Launching an AI Product Without the Hype
AI · 10 min read ·
How to describe an AI product accurately at launch: what it does, where it fails, how you tested it and which claims to avoid. A plain-language checklist.
Every era of technology has its vocabulary of inflation. For AI products, the words are familiar: intelligent, autonomous, human-level, revolutionary. They are used so often, and checked so rarely, that readers have learned to discount them. A launch that avoids them stands out, and, more importantly, can be believed.
This guide is a plain-language approach to launching an AI product honestly: describing what it does, naming where it fails, backing claims with tests and avoiding language that invites disappointment. It is general information, not legal advice.
Why precision pays
AI products are unusually easy to over-describe. Their behaviour is probabilistic: the same input can yield different outputs, and quality varies across tasks in ways that are hard to predict. A demonstration can look astonishing and a week of real use can look disappointing. When claims run ahead of reality, users feel misled, and the word spreads quickly.
There is also a regulatory dimension. Advertising authorities in several countries have made clear that claims about AI products must be truthful and substantiated like any other claims. A product that says it "detects fraud with perfect accuracy" or "replaces your analyst" needs evidence for the statement, and the evidence has to match the claim.
The commercial argument is simple. Precision builds trust, and trust is what turns a curious visitor into a user who keeps using a tool whose output they must occasionally check.
Describe what it does, in verbs
Replace adjectives with verbs. "An intelligent assistant for customer support" says little. "Reads an incoming support email, drafts a reply from your help articles and shows the articles it used" says exactly what happens.
A good description answers four questions.
- What goes in? The type of input, and its limits.
- What happens? In one or two plain steps.
- What comes out? The form and use of the output.
- What does the human do? Review, approve, correct.
If you cannot answer the fourth question, think about whether the product is as autonomous as you imply. Many useful AI products are assistive: they draft, suggest, summarise and flag, with a person making the decision. That is a strength, not a limitation.
Say what it is built on, accurately
Users and buyers increasingly ask what is underneath. A large language model is a type of machine-learning model trained on vast amounts of text to predict and generate language; many products are built by calling a model provided by someone else, with additional instructions, data and logic around it.
Be accurate about this.
- If you use a third-party model, say so if it matters, and do not imply you trained it.
- If you fine-tuned or built your own, say what that means in practice, not as a boast.
- If data goes to a third party, tell users clearly. See the separate guide on privacy statements for AI tools.
- If the model can change, say how you handle changes, because behaviour may shift.
Describing the architecture in a sentence or two is enough. It reassures technical readers and pre-empts the question.
Name the failure modes
Every AI product fails in characteristic ways. Naming them before users find them is the single most credible thing you can do.
For language-based products, a well-known failure is hallucination, the generation of statements that sound plausible but are false or unsupported. For classification products, errors may cluster in particular groups or conditions. For products that extract data, formatting variations can cause misses. For products that act on the world, small errors can compound.
Write a short "Known limits" section. Include:
- Types of input the product handles poorly.
- Observed error patterns from your testing, in plain words.
- Things you do to reduce errors, such as showing sources or asking for confirmation.
- What users should check.
- What the product is not suitable for. For example, medical, legal or financial decisions made without professional review.
This is not weakness. Users who understand the limits use the product more safely, and your support load falls.
Back claims with tests you can describe
If you claim performance, show how you measured it.
- State the task and the data. What were you testing it on?
- State the measure. Accuracy, error rate, time saved, agreement with human reviewers.
- State the sample size. Fifty examples tell you less than five thousand.
- State the conditions. Date, version and settings.
- Say who ran the test. Independent evaluation carries more weight than your own.
- Publish what you can, such as the test set description or the method.
The NIST AI Risk Management Framework, a voluntary framework from the US National Institute of Standards and Technology, describes practices for mapping, measuring and managing risks from AI systems. You do not need to adopt it formally to borrow its habits: write down what the system is for, how it could fail, how you measure it and who is responsible.
Words to avoid
- "Human-level" and "superhuman" without a named task and a measure.
- "Autonomous" unless no human is involved at any point.
- "Understands" unless you can explain what that means here.
