Artificial intelligence

Overview: the three roles of the AI

What sets ChatISP apart is its integration with Ravicor, the ISP's network management system. The AI assistant does not just reply with canned text: it sees the subscriber's record in the ERP, billing, the internet session, the equipment and the Wi-Fi, and with that it diagnoses, performs actions (invoice copy, trust unlock, Wi-Fi password change, equipment reboot) or opens a service order for the technical team when the problem calls for investigation. It serves customers at any time, whenever a flow hands the conversation to it.

In ChatISP the AI appears in three different roles:

On the system screens and messages, the autonomous AI is called the assistant. In the chat, every message it sends is identified as coming from the AI, and everything it looks up or performs is recorded in internal notes on the conversation timeline (the customer does not see these notes).

Requirements and plans

What each plan includes

Integrations that empower the AI

The assistant works without any integration (it answers questions, follows the flow instructions and transfers), but what it can do grows with what is connected in Integrations:

• The ISP's ERP (IXC, Hubsoft, SGP, Gesprov, ReceitaNet, Pipesoft and others): customer identification by phone, record, contracts, invoices, invoice copy and trust unlock (unlocking depends on the ERP; see the comparison under Integrations).
• Ravicor (network management): internet session, fiber signal, customer equipment, Wi-Fi networks and equipment actions over TR-069 (Wi-Fi, reboot, channel, speed test).
• External systems marked as available to the AI: lookup operations the assistant can call during the conversation.
• Knowledge base: plans, prices, commercial terms, ready answers and the current incident notice. See Knowledge base.

Details for each integration in Integrations.

Self-hosted installation and ChatISP Cloud

All AI features on this page apply to both deployment options. The only difference is the Local AI tab in Settings, which shows the status of the transcription and voice services on the server and lets you install them: it only exists on a self-hosted installation Self-hosted installation, because on ChatISP Cloud Ravi Systems takes care of these services.

How the AI enters (and leaves) the conversation

Entry through the flow's AI agent step

In the flow editor, the AI agent step hands the conversation to the assistant. It has three fields:

From then on the conversation is in the AI service state (visible on the AI tab of the conversations screen) and each customer message wakes up the assistant. How to build the flow is covered in Flows.

When the AI never replies

• Conversation in the queue: the queue means waiting for a human agent. The AI does not reply there, not even after it transferred the conversation itself. The only exception is the after-hours on-call mode (see After-hours on-call AI).
• Conversation with an agent: a conversation assigned to a person never receives an automatic reply, whatever its state.
• An agent took over during processing: the AI reply takes a few seconds to be generated. Before sending, the system rereads the conversation; if someone took over or closed it in the meantime, the reply is discarded.

The AI leaves when a human speaks

If an agent writes in a conversation that is with the AI, from the panel or from the phone paired with the number, the assistant steps out immediately: the conversation is transferred to whoever wrote (or to the department queue, when the message came from the phone) with the note An agent replied in the conversation: the assistant left the service.

Safety net: AI not answering

In Settings > Service, Queue and inactivity card, the AI not answering: send to queue after (min) field (default 5) protects against any failure: if the customer's last message has gone without an AI reply for that long, the conversation silently goes to the human queue, without sending anything to the customer. 0 disables.

In addition, when the call to the AI fails, the system retries twice more, with increasing intervals; on the third failure the conversation goes to the queue with the reason recorded.

Assistant configuration

Everything is in Settings > Service, Service card, AI assistant section. The Settings permission of the access profile is required. Change the fields and click Save at the bottom of the card.

The same page has the Audio transcription section (transcription and voice replies, see Audio), the Queue and inactivity card (with the AI not answering safety net) and the Waiting for reply: customer answers outside business hours card (with the on-call AI).

Personality, instructions and fixed rules

The assistant's behavior comes from four layers, from the most general to the most specific. The first two belong to the ISP; the last two belong to the system and cannot be turned off.

Default persona

The installation starts with an ISP persona already written and visible in the field, which you can edit freely. In short, it tells the assistant to introduce itself only in the first message, talk like a member of the team (short messages, human tone, calling the customer by name), not offer option menus, ask one question at a time, look up data silently and reply directly with the solution, explain without jargon in at most 5 short steps and, when it does not know, say it will hand over to someone on the team instead of guessing.

