AI Answers

Sidebar: Communication → AI Answers

Module: AI Answers — Plus and Prime plans, switched on by the owner or an admin on the Modules screen. It works inside the Inbox, so it needs the Inbox module too. If the plan changes to Free, your catalogs and automations are kept, but the automations stop answering and the tester stops running until the plan includes AI Answers again.

What it is

The same questions arrive every day: what is your address?, do you take cards?, how much is the cleaning? AI Answers lets you write each of them down once — with the ways your customers phrase it — together with your answer, and then answers them by itself, on WhatsApp and on FlyB Chat.

How the AI answers — and what it never does

The AI never writes the answer to a question — unless you switch on The AI should always answer on an automation, the one exception, described on its own below. When a customer writes, it compares their question with the questions of your catalog and asks itself one thing: which of these prepared questions asks for exactly what the customer wants — if any? When one does, the answer you wrote for it is what leaves, with its variables filled in for that contact. Nothing invented, nothing rephrased — and when the customer writes in another language than the catalog’s, that same answer is sent translated into theirs, every translation visible and editable in the catalog. See Customers who write in another language.

Before comparing anything, the AI reads what the message is, with the conversation around it:

(This reading of the message, the rewritten questions and the translations are being rolled out gradually — see below.)

Three things stand between a question and an answer that does not fit it:

• The minimum match. Each catalog has one — 90% by default. If the customer’s question does not reach it with any question of the catalog, no message is sent and the conversation is left to your team (unless the automation has The AI should always answer on). See The minimum match.
• Questions that only make sense with the conversation before them are never answered as they are. “And how long does it take?”, right after you described your sofa cleaning, does not name what it is about; the catalog’s “How long does the cleaning take?” might be about something else. The AI recognises these follow-ups: it first rewrites them from the conversation into a complete question and judges that one as strictly as any other — or, when the conversation does not say what the message refers to, leaves them to your team. See Follow-up questions.
• A question that is more specific than yours is not a match. “How much is the carpet cleaning?” is not “How much is the cleaning?”, and “Can I pay after the service?” is not “Which payment methods do you accept?”. The AI is told to answer only when your question covers what the customer asked — not when it merely shares the topic.

And the automation adds its own: it can stay quiet while somebody from your team is in the conversation (if you switch that on), it never sends the same answer twice within an hour, and it never answers a question the team has already answered. See Answering from an automation.

Catalogs

A catalog is a group of questions on one subject — General questions, Sales, Support, or one per service. Your catalogs are listed on the left of the screen, each with how many questions it holds, its minimum match, how many answers it sent this month and how many automations answer from it (or Not in an automation, when none does yet). Click one to open its questions on the right.

New catalog asks for a name and opens the new catalog ready for its first question. The pencil beside the name renames it, and the bin deletes it — with its questions; the automations that answered from it keep running with their other catalogs, and one left with none answers nothing (its row says so). Answers already sent stay in the conversations.

The catalog’s language

Each catalog is written in one language — its questions and its answers — shown under the catalog’s name: Written in Português (Brasil), beside the reminder that customers who write in another language receive your answer translated. A new catalog starts in your organization’s language, read from its country (a Portuguese-speaking country reads Portuguese, any other English); Change language picks any of the platform’s eight — Português (Brasil), English, Español, Français, Italiano, Deutsch, Polski and Türkçe.

The language belongs to the catalog, not to the organization: a business may keep a Spanish catalog for its Spanish-speaking customers beside a Portuguese one, and one automation may answer from both. Changing a catalog’s language deletes the translations already made from it — they were translated from the old language — and each is made again the next time a customer needs it; the screen tells you how many before you confirm.

Greetings and farewells

A customer who writes only “oi”, “bom dia”, “obrigado!” or “até mais” asks nothing a catalog can answer. Each catalog can hold standard replies for them — under the catalog’s language, in Greetings and farewells:

• Reply to greetings — the text sent to a customer who only greets (“Olá! Como podemos ajudar?”);
• Reply to farewells — the text sent to a customer who only says goodbye or thanks (“Obrigado pelo contato! Estamos à disposição.”), with one more option, Resolve the conversation when the farewell is sent (off by default): the conversation then leaves the Inbox queue as resolved — unless the customer wrote something else in the meantime, which keeps it open for your team — and if the customer writes again a new one opens.

Both are off by default. Switch one on, write the text — up to 1,000 characters, with the same variables an answer takes — and save. The AI only recognises that the message is a greeting or a farewell; it never writes the reply: what leaves is your text, translated into the customer’s language like any answer when they write in another one (the translation is listed under the reply, to read, edit or delete). It replies only when the message calls for it — not to a customer who answers the business’s own greeting or goodbye, not to an “ok”, and not twice to the same contact within an hour.

