Masculine or Feminine: How an AI Agent Addresses Every Customer Correctly

In Hebrew and in Russian you can barely write a sentence addressed to a customer without deciding whether that customer is a man or a woman. The verb demands it, the adjective demands it, and sometimes the preposition does too. In English the decision simply does not exist, which is why a team that built its messaging in English has never met the problem. It sounds like a grammatical detail right up to the moment an automated system writes on your behalf, because a system has to choose a form even when it has no idea who is on the other side. The common default is masculine, and that default meets the female half of your enquiries dozens of times a month.

The arithmetic that makes this a commercial problem

We have no industry figure for how many customers walk away over a wrong form of address, and we are not going to invent one. What we can do is arithmetic. In a typical conversation the agent sends between six and twelve messages, and in Hebrew a large share of them contain at least one form that must pick a gender. If the default is masculine and roughly half your enquiries come from women, the chance that a female customer gets through a whole conversation without hitting a single wrong form is very low.

What happens next depends on the person. Some do not notice. Some notice and say nothing, but the impression left behind is of a generic system that is not really talking to them. And some reply with a short line to the effect of "I am a woman", then wait to see whether the system understands. That is the real test: a customer who corrects once and sees no change almost always stops treating the exchange as a conversation.

The other side deserves saying too. Correct address does not close a deal on its own, and it will not rescue an agent that answers the wrong questions. It works like spelling a customer's name right: nobody buys because of it, but it is the first thing that gives away that there is a template on the other end rather than a business.

Four broken sentences, and what they look like fixed

The fastest way to see the scale of this is to look at real sentences an agent writes, of the kind found in every Israeli business's conversation script:

• "אתה מעוניין לקבוע תור לשבוע הבא?" versus "את מעוניינת לקבוע תור לשבוע הבא?" ("Would you like to book an appointment next week?") Two broken words in one sentence, and it appears in nearly every conversation.
• "שלח לי בבקשה את מספר הטלפון שלך" versus "שלחי לי בבקשה את מספר הטלפון שלך" ("Please send me your phone number"). The imperative is where the error is most glaring, because it reads like an instruction meant for somebody else.
• "רשמתי אותך לשלישי בעשר, אתה מאשר?" versus "את מאשרת?" ("I have you down for Tuesday at ten, can you confirm?") This is the message that closes the appointment, which makes it the most important message in the thread.
• "אם תרצה, אשלח לך את המחירון" versus "אם תרצי, אשלח לך את המחירון" ("If you'd like, I'll send you the price list"). Note that "אשלח לך" is written identically in both forms, and only the first verb breaks.

That last detail is the key to the fix: unvocalised Hebrew has a good number of forms that read as both masculine and feminine. "פנית אלינו", "אשלח לך", "נשמח לעזור" and "אפשר לקבוע" are all neutral in writing. The question is whether whoever wrote the script knew that and built on it, or wrote in the masculine and then tried to bolt a feminine variant on afterwards.

Russian has the problem too, and it hides better

The common assumption is that Russian is exempt because polite address uses the second-person plural. That is true for a large share of sentences, which is why a team reviewing ten Russian messages comes away convinced all is well. That review misses exactly the places where Russian does break:

• «Уважаемый Анна» versus «Уважаемая Анна». An honorific opener in a fixed template, and precisely what breaks.
• «Вы наш новый клиент» versus «Вы наша новая клиентка». A noun describing the person carries gender even when the verbs around it do not.
• «Вы единственный, кто записался на это время» versus «Вы единственная, кто записалась на это время». A singular construction leaning on plural address, dragging two failures along with it.
• «Ты записался?» versus «Ты записалась?» Any business that chose a friendly second-person singular, and that is common among Russian-speaking businesses in Israel, loses the protection of polite address entirely.

There is an upside worth knowing: in Russian the customer gives herself away very early. Every first-person past-tense verb does it, and the opening message usually contains one, for example «я записалась» or «я хотела уточнить». So Russian has fewer dangerous sentences than Hebrew, but also a stronger opening signal.

English, by contrast, has no gender agreement at all, and that is worth stating explicitly: reviewing the English version teaches you nothing about the other two languages. If you run a multilingual AI agent working in Hebrew, Russian and English in one channel, the three versions need three kinds of review, not the same review three times.

What actually settles gender, and what only looks like it does

This is where most systems go wrong, because they reach for a fast answer instead of waiting for a correct one. The table separates signals you can act on from signals that look convincing and are not:

The rule that follows from the table is short: a correct guess earns nothing, because nobody is moved by being addressed properly, and a wrong guess costs you the enquiry. When the upside is zero and the downside is real, you stay neutral until a genuine signal arrives.

How to stay neutral in Hebrew without sounding strange

"Neutral" immediately raises the fear of slash forms like "לקוח/ה יקר/ה", and the fear is justified: slashes inside a word look like a government form, and in WhatsApp they damage the impression. But they are only the lazy solution. There are four ways to write neutral Hebrew that sounds entirely human:

The first is first-person plural. "נשמח לקבוע לכם תור", "בדקנו ויש מקום ביום שלישי", "נחזור אליכם עם אישור". The business talks about itself instead of addressing the individual, and it sounds completely natural in business Hebrew.

The second is the infinitive and impersonal constructions. "אפשר לקבוע לשלישי בעשר", "כדאי להביא צילום קודם אם יש", "נדרש אישור מראש". "אפשר" and "כדאי" resolve a whole class of sentences that would otherwise have to choose.

