How the bot builds its conversation

Your bot does not follow a script, and it is worth understanding why before you try to configure one.

It is given three things: who you are, what it needs to find out and why, and what has been said so far. It is then asked to have a useful conversation. So it reorders questions, merges two into one when the lead has half-answered both, and rewords everything to suit whoever it is talking to. A lead who opens with "hi, I need a photographer for a wedding in June, budget about 40k" has answered three of your questions before the bot has said anything, and the bot will not ask them again.

Everything the bot knows is configuration, read when it replies. Change a question this afternoon and the next conversation uses it.

What you control

The shape of a conversation

1. Greeting. It welcomes the lead to your business, says in a sentence or two what you do and for whom (below), then asks its first question.
2. Intake. Before every reply, the product works out which question is worth asking next (below), and the bot asks that one: your must-know questions first, then your optional ones. Each reply reacts to what the lead just said, with one true detail about it, before asking the next thing. Where a question has fixed options, it works the answer out ("Mumbai" is your West region) rather than reading your category names out.
3. Clarification. When an answer is unusable ("flexible" for a budget) it asks again, differently. When the lead says they are not sure, it helps them answer instead (below).
4. Raising a mismatch. When an answer is outside what you offer, the bot says so there and then and carries on (below).
5. Selling, all the way through (below).
6. The read-back, only when something was guessed or suggested (below).
7. One verdict. Once every question is settled, your rules decide: qualified, held for your team, or not a fit (how). A lead that reaches your team is matched against your records if automatic matching is on, and your team is told. With options on, a qualified lead is also shown the best few of your records (below); with booking on, it is offered times (below). A lead that is not a fit is marked Lost quietly. The one exception is a lead that has already missed your threshold of hard rules: it is marked Lost at that moment, and what the bot does next is your choice (below).

One reply, four steps

Every reply is made in four steps, so that the part that reads, the part that decides and the part that talks are kept apart:

1. Understand. An AI call reads the lead's latest message and writes down what it says: the answers, any corrections, what they asked your business, and signs such as frustration or a request for a person. It writes nothing to the lead.
2. Decide. The product itself, in code and not by AI, applies that to the lead. It stores the answers, runs your rules, works out the verdict, chooses the next question, picks the facts the reply may use (only those that relate to what the lead asked) and picks the one sales move for this turn, if there is one.
3. Speak. A second AI call writes the reply from that plan, what is already known and those facts, and from nothing else. It is told what has been decided, so a closing is written knowing whether the lead qualified, is held for your team or was turned away.
4. Check. The product reads the reply before it is sent, and refuses what nothing supports (below).

Two AI calls make one AI reply on your plan: the meter counts replies, not calls (Plans, limits and meters).

If the first step fails, one call does the whole turn instead, so the lead still gets a reply. It is held to the same facts and the same check, but it sells nothing. If the check refuses that reply, the lead gets "let me check that with the team" and your next question at once, with no second try. If the second step fails, the lead gets your next question rather than a goodbye.

Which question comes next

The next question is chosen by the product, in code, not left to the model. The same facts always give the same choice:

• A question you marked must know comes before any optional one. Two must-knows are asked together only when both are quick to answer (a number, or a choice of four or fewer).
• Among optional questions, Should ask goes before Nice to have. A question that feeds a rule or your matching goes first, and one just asked drops back so the bot does not circle.
• A question with Only ask when… waits for the answer it depends on, and is skipped if that answer rules it out.
• A sensitive question is asked only if you made it must-know, and always on its own.
• A switched-off question is never asked and never waited for.
• Anything the lead already said, in any message, counts as answered.

The model then words the question. It is told why you ask, what is already known and what is left for later, and it records an answer to a later question if the lead gives it early. See What your bot finds out for how to set each of these.

When a lead is not sure

A lead who says "I don't know" gets help, not the same question again. The bot asks about their situation, suggests the kind of option that fits and checks, offers a simple choice in words when it is hard, and falls back to the safe default you wrote ("a GP can do a first assessment and refer you on"). It records the suggestion only after the lead agrees: until then the suggestion is held, and the team sees it on the lead as Suggested, waiting for their yes. It never offers a figure, range or price band that was not given to it. How to write the help is in What your bot finds out.

The read-back

Before the verdict, the bot reads the key details back in plain words and asks if anything needs changing, but only when it has something to check: an answer it worked out (Inferred), an answer it suggested and the lead agreed to (Suggested, agreed), or a hard rule the lead is failing. When the lead stated every answer and no hard rule is failing, a read-back would repeat what they just said, so it is skipped.

• It happens once per enquiry. A correction updates the answer and is not read back again.
• If the lead's next message does not confirm it, the bot does not read it back again, and the guessed answers stay Inferred.
• Sensitive answers are never in it.
• A confirmed answer is marked Confirmed on the lead.

A guess never turns a lead away

A hard rule can turn a lead away only on an answer they gave, agreed to or confirmed. If the bot only worked out an answer and it fails a hard rule, the read-back comes first. If the lead confirms, the rule counts as usual. If they do not, the miss is treated as a negotiable one: your team sees it and the lead is not marked Lost.