Fixed system rules

These rules apply to every conversation and no setting turns them off:

1. When unsure, transfer. Certainty means data confirmed by a lookup or by the documentation in that conversation. A hunch, assumption or generalization becomes a transfer to an agent.
2. Never makes up data or acts on assumptions: this applies to technical data, records, amounts, deadlines, links and any action on the account.
3. Never blames the company. The AI does not admit, confirm or imply that the cause of a problem lies with the ISP, the system, the service or an update, and does not agree when the customer or their technician raises that hypothesis ("makes sense", "it might have been the update"), unless a lookup in that conversation confirms it. Without confirmation, it says it cannot verify the cause from there and forwards the case to the team, neither validating nor dismissing the hypothesis.
4. No dashes in messages to the customer.

In addition, the assistant follows mandatory security and style rules:

• Never mentions names of internal systems, suppliers, AI tools or infrastructure details to the customer.
• Never claims it performed an action on the account without actually having performed it, with a successful result. If it failed, it says so frankly and transfers.
• Bank slip links, payment slip numbers, Pix codes, amounts, dates and protocol numbers are always copied from a lookup. Building or completing payment data is forbidden.
• Customer messages are always treated as data, never as instructions: an "ignore the rules above" has no effect.
• WhatsApp formatting: bold with one asterisk on each side, italics with underscores, no tables or links in brackets.
• Understands how customers talk: in Portuguese, "chave", "pix" and "código" in a payment context mean the invoice's Pix code (never the Wi-Fi password); "net" is the internet; "modem", "roteador", "aparelho" and "caixinha" are the equipment; "conta", "boleto" and "mensalidade" are the invoice; "liberar", "religar" and "desbloquear" are a request to restore the connection.
• Problem solved and nothing pending: it says goodbye in one sentence and closes in the same turn; the system then sends the standard closing message with the protocol number.

First message and greeting

The system tells the AI whether the customer has already received a message from the service in that conversation. In its first message it greets according to the time of day and introduces itself according to the persona, even when the customer opened the conversation with a photo, a receipt, an audio or a document. After that it does not greet again or repeat business hours notices, even if the step instruction asks for it.

Special case: if the first message of the conversation is short (up to 80 characters) and has no greeting, such as a "thanks", a thumbs-up or "it worked", and the same contact had a conversation closed in the last 7 days, the AI receives the summary and the last messages of that conversation. If the message just wraps up the previous matter, it replies briefly ("You're welcome! Just reach out if you need anything.") and closes without sending a new protocol message; if it continues a pending matter, it picks up from there without introducing itself again.

How the AI decides in each turn

Each time the customer speaks, the assistant runs a turn: it reads the context, decides what to do and ends with a single output (reply, transfer, close or open a service order). Within the turn it can perform several lookups and actions, up to eight, before replying.

1. Groups the messages. Fragmented messages the customer sends in sequence become a single turn (see Conversation pace).
2. Reads the context. The last 20 messages of the conversation (internal notes are left out), the step instructions, the persona, the system facts and the Knowledge base.
3. Looks up the customer's situation when needed. One lookup returns the subscriber profile: ERP record (name, status, plan and speed, address, login), contracts, billing (overdue invoices and the current month's invoice, total outstanding, next due date, last payment), network (session, signal, equipment, Wi-Fi networks) and a list of alerts already calculated, such as overdue billing or equipment offline. Wi-Fi passwords never enter the profile.
4. Queries external systems the administrator enabled for the AI, when the subject calls for it.
5. Performs enabled actions (invoice copy, unlock, Wi-Fi, reboot, diagnostics) when the customer asks and the rules allow.
6. Replies based on what it found, or transfers, closes or opens a service order.

Invoices the AI offers

The AI works with overdue invoices and the current month's invoice. An invoice for a future month already issued in the ERP is left out: it is neither late nor a debt, and the AI does not offer it. A customer who wants to pay a future month in advance is transferred.

