Skip to main content
Confi
Question and answer

How can I check that an AI customer-service agent really handles Hebrew, in chat and on the phone, before I pay for it?

The short answer

Test a Hebrew service agent with a script covering names, numbers, addresses, dates and handoff to a representative. Use chat and the phone line planned for your business. Before buying, define which error requires stopping an enquiry and how to verify corrected details reached the representative.

Written by Confi, an Israeli automation and AI company.

This article draws on research and on the knowledge Confi has built up working with businesses. We checked the facts and figures against the sources linked in the article, and the check date is shown there. AI helped with the writing. Illustration created with AI.

Last checked: Updated:

A service bell and an open notebook on a counter beside a cup of tea.

How do I decide whether the agent passed?

Use a script covering your business needs and record answers and corrections. This is a suggested test, not a product rating. Include a caller correcting a phone number, email address or appointment time, and check how the correction reaches the representative.

Test content What to look for
A Hebrew name with a geresh, then its spelling Read-back and acceptance of a correction
A phone or order number spoken in groups of digits Confirmation before continuing
A street name that is also an everyday word, with city and house number A full address that can be corrected
A relative date such as "next Sunday" An explicit date and time for approval
The caller interrupts the agent's answer Stopping and continuing from the corrected request
Asking for a representative early and mid-conversation A clear handoff to a representative with the conversation so far

Mark each result as correct, corrected, or an unresolved failure. Before talking to a supplier, decide how to record results.

Which mistake should stop the enquiry?

If the agent continues with a wrong number, stop the enquiry and hand it to a representative.

Illustrative example, not a customer result: a caller corrects a digit in a grouped number. The agent must read back the correction and ask for confirmation.

Ask the supplier to demonstrate how the agent handles an angry caller and a request that changes during the conversation. Decide which actions the agent cannot perform alone, and check that it hands them over to the representative.

What should I check before paying?

Before paying, review these points:

  • Call from a mobile and a landline, not only through a demo page.
  • Test chat and voice separately and record differences.
  • Ask which speech-recognition engine and language model the proposal includes.
  • Request a transcript and compare details with what you said.
  • Ask for a representative when nobody is available, and check who receives the request.
  • Agree to rerun the script after every model or system change.

Read OpenAI's speech-to-text documentation:

Accuracy varies by language [1].

Test in Hebrew even if you saw a demonstration in another language.

Read Google Cloud's guidance:

In telephony, the native rate is commonly 8,000 Hz [2].

Test on the phone line itself as well as chat.

Record a call with your business's names, numbers and vocabulary. Compare the recording with the transcript and note whether each detail is correct.

Questions about testing a Hebrew service agent

Is chat testing enough to predict phone performance?

No. To check phone performance, run the test on the line you plan to use, even if you already tested chat or a demo recording.

Can I rely on the supplier's demonstration?

A demonstration shows what the supplier chose to show. Test situations outside the demonstration. Try them yourself and record cases that do not meet your requirements.

What should I do if the agent gets a digit wrong?

Request a read-back and correction before the agent uses the detail. Rerun the test after correction and check what reaches the representative.

Sources

Last checked: Updated: Published:

Let’s talk about your business process

A short fit call: we look at one process in your business together and tell you honestly whether there is something worth building.