Every consumer AI app has one sentence in its marketing that the model cannot fully back. The builder's job is to find that sentence before a one-star review does. I build a photo-identification app, and most of my marketing decisions come down to one question: what can an image honestly tell a person, and what must we refuse to imply?
A photo shows a surface, and buyers hear a verdict
A camera captures what is visible from one angle in one light. An image model can be useful at describing that. It cannot see inside a wall, behind a cabinet, or under a floor. The Minnesota Department of Health says visual inspection "cannot detect mold hidden within wall cavities, inside HVAC ductwork, or beneath flooring." Guidance from the EPA, the CDC, and the same Minnesota agency is the source for the commonly cited range that a visual-only inspection catches 30 to 50% of real contamination.
A photo tool inherits that limit. Marketing that hides it is a promise the product breaks the first time the problem is out of frame. The gap between what you can photograph and what is actually going on is the subject of an earlier piece of mine, Check moisture before you close the wall, and it applies to software as much as to drywall.
Trust design is a claim-writing rule
The answer screen is where marketing and product meet. These are the three rules I would defend on a call.
Show what the model looked at. A verdict with no visible basis asks for faith. A verdict tied to the part of the photo that drove it invites the person to check it.
Put the limits inside the answer, not in a footnote. If the model cannot assess what is behind the surface, the result screen should say so in plain words at the moment someone is deciding what to do next.
Never publish a number you cannot source. At Mold Scanner AI, we keep a written ledger where every statistic needs a named source before it enters copy. A number without a source is marked as needing data, never invented. That is why I will not quote an accuracy percentage. I would rather say less than say something I cannot defend.
The same discipline covers attribution. Brian Karr, an indoor environmental consultant, has said on a podcast that air tests give a false negative seventy percent of the time when there is a hidden mold source in a wall or under a floor. That is his own field observation, not a peer-reviewed study, and the hidden-source condition matters. Strip either qualifier and a fair quote becomes a false claim.
Distribution and pricing are judgment calls before they are data
On mobile you do not own the landing page. The store listing is the ad, the first screenshots are the copy, and a person decides in seconds. I treat the screenshots as the place to show restraint: an answer screen that states its limits reads as more credible than one that shouts certainty. If you are comparing tools in this category, our guide to the best mold detection app explains what to weigh.
On-device versus cloud is also a marketing decision. On-device processing gives you speed and a cleaner privacy story, but a smaller model. Cloud gives you a stronger model and easier updates, but photos leave the phone. Either can be right, and either changes what you may say. You can only write "your photo stays on your phone" if it is true, and your privacy copy must match your architecture line for line.
Pricing a B2C app before you have a crowd is mostly a judgment about who you want in the room. Charging first filters for intent and signals seriousness. Letting people try first lowers the barrier and raises the number of people who never intended to pay. Neither choice is proof of anything. A price is a hypothesis until real buyers push on it. Anchor it against the alternative they already know: Angi's 2026 figures put the average professional mold inspection at $670. A photo app does not replace that. It helps someone decide what to do before or alongside it, and the pricing story should say so.
Small team, small claims
On a small team, the person who ships a feature often writes its store copy too. There is no legal department to slow a bold sentence down. The guard has to live in the process: a written list of approved claims, a source for each one, and a rule that anything off the list does not ship.
That constraint has been good for the product. It forces every promise to be small enough to keep, and buyers reward that more than they reward a big claim that collapses on contact.
The rule I would give any founder: if your model cannot see it, your marketing cannot say it.