Multilingual AI Agent: Hebrew, Russian and English

Ten in the evening, a kitchen showroom in Ashdod. Three messages land on the business WhatsApp in a quarter of an hour: one in Hebrew about a made-to-order stone countertop, one in Russian with a photo of a dining nook and a question about dimensions, one in English from a couple that moved to Israel last month. In the morning the manager answers the first, pushes the second through Google Translate into something that reads like a form, and never reaches the third. One of them ordered a kitchen elsewhere in town, and the showroom never knew it lost a sale: that inquiry was never logged where leads are counted.

Three languages on one phone number

A small Israeli business does not decide to be multilingual, it simply is. In Ashdod, Haifa, Netanya and Be'er Sheva a large share of inquiries arrives in Russian. In the big cities, the tourist areas and around the tech hubs there is a steady stream of English from new immigrants, foreign workers and tourists, and anyone selling abroad works in English from day one.

The team's language is the ceiling of the service. If reception speaks Hebrew only, an inquiry written in Russian gets a late answer, a badly translated one, or none at all. And the customer will not stretch for you: someone writing in Russian at ten at night simply gets a Russian answer somewhere else.

To be precise about scope: at WhaleBiz the agent is configured by default for Hebrew, Russian and English, and additional languages are a project conversation on Custom.

What happens when a customer writes in a language nobody answers

None of these three leaks has a line in any report, which is why they run for years.

• A message that stays unanswered. It is never marked as a loss, because it never became a lead, and it was the cheapest inquiry the business will ever get: someone who already searched, already found you and already wrote.
• An answer that went through a machine. Automatic translation does not know that business Russian defaults to the polite form, does not know your professional terms, and carries over Hebrew idioms that do not survive. When the product is expensive, that is the moment the customer checks one more supplier.
• The employee who becomes a translation department. One person speaks the language and every inquiry ends up with them. In the week they are on holiday or on reserve duty the whole segment gets poor service, and it looks like one person's workload rather than a process failure.

What a multilingual AI agent actually does

• Detects the language from the first message and stays in it until the customer switches, on WhatsApp, the website, Instagram, Telegram and Facebook.
• Answers from one knowledge base. Prices, policies and hours are stored once and the answer is generated in the customer's language, so a price update never leaves an old Russian version running.
• Handles voice notes. Many Russian and Hebrew inquiries arrive as a recording, especially from older customers and people driving. It is transcribed, answered and stored on the record.
• Qualifies the lead to the end: what the customer needs, when and at what budget, then books a meeting and sends a confirmation email, all in the same language.
• Stores the language as a field on the customer record in WhaleBiz CRM with the transcript, while the new-lead email can be set to Hebrew even when the conversation ran in Russian.
• Knows when to hand off. A complaint, an edge case or a topic the agent should not handle goes to the team with the full history.

The model is deliberately hybrid: the agent takes the volume, the human takes what needs judgement. The answering side is the customer support agent; handling every channel in one place is covered in the unified inbox guide.

Language detection: what it looks like from the inside

Language detection sounds like a one-line feature, but the details separate good service from a slightly off feeling:

• Language sticks to the conversation, not the account. The first message decides and the language holds to the end. An agent that re-detects every message starts bouncing between Hebrew and Russian the moment the customer types ok.
• Mixed messages are the rule, not the exception. A Russian sentence with the Hebrew words for invoice, appointment or VAT in it, or Hebrew with an English technical term. Answer in the dominant language and do not translate terms the customer used.
• Names stay, formats stay. Family names, street names and model numbers are never translated, and prices, dates and currency keep one format in all three languages: shekels, day-month-year, Israel time.
• Right-to-left and left-to-right mixed together. Hebrew with numbers, links and Latin words breaks alignment in forms and emails, not only in chat. The technical side is in our RTL challenges guide.

Three ways to answer in three languages

The mistakes that turn a multilingual agent into a nuisance

A knowledge base duplicated into three languages. The expensive one. Once there are three price list files, one gets updated and two do not, and the agent quotes a price that no longer exists. One source of truth instead.