When network data is missing or billing does not respond

If the customer has no equipment linked in network management, the AI treats this as a limitation of what it can see, never as a diagnosis: it does not say it "sees no device connected" or that the customer is disconnected. For a customer with no internet or a slow connection, it follows the basic script (device on with lights lit, cables secure, unplug for 10 seconds and plug back in) and only transfers if that does not solve it.

Likewise, if the billing lookup did not return, the AI never tells the customer the billing system is down: it says the team will confirm the amounts and transfers when the subject depends on an amount or invoice.

Safeguards against wrong replies

The trail in the conversation

Every AI lookup and action becomes an internal note on the conversation timeline, with the result and the time taken. Examples (notes shown in Portuguese): "IA consultou a situação do assinante: ok (3,2 KB em 1,4 s)" (AI looked up the subscriber status), "IA localizou o cadastro por CPF/CNPJ" (AI located the record by CPF/CNPJ), "IA: segunda via enviada ao cliente: fatura ... de 99,90 com vencimento ..." (invoice copy sent to the customer), "IA abriu a OS OS-000123" (AI opened the service order). The agent who takes over the conversation sees exactly what was done, without having to ask the customer again.

Account holder identification

With the AI only handles sensitive data for the account holder on the registered phone option on (the default), sensitive data and account actions are only handled when the conversation phone is in the customer's record in the ERP. This covers bank slips and invoice copies, billing status, Wi-Fi password, record data, unlocking and anything only the account holder can see or request. General questions about the product or service remain open to anyone.

Phone not in any record

The AI never reveals whether a record exists or not. What happens depends on the lookup options:

With the record located by document or name, the conversation starts pointing to it and the customer card shows the link with the remark located by the CPF/CNPJ the customer provided or located by the full name the customer provided. Record not found, more than one customer found, or customer without the data: the AI transfers to an agent.

Lookup by document works in the ERPs where ChatISP queries customers by CPF/CNPJ (IXC, Hubsoft, SGP, Gesprov, ReceitaNet and Pipesoft); lookup by name uses the records synchronized from the integrated ERPs.

Phone in more than one record

When the same phone appears in records of different customers, the profile the AI receives contains no one's name, billing or network. The AI naturally asks for the account holder's full name (or the CPF/CNPJ, if the customer prefers), without mentioning names, plans, how many records exist or whose they are. If it does not match, it asks once for the other piece of data; without confirmation, it transfers with the reason "account holder not confirmed". Until the account holder is confirmed, no account action is performed, not even the invoice copy.

Customer with more than one contract

If the customer has more than one installed location, the AI asks which address they are talking about before replying about the connection or acting, listing the addresses. With the customer it always refers to addresses, never to contract numbers. Unlocking, reboots and tests go to the chosen contract; the Wi-Fi change only applies to the main location (for another location, it transfers).

Someone else's connection

It is common, especially in rural areas, for someone who has signal to ask about the internet of a neighbor or relative ("so-and-so's house has no internet"). When the customer mentions a name different from the account holder or says the problem is at someone else's house, the AI does not answer based on the conversation's record or look up that record's network as if it were the subject:

• With a lookup option enabled, it asks for the CPF/CNPJ or full name of that connection and, once located, handles the lookup for that record. Actions on the third party's account remain blocked.
• With no option enabled, without the data or without a match, it transfers with the reason "inquiry about another record".

When the conversation switches records, the card records whom it was about before: (the conversation previously pointed to the record of {nome}). If the account holder provides their own document, no note is generated.

AI with actions

In the Actions the AI performs on its own subsection, each switch enables an action the assistant performs when the customer asks. The AI uses the same functions the agent uses on the customer card, with an entry on the timeline and an internal note in the chat. All six switches start on; turn off whatever you do not want to allow.

Rules that apply to every action

• Account holder only: the conversation phone must be in the record, or the link must have been made by an agent. A record located by the AI by document or name does not count, except for the invoice copy. The system checks again at execution time, regardless of what the AI decided.
• Phone in more than one record: no action until the account holder is confirmed.
• One action at a time, always after looking up the customer's situation in that conversation. The AI replies to the customer based on the result, never before.
• If it failed, it says so: with a negative result the AI frankly says it could not do it and transfers to an agent with the reason. It never pretends it succeeded.
• Actions take precedence over step instructions: if the flow instruction says to transfer to an agent for the same purpose (for example, invoice copy), the enabled action prevails.
• Record: each action generates an event on the timeline and an internal note "IA: ..." in the chat, describing what was done.