With a reply switched off (or with no text), such a message gets nothing, and the AI does not go on to look for an answer in the catalog either. Switching a reply off removes its text and its translations.

When an automation answers from several catalogs, the first of them (in the order they were created) that has the reply is the one sent.

The minimum match

How sure the AI must be that the customer is asking one of the catalog’s questions before its answer is sent — a percentage from 50% to 99%, 90% by default. It is shown beside Test, at the top of the catalog; the pencil next to it opens the slider, which saves when you release it, and Confirm saves it (a value set with the keyboard too) and folds it away again:

• Raise it for more precision: fewer automatic answers, and only the clearest questions answered.
• Lower it to answer more: below 75% the AI starts accepting questions that are only similar to yours, and the screen warns you.

The default comes from FlyB’s own tests: questions that one of the catalog’s questions really answered scored above it, and questions that only resembled one — “Do you have a physical store?” against “What is your address?” — scored below it. A catalog keeps the minimum it was saved with: one created while the default was 80% keeps 80% until you change it.

Every automatic answer shows its match in the Inbox, so you can see where your answers usually land before changing the number.

Questions and answers

New question opens the form:

• Question and phrasings. The question, and the other ways customers ask the same thing — “What is your address?”, “Where are you located?”, “How do I get there?”. Up to 10 phrasings, each at least 8 characters long (and up to 200). They all lead to the same answer: adding phrasings never splits the AI’s choice, it only teaches it how this question sounds.
• Answer. Exactly what the customer receives — up to 4,096 characters, the length of a WhatsApp message — with the same Insert variable picker as a free-form automation message: {{contact.first_name}}, your company’s details, the contact’s next appointment, and so on.

A phrasing that is already on another question of the same catalog is refused where it is typed: the same words on two questions would split the AI between two answers, and neither would ever be sent.

Each question shows how many times its answer was sent, and when last. Hover over it to edit or delete it. With more than a few questions, the search box above the list finds one by its words or its answer.

Adding many questions at once

Bulk add (beside New question) takes a whole list pasted as text — typed by you, or written by an AI assistant such as ChatGPT or Gemini. Nothing is saved until you have reviewed it.

The format:

What is your address?
Where are you located?
> We are at 100 Flower Street, downtown.
Free parking for customers.
---
Which payment methods do you accept?
Do you take cash?
> We accept credit and debit cards, bank transfers and cash.
---

• One question per line. A question cannot break across lines: every line is another phrasing of the same question, and the first line is the main question.
• > opens the answer. The > character at the start of a line separates the questions from the answer. Only the first line of the answer takes it.
• The answer may run over several lines, with emojis and variables such as {{contact.first_name}}.
• --- separates one group from the next. A line with only --- ends a group, so never use it inside an answer.

The same limits as the form apply — up to 10 phrasings of 8 to 200 characters, answers up to 4,096 characters, 200 questions per catalog. Blank lines are ignored, and if you paste an AI’s whole reply, only what is inside its code block is read.

Have an AI write the list. The dialog holds a ready-made prompt, already carrying your company’s name and sector: copy it, paste it into the AI you use, add what the AI should know about your business (services, prices, address, opening hours — or your website), and paste the list it returns. The prompt asks the AI to write [FILL IN] wherever it lacks a fact instead of inventing one, and an answer that still holds it is not imported until you replace it with the real information.

The text is checked as you type: under the box, each problem names its line — a group with no >, a phrasing that is too short, a forgotten --- between two groups (a second > in the middle of an answer), the same phrasing in two groups — and a click selects that line. Then Review shows every group as it will land:

A group is recognised as a question already in the catalog when they share a phrasing. You choose what happens to those: Keep the one in the catalog (the default) or Replace it with the text. A group whose phrasings belong to two different questions of the catalog is left out, since the AI would be split between their answers.

New questions are added at the top of the catalog in the order of the text. Groups with a problem do not stop the rest: after the import they stay in the box, ready to be fixed and added. The text is also kept in the browser tab until it is imported, so closing the dialog loses nothing.

Start from this catalog fills the box with the questions the catalog already holds, in the same format — to edit them all at once in a text editor or with an AI, and paste them back with Replace it with the text selected.

Testing a catalog

Test opens a box where you type a message the way a customer would. The AI runs exactly the analysis it runs in a real conversation — nothing is sent — and shows:

• whether the catalog would answer, and if not, why (no question matched; the closest one stayed under the minimum; the message is a follow-up);
• the closest questions, each with its match;
• the answer that would be sent.

