Escrozon Buyer's Handbook

Buying an AI chatbot or AI tool: nine questions to ask the seller

AI software · About 8 minutes to read

Most AI chatbots sold by independent developers are a layer of code around a model rented from a provider such as OpenAI, Anthropic or Google, plus a set of prompts and often a knowledge base. That is not a criticism: it is how most AI products are built. It does mean that much of the value depends on things a demo will not show you. These nine questions bring them out.

1. Whose API key does it run on?

A chatbot that calls a hosted model needs an API key. After the sale it must run on your key: billed to your account, under provider terms you have accepted yourself. A tool that only works on the seller's key is a service you are renting from the seller, not a product you own. Ask to see where the key is configured, and make swapping in your own key part of the handover.

2. What does one conversation cost to run?

Hosted models charge by usage, measured in tokens, so every conversation has a running cost. Ask for the average cost per conversation, or per thousand requests, and check it against the provider's current price list. A bot that looks cheap at demo volume can lose money at scale, especially if it uses an expensive model or sends a large chunk of the knowledge base with every message.

3. Which model, and what happens when it is retired?

Providers retire older models on published schedules. See OpenAI's deprecations page and Anthropic's model deprecations page for examples. Ask which model and version the tool uses, whether it has been tested on a newer one, and how much work a switch would take. Prompts tuned for one model often need rework on the next.

4. Can I test it on my own questions?

The seller chooses the demo questions. Write twenty or thirty of your own from your real use case, and run them before you pay. Include awkward questions and ones the bot should refuse. Note wrong answers, invented facts, and any answer that reveals the hidden instructions. Escrozon's guide to evaluating an AI model before purchasing shows how to build a simple test set.

5. What is in the knowledge base, and who owns it?

If the chatbot answers from documents, those documents are part of what you are buying. Check that the seller has the right to pass them on. Scraped websites, client files and paid reports often cannot be resold.

6. Where do the conversations go?

Chat logs often contain personal information. Ask where they are stored, for how long, and which outside services see them: the host, analytics tools and the model provider. After the sale you are responsible for them, and if you serve customers in the EU or UK, data-protection law will expect you to be able to answer these questions.

7. Hosted API or open-weights model?

Some tools run an open-weights model on their own server instead of calling a hosted API. That removes per-token fees and dependence on one provider, but it adds server costs and upkeep. Open-weights licences also differ: some limit commercial use, and some add conditions for very large companies. Ask which licence applies. The trade-offs are laid out in open-source vs proprietary AI.

8. What exactly is included?

Get a written list:

“Lifetime licence” on a script usually means you may use the code indefinitely. It rarely means exclusive rights, and it never covers the model provider's fees.

9. Is there revenue, and can I see it at the source?

If the tool is sold as a business with paying users, treat it like any SaaS purchase. Check the revenue in the payment processor's own dashboard, not in screenshots. The guide to pricing source code and running SaaS covers how.

Before you pay

Escrozon has a category for AI chatbot scripts and tools, most of them sold as source code with a licence. Wherever you buy, keep the payment in escrow until three things are true: the code and prompts are in your hands, the tool runs on your own API key, and your test questions have been answered to your satisfaction.