Private answers

A question marked Sensitive is never read back, and its answer is never put in the running summary, in the facts the bot keeps for your team, in a follow-up, in an alert or in a later message. If you also made it must-know, the bot does ask for it, gently, and may acknowledge it once, in its reply to the message that gives it, because it needs it to suggest the right person. Your team reads it on the lead's page.

The one exception: a lead who asks what they told the bot ("what did I say was wrong with me?") is answered on that reply, from their own answers and only the ones they asked about (what the bot may tell leads). Answers the bot only worked out, answers to switched-off questions, your team's notes and any score are never read back. On every other reply the rule above holds.

The bot never sounds finished with a question open

The bot does not say goodbye, "that's everything" or "someone will be in touch" while any question is still open. The prompt says so, and the reply is also checked before it is sent: the sentences that sound like a closing are taken out, the rest of the reply (a real answer to the lead, say) is kept, and the next planned question is added in the lead's language. This is a separate check from the one for Lost leads, below. When the lead has just asked something ("when will someone get back to me?"), a sentence that promises contact is the answer and stays; a goodbye still goes.

A lead's question is answered first

When a lead asks something, the reply opens with the answer and only then asks its own question. If you have not written the answer down, the bot says so plainly and that the team will confirm it, rather than moving on. A lead who asks the same thing a second time, while a must-know is still open, gets that answer and nothing else in the reply: no question, no offer, no limit raised. The conversation carries on from the next message.

What else the bot notices

Besides the answers, the bot reports a handful of things about each message. Four open a reason under Needs you; none has a switch in Settings, because they are always on:

The bot also keeps a running summary of long conversations (after about ten messages) and a short list of facts the lead mentioned that none of your questions asks about. The facts appear as Also mentioned. The bot is shown the facts already kept and says when a new one only repeats a kept fact in other words, so the same fact is not written twice.

A lead who says goodbye is not asked one more question

When a lead signs off ("thanks, bye", "that's all", "maybe another time"), the bot stops waiting for your optional questions and closes warmly with no question at all, a must-know that is still open included: it is asked on the lead's next message, not in the same reply as their goodbye. A limit their answer misses is raised on the next message too, never in the goodbye. This holds whether or not the lead was marked Lost. A message that asks something ("bye, what's the price?") is not a goodbye. The bot reads English, Hindi, Spanish, French, Portuguese and German goodbyes in code; in another language it reads the message as it always did. The goodbye is closed with thanks and nothing else: no "shall I pass this to the team?", no upsell, no offer.

The greeting introduces your business

The first reply of a conversation welcomes the lead to your business and says what you do and for whom. It welcomes them to your business only, even when their message names their own company. (A lead who comes back to start something new is not introduced to your business a second time.) When the very first message already misses a hard rule and you chose to end the conversation, the reply still opens with a short welcome that names your business and a thank-you, then gives the rule's message.

It uses How your bot introduces you from your profile (up to 300 characters). If that is blank it falls back to What you do. Templates can start you with an introduction written for their industry, and you can change it. The profile page shows a preview of how it will read in a message.

Sounding like you

Fill in How your bot should sound with a short sample chat, a few Lead: and Bot: lines. The bot matches its warmth, rhythm and length, and never copies its facts, places or prices. Each reply follows the same pattern: react to what the lead just said, with real feeling and one true detail ("Great choice" alone is not enough; it says why); ask the next thing so it follows from what they said, usually one question; and keep it to three short sentences at most, with no lists and no gushing.

The bot never assumes anyone's gender. Asked "are you a bot?" or "am I talking to a person?", it says honestly that it is your business's assistant and offers to pass the lead to your team, whatever personality you gave it. The full guide, with examples by industry, is How your bot sounds.

What the bot knows about today

Every time it replies, the bot is told today's date, the day of the week and the time where you are, and whether you are open. That is what lets it answer "can someone call me today?" from facts instead of guessing. What it is told depends on what you filled in under When you're open on your profile (both fields are optional):

So a lead who writes at 10:30 pm on a Saturday, to a business open Monday to Friday, is not promised a call before Monday morning. Each day is closed or has up to two periods, and a period cannot run past midnight, so a late night is two days. Hours only tell the bot when your team is in: they do not stop it answering and do not change when your team is notified. Booking needs a time zone too: without one the bot offers no times.

One reply for a run of quick messages

People send three short messages instead of one long one. The bot waits a moment (about two and a half seconds) after each message in case more are coming, then answers the whole run once, reading all of it. It also finishes one reply before it starts the next for the same lead, so two quick messages never get two overlapping answers, even when your workspace is busy. The playground does not wait.

Voice notes, photos and other things that are not text

The bot reads text only. It does not listen to voice notes, look at photos or video, or open documents or contact cards. A WhatsApp location pin is the one exception: its place name and address (or its coordinates) are read as words, and the bot never opens or looks up the map.

