9 Questions That Expose Fake AI on a Product Roadmap

A checklist exists to verify the authenticity of AI claims on a product roadmap. Nine questions help determine whether a feature qualifies as an autonomous agent or merely a rules-based system. If a feature fails any question, the label must be removed.

1. The feature must make a decision without human intervention. An agent acts and accepts the consequence, while a chatbot only returns text and waits for a person to act. Ask the team for an example of an action the system takes on its own. If the answer is only suggesting, it's a recommendation feature, not an agent.

2. Determine how the system behaves with unseen inputs. A real agent degrades gracefully, while a static feature throws an error or returns a default. Testing with an unexpected input can expose fake AI.

3. The system should be able to explain its reasoning, not just what it did. A model providing a confidence score or ranked features is acceptable, but if the answer is a flowchart, the intelligence resides in the flowchart, not the model.

4. The system must adapt to changing data. If the same behavior persists even after feeding new outcomes, the system lacks learning capabilities. Real agents evolve with data, while static configurations do not.

5. Trace the model's location in the request flow. Real AI features involve the model in meaningful decision-making steps. If the model's output never changes the outcome, it's likely just decorative.

6. Understand the cost per action and track it. Real AI use comes with a real bill. If a vendor cannot provide the cost per transaction, it may indicate limited model usage or lack of monitoring.

7. Test the system's ability to handle multi-step tasks. Agents can chain steps and recover from errors without restarting, while scripted flows break upon unexpected inputs. Try giving the system a task with multiple steps, ensuring it adapts to surprises.

8. Consider the data preparation costs. Real machine learning relies on clean, labeled, structured data. If a team claims sophisticated agents but neglects data preparation, the claim may be hollow. Ask about the training and grounding data process and who cleaned it.

9. Have an architect review the code. The final filter. A clean, labeled roadmap will withstand architectural scrutiny. A team that won't let engineers inspect the code may be masking the true nature of their AI features.

To run the detox, review the current roadmap and answer the nine questions for each AI, agent, or intelligent labeled feature. Re-tag the features into three categories: Agent, Model-assisted, and Rules-based. Agents make unsupervised decisions, adapt to data, and involve a real model in the decision path, with measurable inference costs. Model-assisted features use a model for a real step but require human decision-making. Rules-based features use deterministic logic and are often the right tool for the job.

This detox process separates true AI agents from mere marketing gimmicks, helping stakeholders make informed decisions about funding and development.