Back

20/11/2026

AI Accuracy Rate

AI accuracy rate is the share of an AI agent's answers that are correct for the customer and supported by the source they cite. Teams measure it by grading a random sample of real conversations each month.

How do you grade an AI answer?

Sample a month of the AI agent's conversations and sort every question into one of three outcomes:

  • Answered: the agent stated a fact or gave an instruction
  • Honest miss: the agent said it can't find that and offered a person
  • Nothing to answer: a greeting, spam, or a request for a person

Grade only the answers, on whether each is right for this customer today and supported by the page it cites. An AI citation shows where an answer came from, and an AI hallucination can still match an out-of-date page.

When no page covers a question, Wonderchat's AI agent says it can't find that and offers a person, which this method counts as an honest miss.

How do you calculate AI accuracy rate?

Three numbers come from the same sample:

AI accuracy rate = right answers ÷ graded answers × 100

Honest-miss rate = honest misses ÷ questions × 100

Wrong-answer rate = wrong answers ÷ questions × 100

SearchUnify puts the sample at about 385 graded conversations for plus or minus 5 points at 95% confidence, assuming accuracy near 50%. A 3-point move on 100 answers is noise.

Vendors also count different things. Maven AGI's example counts a reply as accurate only when no escalation was needed.

For example, a sample of 220 conversations has 20 with nothing to answer, 40 honest misses, and 160 answers, of which 144 are right and 16 wrong:

  • Accuracy: 144 ÷ 160 = 90%
  • Honest-miss rate: 40 ÷ 200 = 20%
  • Wrong-answer rate: 16 ÷ 200 = 8%

Count every honest miss as a failure and the same month reads 144 ÷ 200 = 72%.

Why does AI accuracy rate matter?

A wrong answer costs more than an honest miss, because the customer acts on it. Air Canada's website chat told a customer he could claim a bereavement fare after booking, and on February 14, 2024 the tribunal ordered it to pay $812.02, holding that "it is responsible for all the information on its website."

In 837 reviews of 18 vendors, read in September 2026, Wonderchat counted 1,309 complaints, and 128 were about trust. Before you compare AI agents on accuracy, ask each vendor what its agent does when it doesn't know.

Is an honest "I don't know" a wrong answer?

A formula with only right or wrong counts an honest "I can't find that" the same as a wrong answer, which rewards guessing. An agent that answers everything can outscore one that admits it doesn't know while sending customers more wrong answers.

Researchers at Stanford kept refusals apart in a 2024 test of legal research AI tools. Lexis+ AI answered 65% of queries accurately, while 62% of Ask Practical Law AI's answers were incomplete, meaning refusals or answers without grounding (Magesh et al.).

So drive down wrong answers per 100 questions. A rising honest-miss rate with a steady accuracy rate points at gaps in the docs.

Wonderchat answers from a team's own help center and shows the source on every reply. The source lets a grader check an answer in seconds.

Signpost sorting a sampled question into a graded answer, an honest miss and nothing to answer
Honest misses get their own line. Illustrative example.

Frequently asked questions

Is any AI 100% accurate?

Tools sold as hallucination-free still make mistakes. In a 2024 Stanford test, legal research tools whose makers said they avoid hallucinations made things up on 17% to 33% of queries.

Which AI has the highest accuracy rate?

No shared benchmark answers that for customer support, because vendors count different things. The fair test is one sample of your real questions, graded the same way for every tool.

What is a good AI accuracy rate for customer service?

Customer service has no shared standard for it, and Maven AGI names 80% to 95% for modern AI platforms, a vendor's own figure. Set your baseline from one month's sample, then watch wrong answers per 100 questions fall.

See it on your content