• With no words. The lead gets one fixed reply in their language: "I can only read text for now — could you type that for me?" No AI reply is used. The message is stored, and the Inbox shows it with a label (Voice note, Photo, Video, Document, Sticker, Location, Contact or Attachment), so your team can see what arrived and reply themselves if it matters. The bot does not store the file. If it is the first thing the lead sends, the reply is in English, because the bot has not seen them write yet.
• With a caption. A photo, video or document sent with words is answered from the words, as any message. The label still shows on it.
• Three in a row. Three messages in a row with nothing the bot can use raise Bot going in circles under Needs you (when that switch is on).
• When the bot is not answering (a person has the conversation, or it waits on your team), media is handled like any message the bot does not answer: your team is flagged under Needs you. A locked workspace stores it and says nothing.

Fixed replies follow the lead's language

The bot reads and writes freely in whatever language the lead uses. A small set of sentences is fixed instead, written by the product so that nothing unexpected is said at the wrong moment: the reply to "stop", the holding reply when the AI is unavailable, the full-capacity notice, the default hold message, the turn-away line for a rule with no message of its own, the thank-you at the end of intake, the closing for a Lost lead, "let me check that with the team", "the team has your message", the welcome back's fallback, the follow-up when the AI cannot write one, the "please type" reply, the lines around options and times, and the emergency and under-18 messages. These come in twelve languages (English, Hindi, Hindi in Latin letters, Tamil, Telugu, Marathi, Bengali, Arabic, Spanish, French, Portuguese and German) and are in English when the lead writes in another.

The language changes only when a message is mostly in the new one (about seven words in ten, and more when several words are in another language), or the lead writes in it twice in a row. A single phrase in another language inside an answer ("is there a special price?" in Tamil, say) does not move the fixed replies. The first message sets the language from whichever language most of it is in.

The fixed sentences a lead is most likely to get twice (the next question after a goodbye, "could you say that again?", "let me check with the team", the default hold, the safety messages) each have a few ways of being said. The first time it is the usual wording; after that the bot uses a form it has not just used, so a lead is never sent the same sentence twice in a row.

Anything you wrote is sent exactly as you wrote it: a rule's message, the message while your team decides, the reply when your lead allowance is used up, and your own emergency message. The product does not translate your words, so write them in the language your leads use.

Emergencies

When a lead writes that someone is in danger ("I can smell gas", "he has collapsed"), the product, not the AI, sends a fixed emergency message at once and tells your team it is urgent. A figure of speech ("I'm dying to see it") or something long past ("we had a gas leak last year") is not taken for an emergency; talk of self-harm always is. Every business has a basic safety net for life-threatening phrases; clinics and home-safety businesses have a fuller list and their own message. It is sent even when your team has the conversation, and it costs no AI reply.

After an emergency (or a lead who says they are under 18) the bot asks nothing more, decides nothing and sells nothing in that enquiry. Each message the lead writes gets one short fixed line, until a person hands the conversation back. See Emergencies and safety.

A lead who asks to stop

A lead who writes "stop", "unsubscribe" or "quit" on its own, says "not interested", "take me off your list", "stop messaging me", "don't message me" or "please don't contact me again", writes "wrong number", or asks not to be messaged any more in English, Hindi, Tamil, Telugu, Marathi, Bengali, Arabic, Spanish, French, Portuguese or German, is closed at once as not interested and gets one polite fixed reply in their language. A lone "cancel" or "end" is not an opt-out (it is usually about a call or a time); the bot reads it, and "cancel, stop texting me" still is one. A message that also names a time or a condition ("don't message me until Monday") is left for the bot to read, because that lead has not left.

Anything that was waiting on your team for that lead (a human request, a price question, a reply nobody answered) is resolved, so the lead leaves Needs you and nobody is alerted about the opt-out; the closed requests stay in the lead's history. A held "Possibly a minor", an "Urgent request" or a complaint is never resolved this way: a person still has to see it.

Optional questions are asked of every lead

A question you mark optional is still asked, after the must-knows. The lead may say they would rather not answer, and the bot accepts that and does not ask again. So does a lead who answers an optional question with a shrug and nothing else ("not sure", "no idea", "whatever", in the languages the bot speaks): it is let go at once and not asked a second time. A question asked twice with no answer is also let go. A question a rule reads, or that your matching filters on, is never let go on a shrug. Should ask questions are asked up to twice. Nice to have questions are asked once, and always alongside another question, so they never cost a message of their own. A closing question like "anything else I should pass on?" (the Notes question) is added to the end of the last real question. Two quick optional questions may share one message, and a lead who gives several answers in one message is not asked them again.

The cost is a few more messages per lead, depending on how many optional questions you have. They count towards your AI allowance like any other. The one time they are not asked is when Safe mode is on.

A question the lead does not answer is not asked for ever

A person does not ask the same question again and again, and neither does your bot. A must-know is asked at most three times. Asked twice with no answer, it rests: the bot asks the other questions first, and comes back to it once, when nothing else is left. If there is still no answer, the bot lets it go and holds the lead for your team instead of qualifying them, with the question on their card as "Not answered".

