Understanding grounded answers in IrisAgent

What grounding means, which sources IrisAgent checks, and when it refuses to answer.

A grounded answer is built only from knowledge connected to your account (articles, tickets, procedures, and other indexed sources) and cites those sources. If IrisAgent cannot find support for an answer, it declines or asks a clarifying question instead of guessing.

​

How grounding works

For each question, IrisAgent:

1. Retrieves relevant articles, tickets, or procedures.
2. Generates the reply from that material only.
3. Shows the sources so a person can check the reply.

This loop (retrieve, generate from sources, refuse when support is missing) is how IrisAgent reaches validated answer accuracy above 95% on its accuracy evals.

​

Sources IrisAgent checks

Source type	Examples	Docs
Ticketing system	Zendesk, Salesforce, Intercom, Freshworks, HubSpot, Jira Service Desk, Zoho Desk	Connect your helpdesk
Built-in helpdesk KB	Help center articles synced from your ticketing system	Ticketing System KB
Knowledge bases	Confluence, public websites, uploaded CSV/PDF	Confluence, Public Websites, Upload Content
Procedures and workflows	Procedures and automations you enable	Onboarding, Workflows

IrisAgent uses only content indexed for your account, never another customer’s data. If you segment by product, retrieval is limited to that product’s knowledge. See Product Segmentation.

​

When IrisAgent refuses

1. No matching source. Nothing indexed supports a confident answer, so IrisAgent does not invent policy, pricing, or product behavior.
2. Out of scope. The question is about something your sources do not cover, such as a product line you have not indexed.
3. Missing context. The ticket or chat lacks details needed to find the right article. IrisAgent asks a clarifying question.
4. Failed faithfulness check. Pre-delivery validation finds a claim the cited sources do not support, so the answer is not delivered as grounded. See Hallucination Removal Engine.

A clean handoff to a human is better than a confident wrong answer.

​

Check grounding after delivery

In Support Analyst, ask it to check the grounding of the AI answer for a case or chat ID. It grades:

• Retrieval relevance: were the cited articles about the question?
• Grounding faithfulness: is every claim supported by those articles?

Failures are labeled as hallucination, wrong article, incomplete answer, or knowledge-base gap, so you know whether to fix the model, retrieval, or the KB.

​

Related settings

• Tone and phrasing: AI Response Style
• Map customer phrasing to KB terms: Search Term Rewrites
• Grade conversations for made-up features or policies: AutoQA

Need help? Email [e-mail].