Side menu, top bar, and dashboard shortcuts: Ctrl K search, notifications, dark theme, and profile

"O que a IA faz por dentro" (What the AI does behind the scenes)

Why it replied that way, what it can and cannot do on its own, and where you change its behavior.

The reply the lead receives does not come from one place. Before it exists, the message passes through several specialists, each with a narrow job: one finds out what the person wants, another gives the score, another writes, another checks what was written before sending.

You do not need to memorize this to use the dashboard. You need it when one of these three things happens: it replied something you did not expect, it called you at a strange time, or you want to change the way it behaves and do not know which button controls that. This guide is for those three moments.

Where you see the AI at work

Each decision is recorded on the lead record, in the "Jornada do lead" (Lead journey) card — the timeline on the right. That is where the names in this guide are. The same list, shorter, appears in the "Jornada" (Journey) section of the drawer that opens when you click a contact in the "Base de contatos" (Contact base).

Not every line has the third part. Whoever has something to justify justifies it: the qualification says which signals it captured and which are missing, the audit says what it cut, the handoff says why it called a person. The others conclude and move on.

What it understands before replying

Every incoming message is read by five different readers at the same time, before any reply is written. None of them talks to the lead: they only draw conclusions.

Intent classifier

What the person wants in this message: greeting, product question, price, purchase intent, scheduling, support, asking for a human, or leaving. It is what appears as "Intenção prevista" (Predicted intent) in the conversation's right panel, and it is what matches the lead to an offer in your catalog.

Sentiment analysis

The mood of whoever wrote: positive, neutral, negative, or frustrated, with intensity. High frustration calls a person on its own, without waiting for the lead to ask.

Data extractor

It notes what was said in passing and that no one would type by hand: company, job title, email, another phone number, the deal value. That is why the record fills itself in without you touching it.

Objection Detector

It recognizes when the message carries sales resistance — “that's expensive”, “I'll look at it later”, “I already use another one” — and stores the type and the sentence said. It feeds the "Pergunte à IA" (Ask the AI) screen; a price question does not count, only resistance.

Audience recognizer

It finds out which of your audiences that person resembles, so the right questions get asked. It only comes into play if you registered more than one audience, and it never invents one that does not exist.

Where you are in control: the audiences and what each one needs to find out are in "Qualificação" (Qualification); the offers and the intents that match each one, in "Intent & catálogo" (Intents & catalog). Sentiment, data extraction, and objections have no settings — they read what is written.

How it decides who is worth your morning

The "Qualificador BANT" (BANT Qualifier) rereads the entire conversation with each new message — not just the last sentence — and looks for the signals you declared: what the person needs, how much they can spend, whether they are the decision maker, and by when. From this comes the score from 0 to 100 and the temperature.

Where you see it: the "Qualificação ao vivo" (Live qualification) ring in the conversation's right panel, the qualification card on the lead record (*Score 82 (Hot) — 3 of 4 signals*), the color in the Inbox list and on the pipeline card, and the line pontuação 82 · banda Quente in the journey.

The reason behind the score is not a justification text: it is the lead's sentence that proved each signal, stored next to it. A signal with no sentence shows as *not yet captured* — it means it has not asked yet or the person has not answered, not that the answer was bad.

Where you are in control: everything is in "Qualificação" (Qualification) — which signals apply to each audience, how it asks, and how much each one weighs in the score. Changing it there does not rewrite anyone's score: it applies from each lead's next message onward.

With the Revenue AI module on, a second axis joins the score: the "Aderência ao cliente ideal" (Fit with the ideal customer). The Ideal Customer Profile compares the lead with what you declared in "Configurações" (Settings) → "Cliente ideal" (Ideal customer) and returns 0 to 100 with a sentence — on the pipeline card (ICP 82), on the record, and in the dossier. It runs at most once a day per lead, and only when the lead said something new.