A must-know the lead plainly refuses ("I'd rather not share a budget") is let go the same way, sooner: the bot accepts it once, does not ask again, asks the rest, and then holds the lead for your team. It never turns a lead away for it.

A lead who says they have not decided yet ("we haven't decided on a region", "I'll leave that with your team") is not pressed either: the question rests at once and is tried once more at the end. "Not decided yet" is a valid answer. A first "I'm not sure" is different: the bot still helps the lead work it out, as When a lead is not sure describes.

An answer the bot cannot store twice in a row (an amount in a currency you do not work in, or an option that matches none of yours) ends that question the same way: your team is told the bot could not read it, the figure the lead gave is left on their card as a note, and the bot goes on with the rest.

When an answer is outside what you offer

The bot never says a lead "does not qualify". It does this instead:

• A negotiable rule. It states the limit ("our trips start from 400 per person"), asks warmly whether there is some flexibility, and if there is not, offers what you allowed, one thing at a time (how). When those are used up, declined, or you chose only the check, it says the team will check whether something can be arranged. Then it moves on.
• A hard rule. It states the limit and asks once whether the lead can meet it ("Could three nights work for you?"). It offers nothing and promises no check with your team, because nothing is being reviewed. A reply that promises one is cut and the next question is asked instead.

It states the limit only if the rule's Tell the lead this limit switch is on (it is by default) and the number is one you have confirmed on the rule's card, and it never states a figure it was not given. Safe mode never states a limit, whatever that switch says: the bot only says the answer is a bit outside what you usually do.

The bot raises a limit only when the lead's answer really is outside it. The system compares every answer with every rule on each turn and tells the bot which limits are missed; the bot is not left to judge an answer against a figure it was not shown. If no limit is missed, the bot will not say an answer is "outside what we usually do", above or below your range, or more or less than you usually take on, and a reply that does is rewritten before it is sent. What the lead said is kept: your team reads "Not flexible" and their own words on the lead, and if the lead later gives a better answer, "moved from 300 to 420".

A lead who goes quiet

If a lead answers every must-know and then stops replying, the bot does not leave them waiting. After Decide after a lead goes quiet (minutes) in Settings (30 by default, 5 to 1,380) the optional questions still open count as declined and the verdict is reached with what there is. No closing is sent: a qualified lead is sent the options or times if you have those on, a lead turned away under End gets the rule's message, and otherwise your team takes it from there. Any message from the lead cancels the timer.

A lead who goes quiet earlier, with a must-know still open, is nudged instead: Nudge a quiet lead every (hours) (8 by default, 1 to 23), up to Nudges before the bot stops (3 by default; 0 turns nudges off). The nudge is a short question the AI writes from the conversation, which counts as one AI reply; when the AI cannot write one, or your allowance is spent, a fixed line goes instead and costs nothing. A nudge is checked again just before it is sent, and is dropped if the lead has written meanwhile or a person has taken over. When the nudges are used up, or the 24-hour messaging window has closed, the lead is held under Review before closing, or closed as not replying if Days your team has to decide is 0. Why a nudge sometimes does not go: Why a follow-up was not sent.

After the intake: the bot can keep talking

When every question is settled but the bot still owns the conversation (today, a lead that missed enough hard rules while you chose Keep talking), the bot:

• answers what the lead asks, warmly, from what it knows;
• records anything new they tell it in the right field;
• never says they have been turned away and never promises a yes. For a lead who was held or qualified it says your team will look at what is new. For a lead who was marked Lost it only says the news is noted and passed on, as below.

If a new answer changes the verdict ("actually, three nights works"), the rules are checked again. The lead reopens as held or qualified if it now passes; otherwise nothing changes. Follow-up nudges are not sent. In the thread your team sees a note where the verdict landed: "Marked Lost — the bot kept talking (your setting)".

Selling in a reply

Selling follows the same split as everything else: the product decides, the model words. Each message carries at most one sales move, chosen in this order: the answer to a pushback (a price pushback on a limit the bot has raised is answered by the offer or the team check for that rule instead), then an offer or the team check for a missed rule, then the next step, the reason, a selling point, and an offer you make any time. How your bot sells has the detail. What matters here:

• Your push level sets how many offers and selling points one enquiry gets and how often the next step is asked. Until you choose, the bot sells like a good salesperson.
• The reason is asked once, in a few words, beside another question, and kept under What (your assistant) learnt on the lead.
• A selling point (or a testimonial you ticked) is chosen by what the lead has said. The bot uses at most one in a message, in its own words, and never quotes it.
• A pushback is answered with your written answer if you have one, otherwise with a short, honest one that uses no figure it was not given.
• An offer is picked by the product, never by the model: the first switched-on offer for the rule whose condition holds, and that was not made before, up to your push level's count. A "yes" is recorded only against an offer the bot really made. Asking "what else can you offer?" after an offer counts as a no. A lead who proposes a Suggest something else offer themselves ("fewer nights is fine") has it recorded as accepted, and it is not offered again.
• What you wrote is treated as material, not orders: it reaches the model in marked blocks that say whose words they are, and what a lead writes in a pushback is used only to find your answer.

