How to Spot a Fake AI-Agent Vendor

"AI agent" has become one of the most overused labels in the industry. Plenty of vendors are selling a chatbot with a system prompt and calling it an agent. It is worth asking a few direct, technical questions before signing anything — the answers usually tell you which one you are dealing with.

Five Questions to Ask

1. "What happens when the AI is overloaded?" A vague "our AI never gets overloaded" is a red flag — no system is infinitely scalable. A credible vendor describes actual fallback behavior: queueing, rate limiting, a backup model, or a defined degradation path.

2. "Show me how the AI decides." Being shown a polished, pre-scripted demo is not an answer. A credible vendor can walk you through an actual reasoning or tool-use trace — which steps the agent took, which tools it called, and why.

3. "How large is your context window, in practice?" "Unlimited memory" is marketing, not an architecture. A credible answer is concrete: a token count for the context window, plus how retrieval (typically a vector database) extends beyond it.

4. "Does it work with our existing tools?" A blanket "yes, with everything" deserves scrutiny. A credible vendor can hand you API documentation or a list of supported integrations, not just a verbal assurance.

5. "What can it do that a general-purpose chat assistant can't?" Vague deflection — "it's customized" with nothing more specific — is a warning sign. A credible vendor names concrete use cases for your business and can show what changes when the agent, rather than a person, performs the task.

Warning Signs Worth Taking Seriously

• Talk of "proprietary secret technology" with no willingness to explain, at a reasonable level, how it actually works.
• Pricing denominated in obscure "AI units" or "credits" instead of a comparable, transparent cost basis.
• An inability to explain the architecture, or to share even a high-level diagram of how a request flows through the system.
• No answer to what happens to your data once it leaves your systems.

What Is Worth Paying For

A system that genuinely learns from your own company data rather than a generic model with your logo on it. One that performs real actions — not just conversation — and can demonstrate that behavior with a traceable sequence of steps, not a scripted showcase.

If you want to see what a transparent architecture looks like end to end, KeepUrAi documents its own request flow and model choices openly, and the hosted edition at qcg-ai.ch is a Swiss-operated service you can evaluate directly rather than take on faith.