The AI agent demo worked. Here is what breaks between week one and week three.

Each AI agent demonstration succeeds in impressing participants by responding to questions using a limited set of documents. However, when those demos transition to real-world scenarios involving actual customer interactions, payments, and queries, they begin to struggle in ways the initial showcase did not reveal. Building AI agents for business applications through neu Intelligence, the author emphasizes the significant engineering challenges that arise during this transition.

The checklist the author recommends before labeling an AI agent prototype as a usable product includes:

1. Data layer quality: The performance of the agent is heavily influenced by the quality of the data it accesses. Unlike demos that use curated, high-quality PDFs, production systems often contain messy CRM data, conflicting spreadsheets, and unstructured communication channels like WhatsApp. Mapping out every system the agent interacts with and establishing clear authoritative sources is crucial. Inconsistencies between sources can lead to the agent generating incorrect or misleading information.

2. Define clear job descriptions: Rather than simply providing an agent with a feature toggle, it should have a written contract outlining its end-to-end responsibilities, the systems it can access, the actions it can perform independently, and when it must involve human oversight. Clearly defining these boundaries helps prevent "hallucinations" or missteps that occur when an agent attempts to operate outside its authorized scope.

3. Start with shadow mode: The safest initial deployment involves the agent performing all tasks except for taking any action, such as making payments or sending communications. The agent drafts replies, which are then reviewed and approved by a human operator. This approach allows for measuring the agent's error rate using real-world inputs before gradually increasing its permissions.

By establishing a least privilege model based on empirical evidence, teams can more accurately assess the agent's capabilities and limitations.

4. Establish a clear escalation path: At some point, the agent will encounter a query or circumstance it has never seen before. The design choice regarding how to handle such situations—whether to freeze, provide a guess answer, or hand off the issue to a human with full context—is a critical design decision. Prioritizing the development of a robust escalation path ensures that the agent can gracefully handle unknown scenarios, making it production-ready sooner than an agent that lacks this capability.

5. Monitor cost curves: While a single pilot query may have no cost implications, the true measure of an agent's efficiency lies in its cost per completed task as the volume of interactions grows. Factors such as retries, handling large context windows, and frequent tool interactions can lead to escalating costs if not properly managed. Tracking these costs from the outset and treating cost reduction as an ongoing feature enables teams to optimize the agent's performance and scalability.

6. Ensure ownership and control: In cases where the agent malfunctions outside of regular business hours, it is essential to have the ability to diagnose and rectify the issue promptly. This means having the necessary access to logs, the ability to modify prompts and tool configurations, and the authority to make changes without requiring external vendor permissions.

Ownership of the agent's code, prompts, schemas, and integration details should reside within the organization that controls the system. Understanding what is retained when a contract ends is crucial in determining the true value of the AI agent solution.