By Bram van Zanten, CEO of AssistYou

The core challenge in building enterprise-grade voice AI is mastering a paradox. On one side, you need to deliver a natural, human-like conversational experience. On the other, you must maintain absolute, uncompromising enterprise control over the flow of data.

Right now, there is a dangerous misconception in many C-suites that you can simply take a Large Language Model (LLM), connect it to your company’s APIs, and deploy it to handle customer service.

In a sandbox, this looks like magic. In production, it is a liability.

Out of the box, LLMs do not inherently understand the operational guardrails of your business. If you want reliable, safe, and scalable performance, you cannot rely on the model’s default reasoning. You have to engineer the boundaries yourself.

At AssistYou, we achieved this by building our architecture entirely around the concept of the Schema.

What is the Schema?

Think of the schema as a highly optimized dialogue blueprint. Under the hood, it is a complex JSON file that dictates exactly where the AI agent has conversational freedom and where it is strictly locked down.

Every agent we build operates within these predefined schemas. It is the layer that sits between the conversational AI and your backend systems, ensuring that human unpredictability never breaks your enterprise logic.

Context Dictates the Rules: Integrating the Geovalidation API

A voice agent cannot operate on a one-size-fits-all set of rules because different industries require entirely different levels of control. The schema allows us to define those specific rules of engagement, frequently connecting to a core geovalidation API to parse location data accurately based on industry needs.

The Utility Paradigm: If you are a utility company processing a change of address, the process is rigid. The agent cannot guess or accept vague landmarks. The schema forces a strict lookup through the geovalidation API, capturing a highly structured, verified street name and house number before moving forward.

The Mobility Paradigm: If you are a taxi company, the rules change entirely. A caller might request a ride to “the hospital in Rotterdam.” Here, the schema instructs the geovalidation API to accept this Point of Interest. If there are three hospitals in Rotterdam, the schema allows the agent to dynamically clarify which one the caller means, rather than outright rejecting the input.

Without a schema defining these boundaries and managing the geovalidation API integration, the LLM treats both interactions exactly the same, resulting in failed API calls and frustrated customers.

Protecting Your APIs Through Strict ID&V

Perhaps the most critical function of the schema is protecting your backend systems from the AI itself and malicious external actors.

Consider the Identification & Verification (ID&V) process. How do you identify a customer securely? You cannot simply hand your CRM API over to an LLM and tell it to figure it out. A poorly guarded but naturally conversational LLM could be manipulated into brute-forcing your CRM; using repeated verification attempts to uncover a customer’s house number the attacker never knew to begin with.

The schema prevents this. It acts as a gatekeeper, explicitly instructing the AI that it must collect all required, verified data points from the user first. Only when the schema’s conditions are perfectly met is the system allowed to execute the call to your API.

The Balance of Freedom and Scripting

Even when it comes to the conversation itself, control is paramount. We want the LLM to have the freedom to speak fluidly and adopt a brand’s unique tone of voice when answering complex questions.

However, certain moments in a customer journey carry legal or brand-critical weight. The schema allows us to completely hardcode elements like the opening statement or compliance disclaimers. The AI is not allowed to generate its own greeting. It must read the exact legal script required by the enterprise before smoothly transitioning back into natural language for the rest of the call.

Architecture Beats Raw Intelligence

The future of voice automation is not about finding a bigger, smarter model to do everything at once. It is about building the right structural boundaries around the models we have.

By utilizing schemas, we ensure that the AI is brilliant exactly where it needs to be, while keeping the enterprise firmly in the driver’s seat.

← Back to blog