← All work
Trust & Safety · ThermaSteel / PetDiets

An AI agent that knows what not to say

Customer-facing chatbots for two clients — a panel manufacturer and a veterinary-nutrition practice. The hard part wasn't answering questions; it was the answers the agent should refuse to give.

ThermaSteel's chat widget: 'Hello! How can I help you today regarding ThermaSteel products, construction, or company services?' PetDiets' chat widget: 'Hi! I'm the PetDiets Assistant. I can help with general nutrition questions and finding the right service. How can I help you today?'
The same agent framework, grounded in each client's own content — ThermaSteel on the left, PetDiets on the right.

A confidently wrong answer is worse than no answer.

A support agent that confidently makes something up is worse than no agent — especially in vet nutrition, where a wrong answer about an animal's diet can cause real harm.

Retrieval-grounded, not open-ended.

Playbook-driven agents (Dialogflow CX) grounded in a Vertex AI Search data store scoped to each client's own site — so answers come from real content, not the model's guesses.

Grounding over fluency.

I'd rather the agent say "I don't have that" than improvise. The agents are constrained to their data store and route users to real resources instead of inventing specifics, so confidence never outruns what the source material actually supports.