AI support accuracy you can verify

How accurate is IrisAgent, and how do you measure it?

IrisAgent delivers validated answer accuracy above 95% by grounding every response in your knowledge base, past tickets, and approved procedures, then attaching the exact sources it used. Its Hallucination Removal Engine delivers zero hallucinations: if a claim is not grounded in your content, the answer does not go out.

• Grounded and cited: every answer is built from your verified content and shows its source. If the grounding is not there, the answer does not go out.
• Restraint by design: when your knowledge base lacks the answer, IrisAgent declines or asks a clarifying question instead of guessing.
• Gated on two eval sets: every model and prompt change is tested for both correct answers and correct declines before it reaches a customer.

By the IrisAgent team · Last updated June 15, 2026

How IrisAgent measures accuracy: two eval sets, not one

Most teams test their AI only on questions it should answer. IrisAgent tests both halves of reality, because knowing when to stay silent is as important as knowing the answer.

1. The resolution set. Real customer questions the AI should resolve, drawn from production support conversations. We score whether the answer is correct, grounded, and attached to the right source articles.
2. The hallucination set. Known-unanswerable questions: missing context, out-of-scope requests, or cases the knowledge base does not cover. Here the correct behavior is to decline or ask for clarification, never to guess.
3. Restraint scores as success. On the hallucination set, a confident wrong answer is a failure and a polite decline is a pass. We grade the discipline of not answering, because a wrong answer costs more than no answer.
4. Every change is gated. No model upgrade or prompt change ships if it lifts resolution at the cost of any fabrication. That tradeoff is the entire game, and we refuse to optimize one side of it in the dark.

Accuracy claims are cheap. This is the live signal from production. Over a recent 90-day window, support agents left roughly 118,000 explicit accept-or-reject judgments on IrisAgent output. All time, that feedback corpus exceeds 1.5 million human judgments, and every one of them sharpens the eval sets and the models behind them.

Where agents accept AI output

Accept rate, trailing 90 days

These are decisions on live customer output, not a lab benchmark. An ignored recommendation is not necessarily wrong, so real accuracy sits at or above these rates.

Grounded in your truth, not the internet

Retrieval-augmented generation means IrisAgent pulls answers from your specific knowledge base at the moment of the query, instead of making them up from training data. The system retrieves the exact source material, generates the answer from it, and shows the article it used.

This is why grounding is a mechanism for us, not a tagline. It is what drives hallucinations to zero, and it is what lets a support leader audit any answer back to its source in one click. Learn how it works across AI for customer support, agent assist, and voice AI.

Three questions to ask any AI support vendor

Whether you are evaluating IrisAgent, Decagon, Sierra, or Intercom Fin, these three questions separate proven accuracy from marketing claims.

1. How do you prevent hallucinations, and how do you measure it? If a vendor cannot show a method, they are not preventing them.
2. What does your AI do when it does not know? "It always answers" is the wrong answer. The right behavior is a clean decline or escalation.
3. Can the AI show its sources for every response? Grounding you cannot inspect is grounding you cannot trust.

Comparing a specific platform? See IrisAgent vs Decagonand IrisAgent vs Sierra.

Accuracy and trust, certified

IrisAgent is SOC 2 Type II certified and GDPR compliant. Your data grounds answers for your own customers and is never used to train models for anyone else. Read more on our securitypage.