How it writes the reply

The reply is written by the "Conversa SDR" (SDR conversation), and it writes with what you gave it. Nothing more than that:

• Who it is — its name, the company it says it represents, and the tone of voice, in Personas A/B and "Voz & tom" (Voice & tone). If the account the message came in through has a linked persona, that persona is the one speaking there; with no link, it is the company's persona.
• What your company does and the ready-made answers you wrote in the persona.
• The active offers in the catalog, with the price the way you wrote it.
• The excerpts from the ["Base de conhecimento" (Knowledge base)](ajuda:conhecimento) that relate to the question in that message — not the whole base, just the pieces that serve. The search also looks at the latest turns of the conversation and at the excerpts the AI received for the previous reply: when the customer asks only "is there a discount?" or "give me the price of the 2" right after the AI offered a product, that product's excerpt stays at hand. The new question still leads: "do you deliver?" brings up what talks about delivery first. And the check of the reply also receives the excerpts from the previous reply: if the AI repeats a price it just gave, it is not erased just because the customer changed the subject. And, if you marked documents for specific accounts, only those that apply to the account that received the message.
• What has already been agreed with that lead, which is the memory explained just below.

Where you see it: the message itself, with the AI's name right above the bubble. When an agent takes over, the name that appears becomes the person's — the lead never sees the AI's name signing a human message, nor the reverse.

Some replies arrive with an audio first — and there it is not the AI speaking. It is a recorded reply: the voice of someone on the Owner's team, recorded once for one of the usual questions ("how much does it cost?", "how long does it take?"). When the AI is sure the lead asked that question, it sends the recording before the text; when in doubt, only the text goes. Where you see it: in the Inbox, the audio bubble carries the label "Resposta gravada" (Recorded reply), with the title and — when you hover — the voice of who recorded it; in the lead's journey, the line "Resposta gravada enviada" (Recorded reply sent) logs which one went out. Where you are in control: in the "Respostas gravadas" (Recorded replies) card of "Intent & catálogo" (Intents & catalog).

It checks its own reply before sending it

Once the reply is written, it does not go straight out. A second specialist, the "Auditor de Resposta" (Reply Auditor), reads what was written with a single question: *does this state any fact about this company that is not in what the Owner registered?* Price, deadline, promise, policy, coverage, capacity, any number.

It was taught to separate what is a fact from what is not, because otherwise it would cut the entire conversation. Stating nothing: greeting, asking a question, repeating back what the lead just said about themselves, saying its own name, or saying it will pass the lead along to the team.

A scheduling commitment counts as a fact

Even without a number. Setting a day, promising a reply "by tomorrow", or agreeing to the deadline the lead proposed is stating a deadline, and it must be on file.

Money only passes with backing

Either the amount is in what you registered, or the lead themselves stated that amount as their budget. A number the lead used for something else (number of customers, employees, followers) does not authorize any price.

A lead's question is not a source

"Do you offer 12 installments?" does not give permission to confirm that you do. If the policy is not registered, confirming it is invention.

When it finds a passage without backing, it can only delete. It does not rewrite, swap words, or adjust a number downward: it removes the piece and delivers the rest, and in its place a sentence may go in saying that the detail will be confirmed with an agent. If what is left does not stand on its own, the entire reply is replaced by a safe reply that states nothing: for a question, it says it will confirm the detail with a specialist. With "Conduzir até o pedido completo" (Guide through to the complete order) turned on, neither of the two sentences promises a confirmation: they say the detail is not there and offer a person from the team, and a customer who is ready to buy and gets the safe reply is sent to the team.

Where you see it: in the journey, the Auditor's row says which case that reply fell into — *"resposta fundamentada" (grounded reply)*, *"preço sem lastro corrigido" (unbacked price corrected)*, *"trecho sem lastro corrigido" (unbacked passage corrected)*, *"resposta trocada" (reply replaced)* — and the justification shows the exact passage that was removed. It is the evidence, not a summary.

Two honest notes about its limit. The check never holds up the conversation: if it fails itself, the original reply goes out, because leaving the lead without a reply is worse. And it judges what the reply states about your company — if you registered a wrong price, it will defend the wrong price with the same effort.

It remembers what has already been said

In a conversation that drags on for days, rereading everything on every message is not viable, and an AI that forgets what was agreed last week loses the lead. That is why the "Memória do lead" (Lead memory) exists: a short summary of the relationship, rebuilt from time to time, with who the person is, what they are looking for, the objections they raised, the amounts and deadlines they mentioned, what was agreed, and what the next step is.

It only kicks in on a conversation that has already grown long — an exchange of two or three messages has no memory, and does not need one. And it is built only from what is written in the messages: no deducing a profession, income, or intent that nobody stated.

Where you see it: the "Memória do lead" (Lead memory) row in the journey. It is the stamp showing that the AI updated itself, with the date — which tells you how recent the understanding behind today's replies is. The stored text itself is not shown on any screen.