The third is a nominal construction, a sentence with no second-person verb. Instead of "אתה יכול להגיע ב-14:00" you write "יש מקום פנוי ב-14:00", and instead of "אתה מעוניין במחירון" simply "אשלח מחירון".

The fourth is exploiting the forms that are written identically: "תודה שפנית אלינו", "אשלח לך", "קיבלתי ממך" and "נעדכן אותך" are valid for both genders unvocalised, and a whole opening message can be built on them.

What remains hard is the imperative. "שלח" and "שלחי" share no common form, so in neutral mode the agent rephrases: instead of "send me your phone number" it writes "a phone number to call you back on would be great". Same content, no decision made.

The rule that breaks most often: hold the form to the end

Most systems that do handle gender handle it once, at the start of the conversation, and then lose it. This happens in three predictable places, and each deserves its own check.

The first is mid-conversation. The agent identified correctly at message three and reverts to masculine at message nine, because by then the thread is long. From the customer's side that is worse than being wrong from the start, because she has already seen that the system knew.

The second is any text sent later: the booking confirmation, the reminder the day before, the follow-up two days out. Those texts were written once in advance, usually in the masculine, so they wipe out all the work done in the conversation. The fix is that the settled form lives on the lead card in the CRM rather than only in the conversation's memory, and every outbound text reads it from there. That is exactly the point of a CRM built around AI agents: what the agent learned in the conversation is stored where the rest of the system reads from, follow-up rules included.

The third is the handoff to a human. The settled form is part of the context the human receives along with the lead summary, and without it you get the absurd result where the bot is precise and the person is not. An AI support and customer service agent that works properly hands over the conversation with everything it knows, this small detail included.

One point calls for honesty. In approved WhatsApp template messages, the kind used to open a conversation outside the messaging window, the text is fixed in advance and cannot be altered on the fly. So those templates are written in neutral phrasing from the outset, or two approved variants are prepared and selected between according to what is stored on the card. That is a channel limitation rather than an agent limitation, and it is better known during setup.

How to test this before the agent talks to customers

The test takes about fifteen minutes and you can run it yourself. Three conversations, each following the same path through to the end of the process.

In the first, write as a woman, with a clear signal in the opening message: "היי, אני מעוניינת לקבוע תור". In the second, write as a man. In the third, write completely neutrally, for example "היי, אפשר לקבוע תור לשבוע הבא?", and add no signal at any point in the thread.

In each of the three, check five points: the opening message, the message that collects details, the booking confirmation, the reminder if there is one, and the first follow-up message. In the third conversation the correct result is that all five are neutral and none of them sounds like a form. If something in there guesses, you have a bug, not a style preference.

Add a sixth check: mid-conversation, write an explicit correction after the agent has already been addressing you correctly, and verify that it accepts it immediately and does not slip back five messages later. This is the most common failure and also the easiest to catch. And if you run Russian as well, repeat all three conversations in Russian, especially the honorific opener and the reminder templates, because that is where the translation was usually done once and never reviewed again. The same logic applies to any AI assistant writing in Hebrew, not only to ours.

Two commercial notes to close. This behaviour is part of the agent's definition on all three plans, not a paid extra. And if you later want to change the style of address, for instance moving every message from second person to first-person plural, that counts as one change against the monthly quota. An agent addressing a woman in the masculine after she wrote about herself in the feminine, by contrast, is a fault, and fault fixing is free and unlimited.

Frequently asked questions

How does the agent know whether to address someone in the masculine or the feminine?

It does not know in advance, so it opens in a neutral register and switches to a gendered form only once the person settles it themselves, almost always through an adjective or participle they use about themselves: ani meunyenet, ani tzricha, ani batuach. From that moment the form is stored on the lead card in the CRM, so a follow-up going out two days later and a human agent joining the thread both use it. If no signal ever arrives, the agent simply stays neutral to the end, which is a correct outcome and not a failure.

Does the agent guess from the name?

No, and that is deliberate. Israel has common given names that settle nothing, among them Yuval, Noam, Rotem, Shachar, Tal, Eden, Lior and Or, and Russian has short forms used by men and women alike. The name shown in WhatsApp is also frequently a relative, the business itself or a nickname. The agent does not guess from the treatment or the product either, because men book a cosmetician and women book car servicing.

What happens if the agent gets it wrong and the customer corrects it?

The agent accepts the correction immediately, without arguing and without a three-line apology, and uses the right form from the very next message. The correction is saved on the card, so it does not slip back later in the thread or in the next conversation. The serious failure is not the first mistake but the return to it: a customer who corrected the agent and is addressed in the masculine again five messages later concludes that nobody is listening.

Is this a Hebrew-only problem, or does Russian need handling too?

Russian too, but there it hides better. The polite second-person plural covers a large share of sentences, which makes it easy to assume everything is fine. It breaks in four fixed places: the honorific opener that has a masculine and a feminine form, nouns describing the person such as klient versus klientka, any singular construction, and any copy written in the friendly second-person singular. English has no such agreement, so an English version that reads perfectly tells you nothing about the Hebrew and Russian beside it.

Does changing the agent's form of address count as a change against the monthly quota?

Yes. Rewording, or changing the agent's rules of address, is a change, and a monthly change quota is included, sized by plan. By contrast, an agent addressing a woman in the masculine after she wrote about herself in the feminine is a fault, not a change, and fault fixing is free and unlimited on every plan. The distinction is simple: altering what was defined is a change, failing to meet what was already defined is a fault.