AI Slop & ChatGPT Skins
A buyer's guide for firms being pitched AI every week
Listen to this article
AI-generated podcast
Most AI products you're being pitched are worthless. They are a logo and a price tag on a ChatGPT subscription you could buy yourself for £20 a month. The industry has a word for low-effort AI output now: slop. Merriam-Webster made it 2025's Word of the Year. The product version of slop is the ChatGPT skin.
Some of what you are being pitched is worth every penny. Some of it is a skin. And some of what you need is a spreadsheet. This is how to tell which is which.
What We're Actually Talking About
A "GPT wrapper" is an application that does little beyond calling someone else's AI model. The vendor adds a user interface, maybe some prompt templates, and charges you monthly for access to capabilities you could get directly from ChatGPT for £20. The technical community calls these "ChatGPT skins" because that's precisely what they are: cosmetic layers over someone else's technology.
The problem isn't that these products exist. The problem is that they're marketed as proprietary AI solutions with enterprise price tags. We have looked under the bonnet of enough of these to say it plainly: most could be rebuilt in an afternoon with a model API, a payment form and a template.
The ground moves under them: OpenAI and Anthropic ship new capabilities every month. Each release makes another batch of wrapper products obsolete overnight.
Crunchbase counted $202 billion of venture money going into AI in 2025, about half of everything invested that year. That much money, that fast, buys a great many demos.
When the AI Claim Did Not Survive Contact
Some of these products were not merely oversold. One ended in insolvency, one in criminal charges, and one in a regulator's penalty for claims that were never tested.
Builder.ai — over $445M raised
Sold an AI assistant, "Natasha", that would assemble your app while you waited. A 2019 Wall Street Journal investigation found most of the coding was done by people. In May 2025 the company entered insolvency, after it emerged that the revenue forecasts behind its emergency loan were about four times what it actually made. Microsoft and the Qatar Investment Authority had both backed it.
Nate — over $42M raised
Sold one-tap checkout run by AI. US prosecutors alleged the app's actual automation rate was effectively zero, with hundreds of contractors in a call centre in the Philippines completing the purchases by hand. The founder was charged with fraud in April 2025.
DoNotPay — "Robot Lawyer"
Faced an FTC enforcement action after never testing whether AI output matched human lawyer quality — a $193,000 penalty.
Three cases, three different ways the same question went unasked: does the thing actually run without people behind the curtain? A demo cannot answer that. A bad day can. This is how Builder.ai raised hundreds of millions for a largely human-powered service with an AI veneer: the people signing the cheques had never seen it run on a bad day. We looked at that gap in an article on why most AI pilots never make it past the demo.
Seven Red Flags to Check Before You Sign
Each one is a question you can ask on the sales call. The answer, or the silence, is the red flag.
"What will be measurably different in three months, and how will we know?"
Agree the numbers before the pilot starts, in writing. Without success criteria set at the beginning, an evaluation can run six months to a year and produce nothing. A vendor who cannot name the number does not expect to hit one.
"Which model does the thinking, and what did you build around it?"
A straight answer sounds like: "Claude, with your order history and your pricing rules wired in." "Harnessing the power of AI" with nothing behind it is the warning sign. If the product's outages, "quota exceeded" errors and price rises track OpenAI's, you are paying a markup on a subscription. That is fine when the markup buys real work around the model, and a red flag when it buys a logo.
"Can you run the demo on one of our awkward cases?"
Simple demos can make solutions seem incredibly capable. What matters is how the product deals with real-world exceptions: the order with a missing field, the customer who writes in a hurry, the supplier who changed their format. If the demo could be replicated in ChatGPT, so could the product.
"What would happen if OpenAI built this feature tomorrow?"
If the honest answer is "we'd lose," the product has no moat. The vendors worth paying have built something a model provider cannot copy by shipping a feature: your data, your workflow, your systems wired in.
"What were you doing three years ago?"
Ask what vendors were doing several years ago. This distinguishes genuine expertise from companies that stuck AI buzzwords in their marketing after ChatGPT launched.
"Where does our data live, and does it train your model?"
Legitimate vendors readily explain where data resides, what certifications they hold, and whether customer data trains their models. Hesitation here is an answer.
"Which customer's numbers can you show me?"
Request quantifiable case studies. Not testimonials. Ask for one customer's numbers: hours saved, errors cut, revenue added. If all you get is "transformation", they have not measured anything for anyone.
When Wrappers Actually Create Value
Renting a model is not the problem. We rent them too. The question is what sits around the model:
The distinction is not wrapper versus non-wrapper. It is thin wrapper versus software built around how a business actually runs.
Our own MushRoom is a small example of the difference. The assistant it talks to is the same one anyone can subscribe to. What makes it worth paying for is everything wired around that assistant: the bookings, the cleaning rota, the guest registrations and the tax records of a short-term-rental business, so hours of a host's admin turn into minutes. Nobody rebuilds that in an afternoon, because the afternoon was never the point. Knowing how a host's week actually goes was.
Harvey did the same thing in law at a far bigger scale. It did not train a foundation model of its own; it built the legal data, the security and the workflow around one. The value of the application is how it delivers value, not whether it calls an external API.
Why a Software Studio Is Telling You This
We build AI into businesses for a living. Here is what that looks like from our side.
We build around how your business runs
The model is rented; the value is in the wiring: your orders, your rota, your customers' history.
Sometimes a ChatGPT subscription is enough
If all you need is ChatGPT with some custom prompts and team training, we'll tell you that—and help you set it up properly instead of selling you something more expensive.
Sometimes you don't need AI at all
If your problem is better solved with a spreadsheet, a simple script, or a process change, we'll tell you that too. We'd rather solve your actual problem than sell you technology you don't need.
Being straight about the cheap answer is how a studio earns the call about the expensive one.