-
-
AgentTrust IQ dashboard and reliability workflow: citations, eval scoring, audit logs, guardrails, and performance metrics.
Inspiration
Small businesses want the speed of AI customer support, but most cannot trust a black-box chatbot that may hallucinate, expose private customer data, or give unsupported answers. A local clinic, repair shop, tutor, consultant, or ecommerce seller needs AI that can answer from business knowledge, cite evidence, avoid leaking PII, and show exactly why an answer was allowed or blocked.
AgentTrust IQ was built to solve that trust gap.
Instead of only building an AI agent that answers questions, AgentTrust IQ builds an AI agent that checks itself.
What it does
AgentTrust IQ is a Gemini-powered reliability layer for small-business AI support agents.
A customer asks a question. The support agent retrieves relevant business knowledge, generates a grounded answer, cites its sources, and then passes the response through a second reliability-checking agent. The evaluator checks whether the answer is supported by evidence, whether citations are present, whether private information is exposed, whether the prompt contains injection attempts, and whether the response should be allowed, revised, or refused.
Each interaction produces an audit log with reliability signals such as:
- Citation support
- Hallucination risk
- PII exposure
- Prompt-injection risk
- Refusal correctness
- Latency
- Cost per request
- Final pass/fail decision
The goal is to help small businesses deploy AI support agents that are fast, useful, and accountable.
How we built it
The product is built as an AI-native support workflow:
- A Gemini support agent receives a customer question.
- A retrieval layer searches business FAQs, policies, product/service information, and support docs.
- The agent generates an answer with citations.
- A Gemini evaluator agent scores the answer for groundedness, safety, and trust.
- Guardrails check for prompt injection, unsupported claims, and PII exposure.
- A JSONL audit log stores the full reliability trace.
- A dashboard shows pass rate, hallucination risk, latency, and customer-support quality metrics.
This turns AI support from “hope the answer is right” into a measurable business system.
What makes it AI-native
AgentTrust IQ does not just use AI as a feature. AI runs the core business workflow.
The product uses AI agents for customer response generation, evidence retrieval, reliability evaluation, policy checking, and support-quality monitoring. Human operators review failures and customer feedback, while the AI handles the repeatable support workflow.
This makes the business scalable: one small business owner can support more customers without hiring a full support team.
Business model
AgentTrust IQ is designed as a subscription product for small businesses.
Initial pricing:
- $49/month for a hosted AI support agent
- Reliability dashboard
- Business FAQ/document setup
- Audit logs for customer interactions
- Safety checks for PII, citations, and unsupported answers
The first target customers are small service businesses, clinics, tutors, local agencies, ecommerce sellers, and consultants with repetitive customer questions.
Impact
Small businesses often lose customers because they cannot respond fast enough. They also cannot afford enterprise AI governance tools.
AgentTrust IQ gives them a practical middle layer: an AI support agent that answers quickly, cites its sources, blocks unsafe behavior, and gives owners proof of what happened.
The impact is simple:
- Faster customer responses
- Lower support burden for small teams
- Fewer unsupported AI answers
- Better privacy protection
- More trust in AI-assisted business operations
Challenges we faced
The hardest part was not generating answers. The harder problem was deciding whether an answer should be trusted.
A normal chatbot can sound confident even when it is wrong. AgentTrust IQ needed a second layer that checks evidence, detects weak citations, blocks unsafe requests, and produces logs that a human can inspect.
The project required balancing speed, reliability, cost, and usability. Small businesses do not want a complex AI governance platform. They need a simple product that works.
What we learned
The biggest lesson was that useful AI products need measurement. A demo answer is not enough. A business owner needs to know whether the answer was grounded, whether citations were valid, whether private data was protected, and whether the agent is improving over time.
AgentTrust IQ was built around that lesson: every AI answer should leave behind proof.
What’s next
Next steps are to onboard pilot businesses, collect real customer questions, measure reliability metrics, and convert the product into a paid small-business AI support service.
The target pilot metrics are:
- 5 pilot businesses
- 1–3 paid customers
- 85%+ eval pass rate
- 90%+ citation precision
- Under 8% hallucination rate
- 0 known PII leaks
- Under 3 second p95 latency
- Under $0.05 cost per request
Log in or sign up for Devpost to join the conversation.