No glossary and no register. Every field has words you cannot improvise: deposit, warranty, measuring visit, lead time, cancellation. That is five to ten terms, approved in three languages once. Tone is the same kind of decision: business Russian defaults to the polite form and a chummy tone reads cheap, while in Hebrew it is over-formality that grates.

A language that does not survive the handoff. A customer handled in Russian who gets a call from a Hebrew-only rep is back at the starting point, and automation behaves the same way: if the language is not stored on the record, the reminder and the confirmation email go out in the wrong one.

Language as a data field: what it does for sales

Once the language is stored as a field, you can segment: how many of this month's inquiries were in Russian, what the close rate is per language, when the English ones arrive. Many businesses discover a segment here they never knew they had, because it never reached a report.

From there it gets practical. A follow-up sequence goes out in the language the customer wrote in. A campaign goes to a list segmented by language, in copy written rather than translated. A lead arriving on Saturday evening gets an immediate answer, and on Sunday the team finds an ordered list with summaries. That side, from qualification to a booked meeting, is the sales agent.

This is also where to check what your shortlisted platform meters. Many chat systems sell automations against a counter of one to five thousand triggers a month, and in a multilingual business the sequence count nearly doubles by itself. At WhaleBiz follow-ups and workflows are unlimited on every plan, with no trigger meter, and the CRM is included.

How to implement it properly: five steps

1. A two-week language map. Tag the language of every inquiry from the past two weeks. It takes an hour and tells you whether this is five percent of inquiries or thirty.
2. One knowledge base plus a glossary per language. Price list, policies, hours and the work process, plus the approved terms in three languages.
3. Define the boundaries. Which languages are in, what happens with a fourth one, which topics always go to a human, and who receives the handoff in each language.
4. Connect the CRM and the templates. A language field on the record, and confirmation, reminder and summary templates in three languages, including the links the agent sends.
5. Measure after three to four weeks. Inquiries and close rate segmented by language, and what the agent could not answer. Update and run again.

What it costs and the ROI

Solo is 490 ILS per month (excluding VAT): one agent on one channel, 2,000 AI messages per month (about 300 conversations), a knowledge base of up to ten documents and WhaleBiz CRM. Pro is 990 ILS per month: up to three channels, 5,000 AI messages (about 750 conversations), an unlimited extended knowledge base and in-chat payments. On both, follow-ups and workflows are unlimited. There is also a one-time setup fee from 490 ILS for full setup by our team; it is non-refundable, and the subscription starts counting only when the agent goes live. Unusual needs, such as a fourth language, are scoped on Custom, and the full list is on the pricing page.

The price is not set per language. Another employee who speaks Russian costs thousands of shekels a month, and a translation plugin answers nobody. Take a showroom with 1,200 ILS of gross profit on an average deal: two deals a month that closed only because somebody answered in Russian at ten at night cover Pro and leave change.

Frequently asked questions

How does the agent know which language to answer in?

Detection is based on the message itself, not on an account setting. The first message sets the conversation language and the agent stays in it until the customer switches, and in a mixed message it answers in the dominant language.

Do we need a separate agent for each language?

No. One agent handles all three from a single knowledge base, and the price follows the plan, the channels and the message volume, not the number of languages: Solo at 490 ILS per month covers one channel, Pro at 990 ILS per month up to three.

What happens when a customer writes in a language we never configured, such as Arabic or French?

Our default set is Hebrew, Russian and English, and further languages are scoped per project on Custom. Until they are configured the agent should not improvise: it replies briefly in a language it commands, collects a name and phone number and flags a handoff.

How long does it take to set up an agent in three languages?

Full setup is done by our team and takes up to 14 business days from receiving all materials, including the price list, the glossary per language and the message templates. A one-time setup fee from 490 ILS applies, and the subscription starts counting when the agent goes live.