The bot does not sell when the lead was turned away, is upset, is complaining, is asking for support or reporting an emergency, may be a child, is being handed to a person, is saying goodbye, is being welcomed back or read their answers back, when Safe mode is on, or after a no to the next step (the full list).

The next step

The question "Shall I pass this to the team so they can put together what fits you?" is asked once, in the middle of the intake, to a lead who says they are ready to go ahead. It comes in that very message, beside the first question still needed, even when it is their first message (the welcome comes first). Once only optional questions are left, the ask comes with at most one of them. Every question is still asked, and the verdict still comes only after all of them are settled. Only the order changes.

• Yes: the bot thanks them and says it is being passed on, and your team sees Ready for the next step: yes on the alert.
• No: the bot accepts it warmly and sells nothing more in that enquiry.
• Not sure, or no answer: the bot does not ask again on the very next reply. Only at the strongest push level may it ask once more, after it has answered a pushback, in different words and lightly. It never asks a third time.

A lead who only says they are looking is not rushed, and one who pushes back in the same message gets that answer first. A ready lead is not asked at all when the next step is already theirs to choose: with options switched on, or with booking switched on and times to follow, the options or the times are the answer.

The closing never asks. When the intake is done, the closing for a qualified lead or a lead held for your team asks nothing: the lead is already with your team, so "shall I pass this on?" would ask permission for what has happened. It says your team will look at what they shared and come back to them, and never says when. A lead marked Lost gets the "noted and passed on" closing instead (below).

Showing options

With options switched on (how), the end of the intake changes for a lead who qualifies. The closing is a bridge instead: "Give me a moment while I pull together the trips that fit you." It asks nothing and promises no time. The conversation moves to a stage of its own, Choosing an option; the bot stays active, the lead stays qualified, and your team is told as for any qualified lead.

• Matching runs, then the options go out, even if you switched automatic matching off. The best few that fit well enough are sent as a WhatsApp list, quick replies, or numbered text. The bot did not write them: each pitch is one short line taken from the fields you made visible, checked by code, and the rest is fixed sentences.
• A lead who asks what you have mid-intake is told the bot will show what fits once it has a few details (never that the team will put options together). Once every must-know is answered, such a lead skips the remaining optional questions (one that a hard rule reads is still asked) and goes straight to the options.
• A tap, a number or "None of these" is read by code and costs no AI reply. Anything typed is read by the model against the options it was shown, and it can only name one of those. The choice is checked against the database before it is saved.
• Questions about an option are answered from the visible fields only; anything else gets the check with the team.
• None of these asks once what would make one right, can lead to one more set of options, and then hands the lead to your team (the second round).
• A lead who goes quiet gets one nudge, whatever the number of nudges is set to, and then reaches your team as No option chosen.
• The bot never offers options to a lead it would not sell to (turned away, complaining, asking for support or a person, an emergency, a paused bot or a person in the conversation, Safe mode, a no to the next step).

The peek. Once at least two of the questions you match on are answered, the bot may name the one or two records that look like a fit, beside its next question. It uses the visible fields only, happens at most twice in an enquiry and never within three bot replies of the last, and counts as the reply's one sales move. It runs plain scoring (no extra AI reply, no match recorded) and skips any comparison you set to be judged by the AI.

Booking the next step

With booking switched on (how), the end of the intake changes for a lead who qualifies. Booking is off until you turn it on, and with it off none of this happens.

• The closing comes before the times. It says that times follow ("Let's find a time for your screening call") and names no day, no time and no promise of when your team will be in touch. The conversation moves to a stage of its own, Booking; the bot stays active, the lead stays qualified, and your team is told as for any qualified lead. With options on, the times come after the lead has chosen one, or after the bot has told them your team will put some together.
• A lead who asks to book does not wait for the optional questions. Once every must-know is answered, the bot goes straight to the closing and the times (an optional question that a hard rule reads is still asked).
• The times are written by code, never by the model: the list of times, the confirmation, a move, a cancel and the reminders are fixed sentences filled with a time from the diary, in the lead's language, with your zone named. A tap or a number is read by code and costs no AI reply; anything typed is read by the model against the times it was shown, and it can only report one of those, a day or part of a day to look in, a zone, a phone number or address, a question, or a no. Before it books, the product looks the time up again and re-checks the host's hours, the notice and every other booking; if someone else took the time first, the lead is offered new times. A lead's zone ("I'm in London") is kept for the enquiry, and every later time shows both clocks.
• While times are on offer, the bot answers a question about the call from the call you described and your facts, and gives out no video link before a time is booked. It offers at most four messages of times; after that, or if nothing is free, or if the lead says no, the lead reaches your team as Didn't book a time. A lead who goes quiet gets one nudge, "Would one of those times work for you?", and then reaches your team the same way.
• The bot never offers times to a lead it would not sell to: turned away, held for your team, complaining, asking for support or a person, abusive, a possible minor, frustrated, an emergency, a paused bot or a person in the conversation, Safe mode, a workspace with no time zone, or a no to the next step. Your team can book such a lead by hand from the booking card. In Safe mode a tap or a number on times already offered still books.
• A booking link instead of times: if you gave a booking page of your own, the bot sends it after the closing and the lead reaches your team as qualified. Nothing is booked or tracked here.