When it calls a person

The "Política de transbordo" (Handoff policy) is a fixed rule, not an opinion: in five situations the conversation leaves the AI and enters the team's queue, and the reason is written in the queue so you know why. Two of them have a switch: with "Conduzir até o pedido completo" (Guide through to the complete order) turned on, which is how new accounts already start, purchase intent and the "Quente" (Hot) lead no longer trigger a transfer.

Lead asked for human service

They wrote that they want to talk to someone. The conversation leaves right away.

Purchase intent

They signaled that they want to close and are already "Morno" (Warm) or "Quente" (Hot) (or have completed qualification). This is the moment when a human earns money. A "Frio" (Cold) lead who says they want to buy, or who just picks an option the AI offered, stays with the AI: the AI asks the next question and the journey records "Intenção de compra sem evidência" (Purchase intent without evidence). Outside business hours, this rule and the "Quente" (Hot) lead rule put the conversation in the queue without notifying the lead, and the AI stays with them until the team opens. It can be turned off in "Configurações" (Settings) → "IA & persona" (AI & persona): with guidance through to the complete order turned on, the AI stays with them until the order is complete. If the check against your knowledge base rejects the reply, the AI rewrites it up to twice; only on the third rejection does it pass the conversation to the team.

Qualified lead ("Quente" (Hot))

The score reached the hot range and is worth your time. It can be turned off with the same switch as purchase intent.

Low confidence in the intent

It did not understand what the person meant, and insisting would create friction.

Lead needs a person right now

The AI realized that only a person can solve this: the lead asked for a call, got annoyed with the automated service, or asked for something outside its reach.

Frustrated lead

The sentiment soured enough for a person to take over before it gets worse.

"Pedido completo" (Complete order)

With guidance through to the order turned on: the customer answered everything the "Qualificação" (Qualification) asks for. The AI summarizes the order and the conversation goes to the team to confirm; outside business hours, with no business-hours notice to the customer.

Along with the handoff, the "Resumidor" (Summarizer) delivers a ready-made briefing for whoever will take over: the conversation summary, the lead's mood, and the suggested next step. This is what appears in the package on the Handoff screen — nobody has to read fifty messages before replying.

Which language it replies in

The AI serves in Brazilian Portuguese, European Portuguese, Mozambican Portuguese, English, Spanish and French, within what the account accepts (in "Configurações" (Settings) › "IA & persona" (AI & persona) › "Idiomas do atendimento" (Service languages)). It recognizes the language on every message, with memory: a lone word, such as "ok" or a number, does not knock the conversation out of its language. If the lead switches languages midway, the AI switches along, as long as the new language is among the accepted ones. Outside those, it replies in the account's main language. When the lead writes in European Portuguese and the account accepts Portugal and Mozambique, the text does not allow telling one from the other with confidence: the AI uses whichever of the two is the account's main language; if neither is the main one, it uses the first accepted, in menu order (Portugal before Mozambique).

• Fixed messages (out of hours, transfer to the team, goodbye) go out in the lead's language.
• Consent: outside Brazilian Portuguese, the request uses a neutral privacy text; the Portuguese version keeps citing the LGPD.
• Meta templates go out in the lead's language when the account has the same template approved in that language; otherwise, in the template's own language.
• Your automatic messages go out as you wrote them for those who speak the main language; in the other languages, the AI uses the default text in that language.

Where you see it: in the Inbox, the badge with the language code next to the lead's name (only in an account with 2 or more languages and only when the lead speaks another language). Where you control it: Settings, "Idiomas do atendimento" (Service languages) card. The dashboard stays in Portuguese.

The notice and the exit

Compliance LGPD does two things, and both are a legal obligation, not a preference. On first contact, it ensures that the transparency notice goes along with the reply. And when the lead asks to opt out, it ends the conversation immediately: the reply being written is discarded, the farewell is sent, and the AI writes no more to that person.

Where you see it: the rows *"Consentimento registrado" (Consent recorded)* and *"Opt-out registrado" (Opt-out recorded)* in the lead's journey, and the full trail in Compliance. Where you control it: the notice and farewell texts are in "Configurações" (Settings), on the "Atendimento" (Service) tab.

