AI software is the hardest category to evaluate on a marketplace, because the thing that determines its value — whether there is real engineering underneath — is invisible in a demo. A polished interface calling a public API looks identical to a genuinely engineered product until you read the code.
First, Establish What Class of Thing It Is
An application with a user interface that calls a model. Value lives in the interface, the workflow, and the integrations — not the model, which is usually someone else's.
A model — weights you download and run. Value is in the weights and the training pipeline.
API access — a credential to someone else's account. The highest-risk purchase on any marketplace, and rarely an asset.
An automation tool wiring AI into an existing workflow. Value is in the integrations and edge-case handling.
An agent — a system that takes actions autonomously. Value is in the guardrails, and the absence of guardrails is what makes them dangerous.
Sellers blur these deliberately, because "AI system" sounds like more than "a form that calls an API."
The Four Questions That Decide Value
1. Where does inference happen and who pays for it?
Three answers, wildly different economics. Your own hardware — no per-use cost, but you need the machine. A third-party API on your key — you pay forever, so get the per-operation cost. The seller's infrastructure — the worst case, because you have bought a dependency on someone who has just stopped being invested in it.
Establish this before discussing price.
2. Do you get the source?
Without source you cannot fix, extend, or migrate. A hosted app with no source is a subscription with no contract.
3. What is the licence chain?
A fine-tuned model inherits its base model's licence. An app inherits its dependencies' licences. The seller cannot grant rights they do not hold. Ask what base model, and read that licence yourself.
4. Is there evidence it works beyond a demo?
An evaluation set. Test cases. Measured results on data you can inspect. A seller who has done real work has this; one who has not will describe their prompts as "carefully engineered."
Red Flags
"AI-powered" with no model call. Ask which model it calls and what happens when that model is unavailable. Silence is the answer.
A wrapper priced as a product. If the whole system is one prompt forwarded to an API, you are buying a weekend's work. Ask to see the prompt file — its size tells you a lot.
No error handling. Ask what happens when the model returns malformed output. Real software has a path for this; demos do not. In an automation context, silent failure is the expensive kind.
Undisclosed usage costs. Get the current monthly bill, not an estimate.
Cheap for the claimed complexity. A genuine trained model represents real compute spend. Prices that make no sense usually mean the thing is not what it is described as.
Unwillingness to explain the architecture. A builder can describe how their system works in plain language. A reseller cannot.
Verifying Under Escrow
On Escrozon the seller is paid when you confirm receipt (or automatically 30 days after delivery starts if you raise no dispute), so structure the window deliberately:
- Run it on your own machine, from the supplied instructions alone.
- Test on your own data at realistic volume — not the seller's demo inputs.
- Deliberately break it: empty input, wrong format, oversized file, another language. How it fails tells you more than how it succeeds.
- Trace every external dependency and total the monthly cost.
- Confirm the licence chain end to end.
- Read the code, or pay someone to. On any significant purchase, two hours of expert review fits inside the escrow window and is cheap insurance.
- Confirm receipt only when all of that holds.
If the tool turns out to depend on the seller's infrastructure, or carries usage costs that were not disclosed, that is a material misrepresentation, and the deal chat record supports a dispute.
Frequently Asked Questions
Is a tool that uses my own API key a bad deal? No — it is often the honest architecture and it means you control the spend. Just price it as software, and know the per-operation cost before you buy.
How do I evaluate code quality without being a developer? You largely cannot, and this is the category where paying for a review pays for itself. Short of that: check there is a README, that dependencies are pinned, and that the project runs from the instructions alone.
What if the model provider changes their API? It happens regularly. Ask whether the model layer is abstracted so a provider can be swapped. A tool hard-coded to one endpoint has a shorter life than its price implies.
Should I buy or build? Build if the logic is simple and you have the skills. Buy when the value is in integrations, edge-case handling, and testing you would otherwise spend weeks on.
Are AI agents worth buying? Judge them on guardrails, not capability. An agent that can take actions without confidence thresholds, validation, or a human-review path is a liability. Ask specifically what stops it doing the wrong thing.