After the booking the stage is Booked. The bot answers only what a receptionist would: it moves the call (three new times, and the pick moves it; the bot moves a call at most three times), cancels it when the lead asks for that now (a question about cancelling is only answered), answers questions about the call, and answers thanks with a fixed reply. Anything else is passed on to the host and shows under Needs you as "Replied, unanswered". When a move cannot be done, the call stays and the bot says it has asked your team to arrange a new time. In Safe mode the bot does not read a booked lead's words: it thanks them or passes the message on, and your team moves or cancels from the booking card. No follow-up is sent to a booked lead, and a lead who writes after the meeting has ended is treated as a lead who comes back (below).

The check on every reply

Before a reply is sent, the product checks it against what the bot was given: the lead's own words, the limits you allowed it to state, and the facts you wrote under What your bot may tell leads. A reply is not sent if it:

• states a price, a seat, a link, a phone number, a time promise ("I'll call you within two hours") or what your business offers, accepts or bills ("we work with your insurer") that none of those support;
• offers a discount or special terms (a percentage off, a better price, a free upgrade) that none of your offers or facts contain, or calls something good value or cheap on its own say-so;
• invents urgency ("only two left", "offer ends today");
• names one of your records the lead was not shown (the options it has just shown, and a peek, are the exception, and may use only the fields you made visible);
• repeats something private, or promises contact to a lead you turned away;
• runs past about three short sentences (450 characters).

The check reads English, Hindi, Spanish, French, Portuguese and German best, and the other supported languages less well. A reply that only says it will ask your team about a discount, a special price or a seat ("On the discount, I'll check with the team") claims nothing, so it goes out; the same sentence with "10% off" in it does not.

The bot is asked once more for a reply that avoids the problem. If that fails too:

• A reply that is too long is cut at the last sentence that fits. Invented urgency, a private answer and a promise of contact to a Lost lead are cut out and the rest is kept.
• For anything else, the lead gets the fixed line "Good question — let me check that with the team and come back to you." While the intake still has a question to ask, the next planned question follows it, so the lead is not left waiting: "…Could you also tell me about your budget?" The question is named by your question's label without the unit in brackets ("Area (sq ft)" is asked as "area"). Nothing follows for a lead who asked for a person, complained, asked a support question or reported an emergency, for a lead you turned away (who gets the Lost closing), or once the intake is done.

A reply refused for a price or a discount opens Asked about price on the lead, one refused for an availability or seat claim opens Asked about availability, and anything else is noted as Couldn't answer. The lead has been promised an answer, so your team follows up: the bot does not go back by itself.

While times are on offer or a call is booked, the check is stricter about times: a reply may name a day or a clock time only if it is one of the times the bot offered, the booked time, or the lead's own words, and it may say "you're booked" only about the booked time. It also gives out no video link before a time is booked, and no web address other than your booking page.

Five of these fixed-line replacements in an hour switch Safe mode on (below).

A Lost lead is never promised contact

A lead the rules turned away is never told that your team will contact them, reply, or come back to them. Your team has not agreed to it, and for most of these leads nothing is planned. The bot says instead that what they shared is noted and passed on. This applies to the closing on the turn that makes the lead Lost, and to every later answer under Keep talking.

Because a model can slip, the reply is also checked before it is sent (one of the rules of the check on every reply). If it still promises that anyone will act, whatever the wording, that sentence is cut, or the whole reply is replaced by the closing. That covers "the team will", "we'll get in touch", "you'll hear from us", "expect a call", "they'll confirm what's possible" and "I'll pass it on so the team can confirm", in English and in the lead's own language (Spanish, French, Portuguese, German and Hindi too). "I can't promise anyone will call" is not a promise, and neither is "I've noted that" or "I've passed it on to the team". The closing reads: "Thank you, Priya! I've noted everything you shared and passed it on to the team." The one text this check never touches is your own rule message under End, which goes out exactly as you wrote it: do not promise contact in it.

Safe mode

Safe mode is a brake. With it on, the bot asks only your must-know questions and nothing else:

• It uses no facts. A lead who asks something is told you will check with the team, and those questions are not added to the list of questions your bot couldn't answer, because the bot was not allowed to use your answers.
• It does not sell. No offers, selling points, upsell or next-step question, only the check with the team where a rule is missed.
• It asks no optional questions. Once every must-know is in, the intake finishes with what there is. The read-back still happens when something was guessed.
• It suggests no options to a lead who is unsure. It explains the question in a sentence and asks it again, and if the lead still cannot say, it carries on.
• It shows no records (showing options) and makes no peek: a qualified lead goes to your team as they did before you turned options on.
• It offers no times (booking): a qualified lead goes to your team. A time the lead already holds in front of them still books when they tap it.
• It never states your limits. A rule's Tell the lead this limit switch is set aside while Safe mode is on: the bot knows a limit exists, but a figure like "the least we usually work with is $400" is not said, and the check on replies would hold one back if the AI tried. It says only that the answer is a bit outside what you usually do. The switch works again when you turn Safe mode off.