A test message stands alone, with no conversation around it, so a follow-up such as “and how much is it?” is refused here exactly as it would be in a chat (there is no conversation to rewrite it from). Each test counts towards the daily limit.

The result also says what the AI read the message as (a question, a greeting, a farewell…), in which language, and the complete question it compared with your catalog. A message in another language than the catalog’s shows the answer translated — marked translated to EN, with the catalog’s original beneath it — and that translation is a real one: it is saved on the question like any other and reused for your customers. With standard replies, a greeting shows the reply that would be sent — and a farewell whether the conversation would be resolved.

Answering from an automation

A catalog answers nobody until a message automation answers from it. In Communication → Message automations, create a rule with the trigger Customer’s first message or Any message received — on WhatsApp or on FlyB Chat — and, under The message, choose Automatic answer from a catalog (the option with the AI chip). Tick the catalogs it answers from (up to 10, holding up to 250 questions together); New catalog creates one without leaving the rule, and the pencil on a catalog opens it to add questions.

Everything else about the rule keeps its meaning: the numbers it covers, the keyword filter — All messages by default, or only messages with the words you choose — the conditions, quiet hours. A few things work differently:

• It waits for the whole question. Customers often write in pieces — “hi”, “what’s the price”, “of the cleaning?” — and a voice note’s words take a moment to be transcribed. So the AI reads the question 10 seconds after the customer’s last message, and every new message restarts the count. With a delay on the rule, it waits that long after the last message instead.
• It reads everything the customer wrote since the business last answered — up to the last hour and ten messages — with the latest messages of the conversation as context, and answers it once. A question your team (or another answering rule) has already replied to is never answered again.
• It answers every question. Once per contact does not apply — a customer asks more than one question — and is not shown. The minimum gap still can space its answers to the same contact; leave it at zero to answer every question.
• The same answer is never sent twice to the same contact within 60 minutes. A customer asking again that soon was not helped by it, and the question is your team’s.
• Stay quiet while the team is in the conversation (off by default — switch it on if you want it): if somebody from your team wrote in the conversation in the last 30 minutes — from the Inbox, or from the business’s own phone — the AI does not answer.
• On WhatsApp the answer goes out from the number the customer wrote to, as a free-form message (it is always inside the 24-hour window, since the customer has just written — unless a long delay outlives it). Voice notes are read by their transcription when your account transcribes them.

The AI should always answer

An option of the automation, under the catalogs it answers from, off by default. With it off, a question no catalog answers well enough is left to your team. With it on, the AI writes an answer itself — and only then: when the message really calls for a reply and no catalog question reached the catalog’s minimum match (or the message is a follow-up the catalogs could not answer on their own). Everything the catalogs answer is still sent exactly as you wrote it.

The AI writes from your material and nothing else:

• the AI context of your company — a text you write in Settings → Company describing what you offer, prices, hours, policies and what you do not do. The option cannot be switched on while it is empty: the automation says so and Fill it in now opens the company settings in a new tab;
• the company details from the same screen — business area, website, phone, e-mail, address and social media profiles — so a customer asking for your phone or your Instagram can be given them;
• the catalog questions and answers the AI found closest to what was asked;
• the latest messages of the conversation.

When two parts of this material disagree, a catalog answer comes first, then the AI context, then the company details. Keep them consistent anyway: the AI only sees the catalog answers closest to the question, so an answer it was not shown cannot correct an address it was.

A “yes” needs your material to name exactly what was asked — a similar or neighbouring feature is never a reason to say yes. Asked “Do you have an AI chatbot?”, an AI context that mentions “WhatsApp service, a website chat and an AI website builder” should get the holding reply, not a yes. Nor does the AI offer to tell more: it answers what was asked, or holds.

It answers in the customer’s language, briefly. Together with its answer, the AI says whether it actually knew it from your material — and when it did not (or knew only part of it), it does not guess: the customer receives “Vou verificar essa informação e logo te retorno” in their own language (at most once an hour per customer), and the question goes to the AI review for your team to answer and, ideally, to add to a catalog.

Before an answer the AI wrote is sent, every number (price, hour, quantity, phone), link, e-mail address and variable in it must appear in your material — never only in the customer’s own words, so “do it for R$ 50?” cannot become “yes, R$ 50”, and never only in an earlier answer the AI wrote, so a guess is never repeated as a fact. An hour may change clock: “6pm” for your “18h” is the same fact. An answer that fails this goes to the next model of the chain, and when none passes, the holding reply goes instead.

