6 Signs That AI Product Is Just ChatGPT in a Trench Coat

Somewhere right now, a business owner is paying $149 a month for a tool that is a system prompt sitting on top of the same API I call for a few cents per request.

I know because checking these tools is part of my job. A client forwards me a link to some AI product they're about to buy, asks "is this worth it," and I spend an hour finding out. I've done this enough times that the pattern repeats: most AI products pitched at small businesses cost between $49 and $299 a month, and a surprising number of them are ChatGPT wearing a trench coat and a new logo.

Not all of them. Some wrappers are genuinely worth the money, and I'll get to those. But first, the tells.

The Trench Coat Test

Here's the two-minute shortcut that catches most offenders: take the exact input from the tool's own demo, paste it into ChatGPT, and compare the outputs side by side.

If the structure matches, the tone matches, and the "unique insights" match, you're not looking at proprietary technology. You're looking at a prompt. Someone wrote a few hundred words of instructions, put a form field in front of it, and priced the result at $99 a month.

To be clear: there's nothing illegal or even necessarily wrong with that. The entire question is whether the packaging is worth the markup. To answer that, you need to look closer. That's what the six signs are for.

Sign 1: "Proprietary AI" With Zero Specifics

The landing page says "our proprietary AI engine." The documentation says nothing about a model, training data, infrastructure, or benchmarks. That silence is the tell.

Companies that actually build machine learning cannot shut up about it. They publish model cards, benchmark comparisons, technical deep dives. When a product's deepest technical claim is the word "proprietary," the proprietary part is usually a prompt and a Stripe integration.

It's the same pattern I described in How to Spot a Real Web Developer vs. an AI Website Reseller: vague language about the process is almost always hiding a thin process.

Sign 2: The Output Sounds Exactly Like ChatGPT

Every model has stylistic fingerprints. ChatGPT loves certain sentence rhythms, certain transition phrases, certain ways of hedging. If you use it daily, you can smell them.

So when a "specialized AI built for legal professionals" produces the same rhythms, the same bullet-point cadence, and the same closing summary paragraph you get from a free ChatGPT session, the specialization is cosmetic. I've run this comparison during client evaluations and watched outputs match almost word for word, right down to the same three example scenarios in the same order.

A genuinely fine-tuned or retrieval-backed system produces output that differs in substance: it cites your documents, uses your data, knows things the base model doesn't.

Sign 3: Credit Pricing That Maps to Nothing

Credits are the wrapper economy's favorite trick. "1 credit = 1 generation." What's a generation? How many tokens? Nobody says.

Here's why that matters: the raw API cost of a typical text generation is a fraction of a cent. When a tool sells you 100 credits for $49, the markup on the underlying compute can be 100x or more. Markup itself is fine, every business has one. The tell is when the pricing unit is designed so you can't compare it to anything. Honest products price per seat, per project, or per clearly defined unit. Credits that map to nothing exist to obscure the comparison.

Sign 4: No Model Disclosure Anywhere

Tools with real engineering behind them tell you what they run on. "Built on GPT-4." "Powered by Claude." It builds trust and costs nothing.

When a product hides the model entirely, ask yourself why. In my experience there are two reasons: either they're running the cheapest model available while charging premium prices, or they don't want you to realize you could go direct. Neither is a great sign for you as the buyer.

Sign 5: It Breaks the Same Day OpenAI Does

This one is almost funny. Check the tool's status page or their apology posts, then check OpenAI's public incident history. If the dates line up, you know exactly what's under the coat.

I've watched this happen live: an outage hits OpenAI, and within the hour, half a dozen "proprietary AI platforms" post the same "we're experiencing degraded performance" notice. No dependency disclosure needed. The outage calendar is the disclosure.

Sign 6: The Demo Only Works on Their Examples

Every wrapper demos beautifully on its own examples, because the prompt was tuned on those examples. The real test is your input: your messy data, your weird edge case, your industry jargon.

A real product has logic around the model: validation, retrieval, error handling, domain rules. A trench coat has a prompt, and prompts tuned for the demo fall apart the moment reality shows up. If the sales flow keeps steering you back to their sample data, that steering is the answer.

Wait, Is Claude Just a ChatGPT Wrapper?

This comes up in almost every conversation I have about AI wrappers, so let's clear it up: no. ChatGPT is OpenAI's product built on OpenAI's models. Claude is Anthropic's product built on Anthropic's models. Different companies, different models, trained separately, direct competitors.

The distinction people are actually reaching for is AI-native versus AI wrapper. An AI-native company trains and operates its own models: OpenAI, Anthropic, Google. A wrapper calls those models through an API and adds a layer on top. That layer can be worthless or genuinely valuable, which brings me to the uncomfortable part of this post.

The Wrappers I Happily Pay For

Confession: I pay for several tools that are, technically, AI wrappers. After ten paragraphs of roasting the category? Yes. Because the trench coat isn't the crime. The empty trench coat is.

What earns the markup, in my experience, is one of three things:

• Workflow. The tool saves me from rebuilding the same context every time. It remembers my projects, my preferences, my history. Recreating that in ChatGPT would cost me ten minutes per session, every session.
• Integration. The output lands where it needs to go: my CMS, my inbox, my codebase. The model call is 10% of the value. The plumbing is the other 90%.
• Data. The tool brings information the base model doesn't have: live SERP data, my analytics, industry databases. That's not a prompt. That's a moat.

I walked through my actual stack in The AI Tools That Actually Run My One-Person Agency, and most of what survived the year does at least two of these three. My rule of thumb: a wrapper is worth paying for when reproducing the outcome yourself takes real work, not just a saved prompt.

If You're Building One: The Moat Is Never the Model

A quick word for the other side of the table, because half the people reading this are probably thinking about building an AI product rather than buying one.

Everyone has access to the same models you do. That's the whole problem. Your moat has to live somewhere else: proprietary data, deep integrations, distribution, or a workflow so well-designed that people would keep paying even if you swapped the model out tomorrow. If your entire product is a prompt, OpenAI's next feature release can delete your business in an afternoon.

And while you're at it, secure the thing. As I wrote in Everyone's Selling AI Shovels. Nobody's Checking If Their Own Barn Is Locked., the AI gold rush produces a lot of products where the security effort matches the engineering effort: minimal.

What I'd Do Before Signing Up

Run the trench coat test. Count the signs. Three or more, and you're better off with a ChatGPT subscription and a saved prompt.

But be honest with yourself about the other direction too. If a tool fails the test and still saves your team five hours a week because of the workflow around it, then the packaging is the product, and that's a legitimate thing to pay for. Just pay for it knowing what it is.

And if you're not sure whether you need any of this in the first place, I made an honest flowchart for deciding whether you need an AI consultant that applies the same thinking to services instead of software.

Evaluating an AI tool for your business and want a second pair of eyes before you commit? Let's talk. I do this for clients regularly, and an hour of checking is usually cheaper than the annual subscription.