Your rules, the verdicts, the emergency message and the hand-offs to your team work as before, and the lead is not told anything is different. It applies from the next reply, to every lead.

You can switch it on yourself under Settings, Conversation, Safe mode. It also switches itself on when the check above has had to replace a reply with the fixed line five times in an hour in your workspace, which means something keeps asking your bot to say what it is not allowed to say. Conversations in the playground are not counted. When it does:

• the first card of Your bot's to-do says "Your bot is in Safe mode" and when it switched on, and the banner above the inbox says it too;
• the Activity log has a row, "Safe mode switched on by itself after repeated blocked replies," made by the system;
• nothing is emailed or messaged to you.

To turn it off, switch it off in Settings; the card goes with it. Before you do, look at the leads flagged Asked about price, Asked about availability or Couldn't answer, which show what the bot was trying to say, and write the missing fact. If the same thing happens again, five more replaced replies in an hour switch it on again.

A lead who comes back

A lead who goes quiet for longer than the quiet gap (Welcome a lead back after (hours) in Settings: 24 by default, 1 to 720) and then writes again is welcomed back, if the bot owns the conversation:

• an intake that went quiet, or one at the completed stage;
• a lead marked Lost (not a fit);
• a lead closed for not replying;
• a lead who was shown options and went quiet;
• a lead who was qualified, matched or won, where nobody on your team has written since the hand over (after a booked call, once the meeting has ended).

The bot replies once, with a one-line summary of the last enquiry in the lead's own words, and asks whether they want to pick that up or have something different in mind: "Welcome back, Priya! Last time we talked about a December trip to Spiti for a group of friends. Would you like to pick that up, or are you planning something different this time?" If the lead's message already says which, it does not ask. A bare "thanks" or a thumbs up is not a return. A lead who is asked and then goes quiet is nudged like any other quiet lead (above).

• Pick it up. The questions carry on where they stopped, with a few warm words first. "No rush", "still thinking about it" or "just browsing" counts as picking it up, at their pace: it is not a new enquiry and not a decline. A lead who had been marked Lost is reopened with every answer kept, and the bot asks whether anything has changed; the verdict is worked out again when the questions are done. A lead who was qualified, matched or won tells your team they are back, and the bot steps aside.
• Something new. The bot saves the old enquiry, starts a fresh one with a clean slate and asks the first question again, exactly one question and without introducing the business a second time. The old matches stay with the old enquiry. For a lead who was already qualified, your team gets a note that a new enquiry started. The lead page shows Earlier enquiries under the conversation: one line each with its number, dates, outcome and a few key answers, which you can expand to see every answer.
• Not welcomed back: a conversation a person has taken over, a held lead (your team owns it, so it is flagged as before), an opt-out, a lead who is being offered times or whose call is still to come, and any message inside the gap. A lead marked Lost after you chose End who writes inside the gap gets one short reply, "Thanks — I've passed that on.", and nothing is raised to your team.

When the bot hands a lead to a person

Sometimes the bot stops and waits for your team. The lead moves to Needs you in the Inbox, and the Inbox row, the conversation and the bell all say why:

A bare "thanks", "ok" or a thumbs up does not count as a reply that needs answering. Meeting booked is a note, not a request: it never counts under Needs you.

The five reasons from Asked about price to Bot going in circles are signals: the bot does not stop for them, so the lead keeps its stage and the conversation carries on while your team is told. Each has its own switch in Settings, all on to begin with: Flag leads who ask about price, Flag leads who ask about availability, Flag urgent leads, Flag leads who sound annoyed and Flag conversations going in circles. With price or availability switched off, the bot also stops promising a follow-up on it. A budget the lead gives you is an answer, not a price question, and is never flagged.

Replying is taking the conversation over: the bot pauses and stays paused until someone hands the conversation back. Agents press Take over to open the message box; admins and owners can write straight away, and sending pauses the bot. Hand back to (with the bot's name) gives it back when it may carry on; handing back a conversation where the lead asked for a person picks the questions up where they stopped.

A reply has to fit in one message on the lead's channel: 4,096 characters on WhatsApp, 2,000 on Messenger, and 1,000 bytes on Instagram (about 333 characters of Hindi or Arabic). The message box counts what is left and will not send a reply that is too long. It is also switched off once the lead last wrote more than 24 hours ago, because the platform will not deliver a plain message after that; it opens again when they write in.