The holding reply is always the same sentence, never one the AI wrote: “Vou verificar essa informação e logo te retorno”, translated into the customer’s language.

Customers who write in another language

Your catalog is written in one language, and your customers do not all write in it. When a customer asks one of your questions in another language, the AI still recognises the question — it matches across languages — and your answer is sent translated into the customer’s language. Twenty-six languages are recognised: the platform’s eight plus Chinese (Mandarin), Hindi, Arabic, Bengali, Russian, Japanese, Korean, Urdu, Punjabi, Marathi, Telugu, Persian, Indonesian, Vietnamese, Tagalog, Thai, Swahili and Hausa.

• The language is recognised, not guessed. The AI says which language the customer wrote in — by the grammar, not by the names or borrowed words in it — and the answer is translated only when that reading is confirmed a second time, on the complete question the message was rewritten as (a standard reply to a greeting or a farewell, which has no question to rewrite, only when the first reading is certain). A message written in Chinese, Japanese, Korean, Arabic, Devanagari or Cyrillic characters needs less certainty when the AI names a language written in them: the characters already rule out the languages that look alike. A short or mixed message (“ok”, “tem free trial?”, “oi, do you accept pix?”) is read as the catalog’s own language. Regional variants count as the same language: a customer from Portugal reading a Brazilian Portuguese catalog receives your answer as written.
• Your answer is translated once, then reused. The first time a customer asks a question in a new language, its answer is translated and saved on that question; every later customer in that language receives the saved translation. Nothing is translated ahead of a customer needing it.
• What must not change, does not change. Variables such as {{contact.first_name}}, links, e-mail addresses and phone numbers are set aside before translating and put back exactly as you wrote them, and the translation is checked before it is used: it must keep every one of them, the same numbers and prices, and fit a WhatsApp message. Variables are filled in for each contact after translating, so one translation serves every customer (a date a variable fills in still follows the catalog’s language).
• Editing an answer retires its translations. A translation is tied to the exact words it translated; when you change an answer (or a standard reply), its translations are deleted with the change and made again from the new words the next time they are needed — a translation never carries an old price.

Each question lists its translations under its answer — EN · automatic, ES · edited… Click one to read it, and how many times it was sent. Since translation itself is always on, this is where you fix one that reads badly: edit it (your version is kept until the answer changes, marked edited) or delete it (the next customer in that language receives a new automatic translation).

When an answer cannot be sent in the customer’s language, nothing is sent and the conversation is left to your team — never your answer in a language the customer did not use:

• the customer wrote in a language outside these 26;
• the translation did not pass the check (a link, a variable or a price would have changed);
• the organization’s daily limit of translations was reached.

The history says which, and the Inbox chip says when an answer went out translated.

Follow-up questions

Customers often ask in pieces: your team writes “We have a basic and a full cleaning”, and the customer answers “how much is it?”. On its own that message names no subject, so the AI never answers it as it is (see How the AI answers).

So every message that asks something is first rewritten, from the last messages of the conversation, into one complete question — and only that question is compared with your catalog, as strictly as any other, with the same minimum match:

• a follow-up gets its subject: “how much is it?” → “How much is the basic or the full cleaning?”; a customer still waiting (”?”, “any news?”) gets their earlier question back;
• a message that already stands on its own is kept as it is (typos may be fixed), and a greeting around a question is dropped: “Oi, bom dia! Quais são os planos?” → “Quais são os planos?”;
• two questions in one message stay together — and since no single answer of yours covers both, it is left to your team rather than half answered.

The rewrite never sees your catalog, so it cannot steer the question towards one of your answers; at worst it makes a message not match. When the conversation does not say what the message refers to, it is left to your team.

A follow-up that reaches the AI review — with The AI should always answer on, when the AI could not answer it either — shows both texts: what the customer wrote and the rewritten question. The history shows the rewrite too.

In the Inbox

An answer sent by a catalog carries, under the bubble, a purple chip with the match — AI 92% — beside the automation’s ⚡ badge. Hover over it to see which catalog and which prepared question the customer’s message was matched to, and the catalog’s minimum at that moment. So a colleague reading the thread can tell an answer the AI was sure of from one that only just cleared the minimum. An answer sent translated says so — AI 92% · translated to EN — and the hover also shows the catalog’s original answer. A catalog’s standard reply to a greeting or a farewell carries AI · greeting (or AI · goodbye) instead, with the catalog it came from on hover. An answer the AI wrote under The AI should always answer carries an amber Answered by AI — nobody checked it, so read those — and its holding reply AI · holding reply, which means the customer is waiting for your team.

