A note for decision makers

Just admit it.

Every other page on this site is written the way these sites are usually written: confidently, by a language model, and without anyone checking. This page is the exception. It is for the person being asked to approve, buy, or sign up for something like it.

What this kind of site actually is

Someone with no development background asked an AI coding assistant to build an application, pasted the result into a folder, and put it online. That on its own is fine. The problem starts when the same person calls themselves a developer, describes the result as enterprise grade, and cannot explain what any of the code does when asked. If the answer to "how does this work" is a change of subject, that is the answer.

Where your data goes

An app like this does not "use AI" in any contained sense. It sends whatever you type or upload to a third party's model over the internet, often through a personal account on a consumer plan, and shows you the reply. Pricing, client lists, proposals, contracts and staff records all leave the building the moment they touch the box.

What happens to it after that depends on the account type and the terms in force at the time. Depending on those terms, the provider may keep it for weeks for abuse monitoring, have staff review it, or use it to improve future models. You cannot verify which applies, and neither can the person who built the app, because they have not read the terms and the terms were never written with your compliance obligations in mind. There is typically no contract between your company and the provider, no data processing agreement, no audit log, and no way to get the data back or have it deleted. If you or your business puts sensitive or proprietary information into a system like this, treat it as leaked.

The "AI" in an app like this is, in almost every case, a thin wrapper around one of three companies' models: OpenAI, Anthropic or Google, usually in that order of popularity. Nothing was trained, tuned or built. A request goes out with your data attached and an answer comes back. Every one of those models is vulnerable to prompt injection, where text inside the data (an email, a document, a form field) is read by the model as an instruction. The UK's National Cyber Security Centre has said plainly that this may never be fully fixable, because these models cannot separate data from instructions. It has already been used to pull data out of a production system: in 2025 a single crafted email was enough to make Microsoft 365 Copilot hand over a user's internal documents with no click from the victim (CVE-2025-32711). If an app like this keeps what you shared so it can refer back to it, and other people use the same app, that data can be extracted by anyone who phrases a request the right way. The builder will not have thought about this, because the builder does not know it exists.

Which assistant was used does not change any of this. Claude, ChatGPT, Cursor, Lovable, something newer: picking one over another is a brand preference, not a qualification. Someone who used the better regarded tool has not found a secret the rest of the industry missed. The tools produce similar output, and the "creators" share the same inability to read or understand it.

What is usually missing

Off-site backups. Encryption in transit: if the address bar says "Not secure", every password and every document typed into the site crosses the network as plain text, readable by anyone on the same wifi. A certificate is free, takes one click in any hosting control panel, and has been switched on by default almost everywhere for years; a site is still on HTTP because nobody knew to look. Passwords hashed rather than stored as text. Access checks on the server rather than in the browser, where anyone can change them. A second person who has read the code. A plan for what happens when the single server it runs on stops. None of this is a guess. It is the pattern for how these apps handle other people's information, and it is absent because the person did not know it existed.

What to ask before you approve it

  • Which company receives our data when the AI feature runs, and do we have a contract with them?
  • How has the code been reviewed for security and compliance, and by whom?
  • Where are the backups, and when did you last restore one?
  • What happens to the data we have already put in if we stop using it?

If the answers are vague, treat the product as a demo, not a system. Do not put real client or staff information into it until someone who can read the code has done so and written down what they found.

How to recognise one

The tells are consistent because the same tools produce them. Purple gradients on a dark background. A small pill with a pulsing dot above every heading. Cards that lift and glow when hovered. Emoji used as icons. Several Google fonts at once. Em dashes in every paragraph and in every customer review. Reviews from people with common names and no links to anything. A review that does not link to a Google Business Profile, Trustpilot, or some other account you can click through to is worth nothing at all; anyone can type five stars and a first name, and a model will type ten of them in one go. Counters that animate up to round numbers. A dark mode toggle that does nothing. A placeholder such as "[Your City]" or "YourSiteName" left in the footer. A founder page that becomes defensive when the question of how it was built comes up.

If you are an employer and someone on your team who you did not employ as a developer brings something like this to you, be cautious. Ask them to walk you through the code and its security measures. Ask for a portfolio of genuine work that shows they have built and understood systems like this before.

Further reading

Independent sources, none of them selling anything. Read these before you approve an AI product built by one person.

This site was built by Claude, on request, to make the point that anyone can do this in a day. The person who asked for it will tell you so. That is the difference. Using AI is not the problem. The problem is nobody being honest about it, and nobody competent checking the result before it was pointed at real data.

Re-read the security page →Back to the site