Controlling AI Responses: A New Approach to Fix LLM Hallucinations in Customer Conversations

The Challenge of Hallucinations in Customer-Facing AI

In conversations with enterprise technical leaders, a surprising question arose: "Can we use Parlant while turning off the generation part?" At first glance, this seemed contradictory—how can a generative AI function without generation?

High Stakes Demand Precision

These AI agents were not experimental but intended for millions of users monthly. Even a 0.01% error rate could lead to compliance issues, legal risk, or damage to a brand. Traditional large language models (LLMs), despite advances, still produce uncertain outputs like hallucinations, tone mismatches, and factual errors.

Rethinking Generative AI

The enterprises had skilled Conversation Designers who craft agent behaviors, responses aligned with brand voice and legal standards, and design engaging interactions. Their goal was not to avoid generation due to fear but to exert precise control over AI outputs. Generative AI agents are better understood as adaptive systems responding contextually and intelligently. Whether responses come from LLMs token-by-token or a curated bank, the priority is appropriateness, compliance, clarity, and usefulness.

Conversation Designers: The Key to Eliminating Hallucinations

Instead of patching generative models, integrating Conversation Designers into AI development offers a path to eliminate hallucinations. These professionals introduce clarity, intentionality, and engagement beyond what foundational LLMs achieve alone.

Introducing Utterance Templates in Parlant

Inspired by these insights, Parlant implemented Utterance Templates, allowing designers to create fluid, context-aware, yet fully vetted and governed response templates. This three-step process involves:

Drafting a fluid message based on context and guidelines.
Matching the draft to the closest template from the utterance store.
Rendering the template with variable substitutions using Jinja2 format.

This approach balances LLM-like adaptability with strict control, enabling developers and conversation experts to collaboratively build reliable, trustworthy AI agents.

Empowering Human Expertise

The future of conversational AI lies in empowering those who understand brand, customer, and legal requirements to shape AI communication. Parlant provides tools that put control into the hands of the right people, ensuring AI interactions are safe, compliant, and effective.

Final Thought

Controlling or limiting generation in customer-facing AI is not a limitation but a necessary evolution to trust and reliability in real-world applications.