- "Never makes mistakes" and its relatives.
- "Works with any data" or "any language".
- "Unbiased" and "objective". Systems built on human data and human choices carry the properties of both.
- "Secure" and "private" without detail.
- "Powered by AI" as a label with no content.
If a word cannot be tested, it does not belong in a headline.
Show, do not tell
The best AI launch pages let visitors try the product, or see real unedited output.
- Real examples, including a mediocre one.
- A way to try it without registering, if feasible.
- Before-and-after pairs with the actual input.
- A short video showing a full task, not an edited highlight.
- Sample output labelled as such, with the date and version.
Cherry-picked examples are tempting and are easily exposed. If you show your best output, say it is the best, and show a typical one too.
Be clear about humans in the loop
If humans review outputs, either at your end or the user's, say so. If your product relies on human workers behind the scenes, be transparent about it; presenting human work as automated is misleading. If you rely on users to check, say what they should look for.
A worked example
A team launches a tool that summarises long meeting recordings. Their first headline: "Revolutionary AI that understands every meeting." They rewrite.
New headline: "Turn a recorded meeting into a one-page summary with action items." Beneath: "It transcribes the recording, drafts a summary and lists tasks with the names mentioned. You review and edit before sharing."
Known limits: "Accuracy drops with heavy accents, crosstalk and poor audio. In our test of sixty recordings in March, about one in ten summaries missed or misattributed at least one action item. We show the transcript alongside every summary so you can check. Not suitable as a legal record."
Method note: "Tested on sixty anonymised recordings from consenting volunteers; two reviewers compared summaries with the recordings; results and method are on this page. Built using a third-party language model; recordings are processed as described in the privacy page."
Visitors can try a sample recording without signing up. The page is plain, but it sets honest expectations, and the support team report fewer complaints about accuracy than they had feared.
A pre-launch checklist
- Is every claim supported by a test you can describe?
- Does the description say what goes in, what happens, what comes out and what the human does?
- Is the underlying technology described accurately?
- Is there a visible "Known limits" section?
- Are examples real and representative?
- Is data handling explained?
- Have you removed unfalsifiable adjectives?
- Do you have a plan for model changes and for reports of errors?
Questions founders ask
Will honest limits scare people away? Some, yes, and those were unlikely to be good users. Others will trust you more. Most buyers of AI tools have already been disappointed once and are looking for evidence of candour.
What if a competitor makes bigger claims? Let them. Claims that cannot survive a trial tend to produce churn and complaints. Your tests and limits section will outlast their slogans.
How should I handle reports of errors? Provide an easy way to report them, acknowledge each, add reproducible cases to your test set and tell users when you have improved. A visible loop of reports and fixes is itself a trust signal.
Should I label AI-generated output? Where users might mistake it for human-produced or verified content, yes. Clear labelling protects users and you.
Can I use "beta"? Yes, if you mean it: say what is unfinished and what will change, and do not use the word as a way to excuse problems.
Keeping the page honest over time
AI products change faster than most. Models are updated, prompts are revised and data shifts. Put a date on your performance statements and a reminder to retest each time something significant changes. Keep a short public changelog that notes changes in behaviour, not just new features. If a claim is no longer true, correct it the same day. A page that is accurate this month is worth more than a page that was impressive last year.
Summary
Launch an AI product by describing verbs not adjectives, saying what it is built on, naming failure modes, backing claims with tests you can describe and avoiding words that cannot be checked. Be clear about the human's role and let visitors see real output. Honest precision builds the trust that lets people keep using a tool they must sometimes double-check.
Questions and answers
- Why is AI marketing so often criticised?
- Because claims frequently outrun what the product demonstrably does, and because the words used, such as intelligent or autonomous, are vague.
- Do advertising rules apply to AI claims?
- Yes. Regulators in several countries have said that claims about AI products must be truthful and backed by evidence, like any other claims.
- How should I describe failure modes?
- Plainly and early: what kinds of input the product handles poorly, how often you have seen errors in testing and what users should check.
- Should I say which model I use?
- If it matters to users, such as for data handling or capability, say so accurately. Do not imply you built a model you license.