7 Red Flags in AI Product Launches (Learned the Hard Way)
Last spring I gave a company my work email, my use case, and twenty minutes of a discovery call because their launch video showed exactly the workflow I needed. The product, when I finally got access six weeks later, could do roughly a third of what the video showed. The other two-thirds was "on the roadmap." Nobody had lied, exactly. Every scene in that video was real. It just wasn't representative — and the gap between real and representative is where most AI launch disappointment lives.
Since then I've evaluated a few hundred launches for this site, and the same seven failure patterns keep showing up. Consider this the field guide I wish someone had handed me before that discovery call.
1. Waitlist theater
A waitlist is a tool for managing capacity. Waitlist theater is a waitlist deployed as a marketing prop — manufactured scarcity meant to generate FOMO and harvest emails while the product gets built (or doesn't).
Telling them apart is easy once you know the tell: real capacity constraints come with evidence the product exists. A genuine GPU-bound video tool will have a demo gallery, docs, and pricing, because paying users are already inside. Theater has none of that — just a hero image, an email field, and a counter claiming 40,000 people are ahead of you.
I've gone deep on this whole phenomenon in Waitlist Culture: Why AI Tools Launch Before They Exist, but the short version for triage: waitlist plus no demo plus no docs equals close the tab.
2. Demo cherry-picking
Every demo is curated. That's fine — nobody films their retries. The red flag is when curation crosses into fiction: heavy cuts inside a single task, outputs that appear faster than the model could plausibly generate them, or a "typical result" that turns out to be the best result from hundreds of runs.
The fast check is reproducibility. If the tool is available, feed it something close to the demo's input and see how far off you land. A 20% quality gap is normal launch optimism. A tool that can't complete the demo's task at all on a second attempt was showing you a lottery ticket, not a product.
And when the tool isn't available to test — when the only evidence is the video itself — count the cuts. One continuous take is a product. Fourteen cuts in ninety seconds is an edit suite.
3. The GPT-wrapper tells
There's nothing wrong with building on frontier models — as of this writing, most useful AI software does. The red flag isn't the wrapper; it's a wrapper priced and pitched as a platform. You're looking for signs that the product adds nothing beyond a system prompt:
- Latency that exactly matches the underlying model's, every time
- Error messages that leak the base model's phrasing or refusal style
- No capability the raw model lacks — no retrieval, no memory, no tool integrations, no domain data
- Pricing that tracks token costs one-to-one, with a margin stapled on
A thin wrapper can still be worth $10 a month if the interface saves you time. It is never worth $99 a month, and the launches most eager to hide what's under the hood are usually the ones charging platform prices for prompt engineering.
4. Pricing hidden until the sales call
I covered this in the 10-minute evaluation framework, but it earns its own entry because it's the most reliable flag on the list. Teams publish pricing when they have paying customers, because customers force the question. Teams hide pricing when the answer is still "whatever we can get."
The exception people always raise is enterprise software, where "talk to sales" is genuinely the norm. Fair. But note what surrounds it: real enterprise products with hidden pricing still show you the product — docs, security pages, case studies, a sandbox. A launch that hides the price and the product is hiding the same thing twice.
5. Benchmarks without methodology
"Outperforms leading models by 34%" — on what test set? Judged by whom? With how many runs? At which settings? A benchmark claim without methodology isn't a data point; it's a slogan with a number in it.
The thirty-second check: look for a linked eval, a paper, a repo, or even a blog post describing the setup. If the claim's only citation is its own press release, weight it at zero. Honest teams publish their harness precisely because they know unverifiable numbers are worthless.
6. The team is a logo wall
Scroll to the footer. If the strongest evidence for the product is where the founders used to work — a wall of prestigious former-employer logos — be careful. Pedigree tells you the team can pass interviews. It says nothing about whether this product works.
What you want instead is evidence of shipping: a changelog that predates the launch, public responses to hard questions, an actual named human answering bug reports. I'll take a two-person team with a six-month public changelog over a ten-logo alumni wall every single time, and I don't think it's close.
7. The roadmap does the heavy lifting
Read the launch copy and mentally sort every claim into "works today" and "coming soon." Some launches, once you do this, turn out to be 80% roadmap. The agent that "will soon" handle your calendar. Integrations "rolling out over the coming weeks." Enterprise features "planned for Q3."
A roadmap is fine. A roadmap wearing the present tense is not. The sleight of hand to watch for is grammatical: "Connects to 200+ tools" (does it? today? all 200?) versus "Zapier integration available now, native integrations planned." The second team respects you. The first is hoping you won't check.
The quick-check table
| Red flag | 30-second check | Innocent explanation exists? |
|---|---|---|
| Waitlist theater | Any demo, docs, or pricing visible? | Rarely |
| Demo cherry-picking | Count cuts; try to reproduce | Sometimes |
| GPT-wrapper pricing | Compare latency and pricing to base model | Sometimes |
| Hidden pricing | Is there a self-serve tier with numbers? | Only for true enterprise |
| Benchmarks, no methodology | Look for linked eval or repo | Rarely |
| Logo-wall team page | Find a changelog predating launch | Sometimes |
| Roadmap in present tense | Sort claims into today vs. someday | Rarely |
Scoring is simple. One flag: proceed with suspicion. Two: the burden of proof has fully shifted to the launch. Three or more: close the tab and let your watch list handle it — if the product becomes real, it'll still be real in a month.
The pattern under the patterns
All seven flags reduce to one question: is this team describing something they built, or something they intend to build? Building is slow and specific; intending is fast and vague. Every tell above — the missing docs, the cut-heavy demo, the number-free pricing page, the present-tense roadmap — is vagueness leaking through a story that wants to sound specific.
The good news is that the honest launches benefit from your skepticism. When you reward the teams that show pricing, publish methodology, and demo live, you're voting for a launch culture that wastes less of everyone's time. That's the trade this site tries to make every week.
Want the flags spotted for you before you spend the email address? The weekly hunt covers what shipped, what half-shipped, and what's still a landing page. Get it in your inbox.
Frequently asked questions
What is waitlist theater?
A waitlist used as a marketing prop rather than a capacity control. The tell is a waitlist paired with no demo, no documentation, and no pricing — real capacity constraints usually come with proof the product exists.
How can I tell if an AI tool is just a GPT wrapper?
Look for latency identical to the underlying model, failure messages that leak the base model's phrasing, no offline or batch mode, and pricing that tracks token costs exactly. Thin wrappers aren't automatically bad, but they should be priced like conveniences, not platforms.
Why do companies hide pricing at launch?
Usually because pricing isn't decided, which means nobody is paying yet. Sometimes it's an enterprise sales strategy, but on a self-serve product, hidden pricing at launch most often signals an unfinished business, not an exclusive one.
Is a cherry-picked demo ever acceptable?
Every demo is curated to some degree. The line is reproducibility: if you can't get anywhere near the demo's result with the same inputs, the demo was fiction. Look for live, uncut sessions as the honest baseline.
How many red flags should disqualify a launch?
One high-severity flag deserves suspicion; three or more of any severity and you should close the tab. Individually each flag has innocent explanations — in clusters, they almost never do.