mean time to resolution by 60%," is "trusted by teams at" two household-name retailers, offers "native integration" with their ticketing system, and is "enterprise-grade." Her CIO forwards it with one line: should we pilot this?


She does not argue with the release. She sends the five questions below, in writing, before agreeing to a demo. The answers come back in a week.


  • What can it do without a human? It drafts incident summaries and suggests runbook steps. Every action needs an engineer's click. So "agentic" means "suggests," and the risk is engineers clicking through suggestions at 3 a.m. without reading them.

  • A named production customer? The two retail logos are pilots, one team each. The vendor offers a reference at a smaller company that has run it in production for five months. She takes the call. The reference likes the summaries and has not measured resolution time.

  • The failure mode? When the model suggests a wrong runbook step, nothing flags it; the engineer is expected to notice. There is no log of rejected suggestions.

  • Pricing in 18 months? Introductory per-seat pricing for the first year, then "standard list," which the vendor will not state.

  • Where does the data live? Incident text goes to a third-party model provider. The provider does not train on it under the vendor's current plan, but the data processing agreement allows the vendor to change sub-processors with 30 days' notice.


The "60% faster resolution" turned out to be one pilot team's estimate, with no baseline. None of that makes the product bad. The summaries may well save time. But the decision changes from "pilot an agent that cuts resolution time by 60%" to "trial a summarisation tool, measure resolution time ourselves, log rejected suggestions, cap the year-two price in the contract, and require notice and an exit right on sub-processor changes." That is a decision she can defend, and it took one email.


Why this is more than a joke


Overstated AI claims have started to carry legal consequences for the people making them. In March 2024 the US SEC settled its first "AI washing" cases, against two investment advisers that had claimed AI capabilities they did not have. In September 2024 the FTC's Operation AI Comply brought actions against companies using AI claims to deceive customers. Most vendor releases are nowhere near that line. But the same gap between the words and the system is what you are paying for when you buy on the release, and nobody is going to fine the vendor on your behalf.


What to actually do with this


The point of all this is not to be cynical about every announcement. There is genuinely good AI tooling shipping in 2026. Some of it will save you real money. Some of it is being marketed by people who are good at marketing and not as good at the thing the marketing is about.


When a vendor sends you the press release, send them five questions before the demo:


  1. What is the agent allowed to do without a human in the loop? Be specific. A list of actions, not a description.

  2. Can you put me in touch with a named customer using this in production, not in a pilot? A real customer relationship survives this ask.

  3. Walk me through the failure mode. When the model is wrong, what does the user see, what gets logged, and who notices?

  4. What is your pricing today, and what stops you from re-pricing the moment we are integrated? A vendor that has thought about this has a real answer. A vendor that has not is one re-pricing event away from being your problem.

  5. Where does my data live, and who else can see it? This includes the model provider, the analytics provider, the support team, and anyone with production access. If the answer is vague, the answer is bad.


Score the replies the simple way: each answer is either a description of behaviour (actions, data flows, a named contact, a number with a baseline) or a restatement of the marketing. A vendor whose answers are mostly behaviour is usually worth your time. A vendor that deflects has told you what you need to know. For the vendors that pass, the Vendor Evaluation Scorecard Template and Evaluating New Tools Without the Hype take the evaluation from here.


The press release is for the buyer. The questions are for the user. Closing the gap between what the buyer reads and what the user lives with is half the job of a competent tech leader.


That gap is where the budget goes to die.


Sources




Part of the Perspectives guide — ShiftQuality's field notes on building, AI, and business.


Frequently Asked Questions


Is this actually a guide for writing press releases?


It is a guide for reading them. Every AI vendor announcement in 2026 follows the same template. Once you can see the template, you can see the announcement underneath it — which is usually much smaller than the announcement implies. The satire is the delivery vehicle. The translations are the point.


What are the biggest tells that a vendor is overselling?


Five recurring ones. (1) The word 'agentic' without a description of what the agent does autonomously. (2) 'Enterprise-grade' without naming the specific audit reports and certifications, with their scope. (3) 'Native integration' that turns out to be a webhook. (4) Customer logos with no quote or named contact. (5) Performance numbers ('70% faster,' '10x improvement') without a baseline. Our rule of thumb: any single one is a yellow flag; three or more in one announcement is a red flag.


What questions should I ask a vendor at a demo?


Five that will defuse most hype. (1) What is the agent allowed to do without a human in the loop? (2) Show me a customer using this in production, not in a pilot. (3) What is the failure mode when the model is wrong, and how do I see it happen? (4) What is your pricing in 18 months, and what stops you from re-pricing the moment we are locked in? (5) Walk me through the data flow — where does my data live, and who else can see it? Vendors who can answer these confidently are usually worth talking to. Vendors who deflect have told you what you need to know.


Why does this template keep working on people?


Because the people who buy this software are not always the people who use it. The buyer reads the press release and the analyst quote. The user finds out six months later when integration is harder than promised, the agent does less than implied, and the pricing has shifted. The template optimizes for the buyer's pattern-matching, not the user's outcome. Closing that gap — making sure the person who signs the contract understands what they are actually getting — is half the job of a competent tech leader.

 
 
bottom of page