AI wrappers vs real products, and how to spot a SaaS worth acquiring

Roughly half the listings hitting Failedups now have “AI” somewhere in the description. That number was about 15 percent two years ago. Most of these projects are not real AI products. They are wrappers, and the difference matters a lot when you are writing a check.

Here is how I separate the two when I am scanning a listing.

Define your terms before you read the listing

A wrapper, in my book, is a thin UI on top of a public LLM call. No proprietary data feeding the prompt. No fine-tuning. No evaluation system. No agent loop. The user types a thing, the app forwards it to OpenAI or Anthropic with a system prompt the seller wrote in 20 minutes, and the response comes back. That is the entire product.

A real AI product solves a job that the user actually has, where AI is one ingredient among several. The user puts something in. The system improves that input over time using their data, their corrections, their workflow context. There is a moat somewhere: workflow integration the user cannot easily migrate, data the buyer cannot recreate from scratch, distribution that compounds, or a brand the niche actually trusts.

You can run a wrapper to $5k MRR with vibes and a Twitter account. You cannot run a real AI product without solving an actual problem.

The acid test

Before I read anything else in a listing, I run the acid test:

If OpenAI ships this exact feature for free in their next dev day, what happens to your acquisition?

If the honest answer is “the project is worthless,” you are buying a wrapper. Pay accordingly.

If the honest answer is “we still have the customers, the data, and the workflow integration, the model is just one piece,” you might be looking at a real product. Now do the rest of the work.

This is not a hypothetical. In the last 18 months, OpenAI alone has shipped voice mode, structured outputs, file search, computer use, and a dozen smaller features that vaporised hundreds of YC batch projects. The platforms eat their wrappers. They always have. The only question is whether the project you are looking at is on the menu.

Red flags I see in wrapper listings

Some of these are obvious. Some look respectable until you press on them.

Single-prompt apps. The entire product logic lives in one system prompt. You can read it in 30 seconds. Whatever the founder spent six months on, it was not the prompt.
ChatGPT clones with a coat of paint. Same UX as the official app, slightly nicer typography, “for marketers” or “for lawyers” stapled on the front. The vertical wrapper. These were a 2024 phenomenon and most are now dead.
“AI for X” with no domain expertise. The founder built a SaaS for radiologists or tax accountants but cannot tell you a single thing a working radiologist does in their day. The prompt was written from a Google search.
No evaluation suite at all. Ask the seller how they measure output quality. If the answer is “I tried it a few times and it seemed fine,” there is no quality bar. Buyers cannot improve what was never measured.
No caching or cost awareness. Every request hits the API at full price. The seller has no idea what their per-user cost is. This is the engineering tell that there has been no real production thinking.
Free trial only, no paying users. People will try anything once. Conversion to paid is where wrappers die, because once the novelty wears off, users go back to ChatGPT directly.

If you see three or more of these, you are buying a wrapper. The math from there is simple: the project should sell at near-zero premium over the code’s replacement cost. Maybe a $3k codebase plus a $1k domain. Total $4k. Not $40k.

Green flags in real AI products

The flip side is rarer, but worth knowing on sight.

Custom RAG over data the buyer could not easily replicate. Scraped industry data, user-uploaded documents, an integration into a niche tool’s API that took weeks to figure out. The retrieval layer is the moat, not the model.
Agent loops with measurable outcomes. Multi-step workflows where the system completes a job, logs the result, and the seller can show you the success rate over time. Bonus points if the agent learns from corrections.
Fine-tuned models on a niche dataset. Even small fine-tunes on a few thousand examples beat raw prompts in narrow domains. If the seller can hand over training data and a tuning pipeline, that is real.
Workflow depth that creates switching cost. The product is wedged inside how the user already works: their Slack, their CRM, their daily report. Migrating away costs the user time, not just money.
A brand the niche has heard of. Three Reddit threads, a podcast mention, a newsletter that drives traffic. Distribution is the most underpriced moat in AI right now.

A project with two or three of these can credibly command 2 to 3 times the replacement cost, because the buyer is paying for the unfair advantages, not just the working code.

What I actually do when valuing an AI listing

Here is the worksheet I run, in order:

Apply the replacement cost framework to the codebase, ignoring the AI angle entirely. What would it cost to rebuild the app shell?
Run the acid test. If the project fails, stop here. The price ceiling is replacement cost, often less.
If it passes, count the green flags. Each one nudges the multiplier up. Two green flags gets you to 1.5x. Three or four gets you toward 2x or 3x, but only with traction to back it up.
Subtract a “model risk” discount. Even good AI products carry the risk that the underlying model gets cheaper, smarter, or eaten. I knock off 15 to 25 percent for this. Buyers in 2026 should be honest about it.

The number you are left with is roughly fair. If the seller is asking double, you walk. If they meet you somewhere in the middle, you have a deal.

The market is full of AI listings right now. Most are wrappers. The real ones are out there, but you have to read carefully, ask the boring questions, and refuse to pay platform-risk prices for platform-risk products.

Browse the current active listings and see how many pass the acid test. The number will surprise you.