What it does without being asked

Three things happen without anyone clicking anything, and you turn on, turn off, or feed all three:

1. Contacts people who went silent again: A lead stopped responding and the AI writes again, at the interval you defined in the cadences, in "Configurações" (Settings). The message is written on the spot, with that lead's context — it is not a ready-made text that is the same for everyone.
2. Sends your payment link: When the lead signals they want to buy and a matching offer exists, it shares the payment link you registered on that offer, in Intents & catalog. If an offer has no registered link, nothing is sent: it has nowhere to invent a payment address from.
3. Books a meeting on your calendar: An agreed time becomes a real appointment in the Calendar, and that part is verified by rule, not written by AI — it is the difference between a time that exists and a time that merely sounded good in a sentence.

What it does not do

Worth knowing as much as the rest, and it is what usually takes away the fear of leaving it on its own:

• It does not quote a price you did not register. If one slips out in the writing, it is removed before the lead reads it.
• It does not send a payment link that is not yours, registered by you on an offer.
• It does not decide a meeting time on its own — the calendar is checked against what is actually free.
• It does not keep talking to someone who asked to opt out.
• It does not open the transfer through what it writes. What passes the conversation to the team is the handoff policy, before the reply is written. A sentence from it saying it will call someone does not call anyone.
• It does not sign with the name of anyone on your team, and no agent appears to the lead under its name.
• It does not learn on its own from one lead in order to answer another. What it knows about your business is what you registered; what a lead tells it stays with that lead.

The names that appear in the journey

A list for quick reference, just as the names appear on the lead's record. Next to each one comes the moment: "antes da resposta" (before the reply) for the safeguards, "agente" (agent) for whoever reads, decides, or writes.

Intent classifier

Concluded what the lead wants in this message.

Audience recognizer

Chose which of your audiences the lead resembles.

"Qualificador BANT" (BANT Qualifier)

Recalculated the score and temperature using the whole conversation.

"Perfil de Cliente Ideal" (Ideal Customer Profile)

Compared the lead with the profile you declared in "Cliente ideal" (Ideal customer) and said, in one sentence, what matched.

Sentiment analysis

Measured the mood of the person who wrote.

Data extractor

Noted on the record a detail mentioned in passing.

Objection Detector

Noted the objection stated in the message — this is what feeds the Ask the AI screen. In "Modo atendimento" (Service mode) (direct sales) it does not run, unless Closer AI is on.

Compliance LGPD

Let the conversation continue, or stopped everything because the lead asked to opt out.

"Política de transbordo" (Handoff policy)

Decided whether that conversation stays with the AI or goes to the team.

"Resposta gravada" (Recorded reply)

Sent, before the text, the audio your team recorded for the recognized question.

"Conversa SDR" (SDR conversation)

Wrote the reply the lead received.

"Auditor de Resposta" (Reply Auditor)

Checked that reply and said whether it had backing in what you registered.

"Memória do lead" (Lead memory)

Updated the accumulated understanding of that relationship.

"Resumidor" (Summarizer)

Built the briefing for the agent who was about to take over.

"Closer de oferta" (Offer Closer)

Decided whether it was time to send your payment link.

Closer

Negotiated within the discount ladder you defined, handled the objection with your playbook, or called the human with the dossier (Revenue AI module).

"Follow-up SDR" (SDR follow-up)

Wrote the re-engagement message for someone who had gone silent.

It got something wrong. Now what?

Almost every error falls into one of these five, and each has a different place to fix it. Always start with the lead's journey: it tells you which of the five it is.

1. It said something wrong about your company: Fix it at the source, not in the conversation: Knowledge base for information, catalog for offers and prices. From the next message onward it uses the right one — and it now has something to be checked against.
2. It cut too much and became evasive: That is the audit working with little material. Its row in the journey shows the removed passage; register that information and the cutting stops happening.
3. It asked the wrong thing: The question script is per audience, in Qualification. If it used the wrong audience, you can change it on the lead's record.
4. The tone is not your company's: Voice & tone and Personas A/B. If the difference is big, test two versions before switching for good.
5. It should have called you earlier: Take over the conversation in the Inbox and answer "Discordo" (Disagree) in the qualification box — that vote is what feeds the accuracy of the handoff. If it always happens with the same type of lead, the adjustment is in the qualification weights.

In the dashboard: "Abrir a lista de leads" (Open the lead list). Click a lead to open their record and read the conversation journey, decision by decision.