A customer message the AI looked at and could not answer, so that nothing was sent, carries a red Not answered by AI chip under it, with Create answer and an × to ignore it — the same two choices as the AI review.

Adding a reply to a catalog

When your team answers a question by hand, that answer can become a catalog answer in one click. Hover over a reply typed by a team member — not an automation, a template or an attachment — and press Add as an answer in the AI catalog (the book with a plus). With more than one catalog, FlyB asks which. The catalog then opens with a new question already written: the customer’s last question as the question and the team’s reply as the answer. Adjust both — a question usually needs tidying, and the answer may deserve a variable — and save.

The same button sits on an answer the AI wrote and on its holding reply — Create the real answer in the AI catalog — and on a message it left unanswered. There the catalog opens with the customer’s question and a blank answer: what the AI wrote is never copied into a catalog, because nothing proved it right; the answer has to be your team’s. Saving it takes the question off the AI review.

The button appears while the module is active, you may edit catalogs, and at least one catalog exists.

The AI review

What the AI could not answer — and what it answered on its own — waits for your team in one list, the AI review: the button with a list icon in the Inbox toolbar, beside Hunt AI, whose red number counts the questions still without an answer; and the same list under the catalogs. It has two tabs:

• No answer — questions a customer asked that no catalog answered: the AI looked, found no question close enough, and the customer really was asking for something (a “thank you” or an “ok” never appears), so nothing was sent; and, with The AI should always answer on, the questions the AI told the customer it would check (“A IA disse que ia verificar”).
• Answered by AI — answers the AI wrote under The AI should always answer, each with what it said, so you can judge it.

Each item shows who asked, when, through which automation, and the closest catalog question when there was one. A follow-up shows both texts: “how much is it?” and, under it, the question it was rewritten as.

• Create answer opens the catalog with the question already written and a blank answer for you to write — the closest question’s catalog is suggested first, and a follow-up starts from its rewritten question, since the customer’s own words name no subject. Once saved, the item leaves the list.
• Ignore takes it off the list when the question does not deserve an answer in a catalog.
• Open conversation (in the Inbox) takes you to the thread it came from.

The list is not a copy of anything: the state lives on the customer’s message itself, so ignoring or answering it anywhere — the Inbox, the thread, the catalog screen — is the same for everyone. It shows only what you may see in the Inbox: the contacts you own, if you are restricted to your own contacts, and the WhatsApp numbers you have access to. This is how a catalog grows from the questions people actually ask.

History

The 🕘 history of an answering automation lists every time it looked at a customer’s message: what the customer asked (with the language it was written in), what the AI read it as when it was not a question (a greeting, a goodbye, an acknowledgement), the complete question it was rewritten as, the closest prepared question and its match — marked translated to EN when the answer went out translated — the standard reply sent to a greeting or a farewell, the answer the AI wrote (AI wrote:, with how well it kept to your material) or its holding reply, and, when nothing was sent, why, in a sentence:

Limits

Who can use it

AI Answers is its own permission. Writing what the business answers every customer is not the same job as wiring the automations that send it, and teams often hand the two to different people.

See Settings → Team to change it per member.

Privacy

To decide, FlyB sends TypeSafe — the AI provider listed in Sub-processors — the customer’s latest messages (a voice note as its transcription), the latest messages of the conversation, your organization’s name and sector with a short description of it (the first sentence of your AI context and a few of your catalog questions, or the catalogs’ names), and the questions of the catalogs the rule answers from. Before anything leaves, e-mail addresses, numbers of eight digits or more and the contact’s own name are removed from the messages. The contact’s record and your catalog’s answers are not sent — except, with The AI should always answer on, where an answer the AI wrote is checked by TypeSafe: then that answer, your AI context, your company details and the closest catalog answers may be sent too.

Translation and follow-up rewriting use a language model reached through OpenRouter on FlyB’s own account — also listed in Sub-processors, with the model providers it routes to (a primary one, and two backups used only when the one before fails):

• to rewrite a question: up to the last six messages of the conversation before it and the customer’s message, with the same removals — not your organization’s name, not your catalog and not the contact’s record;
• to translate an answer (or a standard reply): the text you wrote, with its variables, links, e-mail addresses and phone numbers replaced by placeholders before it leaves — nothing about the contact;
• to write an answer, only where an automation has The AI should always answer on: your company’s AI context, name and company details (business area, website, phone, e-mail, address and social media profiles), the closest catalog questions and answers, up to the last six messages of the conversation and the customer’s message, with the same removals — not the contact’s record. The answer may then be checked by TypeSafe against the same material.

FlyB asks OpenRouter to use only model providers that neither store nor train on what they receive. See the Privacy Policy.