On every plan, a waiting lead shows under Needs you and on the bell. On the Pro plan, your team is also emailed or messaged on WhatsApp at the addresses in Email addresses to notify and WhatsApp numbers to notify in Settings. WhatsApp reaches a number only within 24 hours of its last message to your business number; outside that, the alert goes by email only. The alert comes once when a lead starts waiting, not again for each message, and at most once per message even when several reasons apply (a lead that qualifies and asks about price in the same message sends the qualified notification only). Turn these off with Tell the team when a lead needs a person; the notice that the AI allowance is used up always goes out, because it stops the bot for every conversation. If nobody has replied after Remind your team after (minutes) (30 by default; 0 turns it off), the team gets one reminder. If someone replied in the last ten minutes, a new message from that lead flags the conversation but does not notify the team again.

The two reviews close themselves if nobody acts: after Days your team has to decide (3 by default) the lead closes as lost and no message is sent. Set it to 0 and a missed-rule review waits under Needs you until someone decides; a lead who stopped replying is then closed as not replying as soon as the nudges run out, with no review. Replying to the lead cancels the close. A missed-rule review keeps its two buttons until you choose one. "Review before closing" sends no alert and no reminder; it waits under Needs you.

Response time on Today shows how fast your team answers these leads over the last 30 days. Qualified → first contact measures speed to lead: the time from qualification to the first reply, or to marking a match as contacted.

When the bot stays quiet

The bot does not answer while a person has the conversation, while it waits on your call (a missed-rule review, or the lead asking for a person), once the month's AI replies are spent, or once the conversation has ended: qualified and handed to your team, or closed as not a fit. (A qualified lead who has just been shown options, is being offered times, or has a booked call is the exception: the bot still answers them.) The other exception is an emergency: it is answered with the fixed emergency message, whoever has the conversation (how).

A message it does not answer costs nothing: no AI reply is spent. It is still stored and shown in the conversation, and the lead waits under Needs you as "Replied, unanswered", unless it is already waiting there for another reason. Two cases differ:

• A lead whose enquiry the bot has handed to your team while the conversation is still the bot's (nobody has taken it over or written in it) gets one short fixed line when they write again, "Thanks — the team has your message and will pick it up from here." It is sent at most once a day, whatever they write, and never when a person has the conversation or has written in it, because then your team is the one talking. It is in the lead's language and costs no AI reply.
• A lead closed as not a fit does not appear under Needs you, because a lead the rules turned away never needs your team. If you chose Keep talking, the bot answers as above. If you chose End, a message inside the quiet gap gets one short acknowledgement, and a message after the gap gets the welcome back. You can also open one yourself with Reconsider on the lead's page.

A lead who went quiet and comes back:

• Still under "Review before closing". The lead writing again ends the review, and the bot answers them and carries on from where the conversation was.
• Already closed as lost for not replying (the review ran out, or you pressed Close as lost). Inside the quiet gap their message reopens the lead: it is back in conversation, the bot answers, and the questions pick up where they stopped. After the gap they are welcomed back. Nobody is notified, because the bot has it. If a person had taken the conversation over, it stays theirs instead: the lead is flagged and your team is told.

A bot whose replies are marked Not delivered is not quiet: it answered, and WhatsApp, Instagram or Messenger refused to deliver it, so the customer saw nothing. The reason is under each reply, with a What to do link. The usual causes are a WhatsApp number whose display name Meta has not approved yet, and a reply after the 24-hour window has closed. The same goes for a message the bot meant to send after the window closed, such as the message of a rule that turns a lead away (sent when the lead goes quiet) or your own booking page link: it is kept in the conversation as Not delivered, with the 24-hour window as the reason, so you can see what the lead was not told and send it yourself if you want to.

"Why you ask" is the steering wheel

Why you ask is one sentence per question that says what the answer is for. It is not shown to the customer, and the bot never copies it. From it the bot decides how to word the question, how much it matters, and what a useful answer looks like.

A must-know question with no reason is the most common cause of a conversation that loops or accepts something useless. Your core questions already have a reason from the platform, and the templates write one for every question they ship. For a question you add yourself, write one.

Question:       What is your budget?
Why you ask:    So we only suggest options that fit what they want to spend
If not sure:    Ask what it needs to cover; never suggest a figure yourself

Question:       When do you want to start?
Why you ask:    So we can check what is free in their month
If not sure:    Offer this month, next month or later

Write the reason as you would explain it to a new colleague, not as a description of the field. "The customer's budget" tells the bot nothing it could not infer from the label. The older AI prompt hint is only used when there is no reason. See What your bot finds out.

Personality is a tone, not a costume

The personality field changes register: brisk, warm, formal. It does not change what the bot asks or what it will accept, and it is not a place to put instructions. It also never overrides the bot's honesty (it still says it is your assistant if asked, and it still invents no discount, no scarcity and no figure, however hard you ask it to sell), its safety rules, or its short replies. The same goes for the push level: it changes how often the bot asks, never what it may say. "Always ask about budget first" belongs in your intake questions and their order; written into personality it competes with the conversation the bot is actually trying to have. To change how it sounds, write a style example (how): showing it one good conversation steers it better than describing one.

Test it, do not reason about it

Every change here is one message away from being observable. Open the playground, answer badly on purpose, and see what happens. That is a faster loop than reading this page twice, and it is the same pipeline your customers get.