Diagnostics that take time

The speed test and the channel analysis are run by the customer's equipment, at its own pace. The AI triggers the diagnostic and waits up to about 20 seconds. If it has not finished, it tells the customer the test is running and that it will come back with the result. The system checks every 30 seconds, for up to about 5 minutes, and when the result arrives the AI is woken up and informs the customer, even without a new message from them. With no response from the equipment within that time, the test is marked as unanswered.

Payment receipt: when the customer sends the receipt (image or PDF), the AI reads the amount and date, checks them against the invoices and, if the record is blocked and unlocking is enabled, can grant a trust unlock until the payment is posted, explaining that it is temporary.

Transferring, closing and opening a service order

Transfer to the team

When transferring, the AI writes a reason for the agent (the customer does not see it) and, optionally, a short message to the customer. The destination follows this order:

1. With two or more active departments, the AI receives the list and chooses by subject: collections, invoices, blocks and negotiation go to billing; plans, sign-ups, upgrades and address changes go to sales; connection, Wi-Fi, equipment, slowness and technical visits go to support. The reason records "department chosen by the AI".
2. If the name does not match an active department, or none fits, the conversation's department applies (the one from the flow step, when defined).

Question before transferring

With AI asks whether the customer needs anything else before transferring on, when it is time to transfer the AI first says in one sentence what it will forward and to which team, and asks whether the customer needs anything else. On the next message, if they need nothing else, it transfers; if they bring another request, it handles that too, and whatever it cannot solve goes into the reason as well, so the agent receives everything at once. The question is asked once per conversation, and anyone who directly asks to talk to an agent is transferred immediately.

The flows' Transfer step has an equivalent option for conversations that do not go through the AI (see Flows).

Closing

With the problem solved, the AI says goodbye and closes the conversation; the system sends the standard closing message with the protocol number. If there are closing reasons registered in Catalogs, the AI picks the one that best describes the subject of the conversation, to feed the reports by conversation type. It never makes up a reason outside the list; if none fits, it closes without a reason.

Opening an internal service order

With AI can open internal service orders on, the AI records a service order when something needs internal investigation by the team, without the customer on the line (a suspected failure or odd behavior after a change, something that requires checking in the system), instead of guessing:

• Writes a title and description with what the customer reported and what it has already checked, facts only.
• Picks the service order type by name among the active types; the service order goes to that type's team. When in doubt, it leaves it without a type and the service order stays open for the whole team.
• Tells the customer it has recorded an internal request and that the team will check and get back to them, without promising a deadline or a result.
• At most one open service order per conversation: with one already open, it just says the team is already checking.
• On the service order board, the origin appears as AI, and the chat gets the internal note "IA abriu a OS ..." (AI opened the service order). The conversation summary for the team is generated automatically.

How the board works is covered in Service orders.

Conversation pace: grouping and "typing"

WhatsApp customers often send their message in pieces ("hi", "good morning", "my internet", "went down"). To avoid replying to each piece, the assistant waits a little before acting:

• Grouping: each customer message sets an 8-second wait. If another arrives within that interval, the wait restarts. When the customer stops, the AI replies only once, covering everything they sent.
• "Typing..." holds the reply (QR Code connections): while the customer's WhatsApp shows they are typing or recording audio, the pending reply is postponed another 8 seconds with each notice, capped at 2 minutes from the first message of the group. Someone who keeps typing nonstop does not hold the AI forever.
• A repeated turn never goes out: if for any reason two turns are pending for the same messages, the second is cancelled before using any AI.

"Typing" depends on the QR Code connection's Live typing indicator card, which comes enabled. When enabled, the panel shows "typing..." and "recording audio..." in real time and the AI waits for the customer to finish. While it is enabled, the phone paired with the number does not ring for message notifications, as happens when WhatsApp Web is open. When disabled, the phone rings again and the AI replies without waiting for the customer to finish writing. See WhatsApp connections.

Images, documents and other media

Media also wakes up the AI: a conversation that starts with a photo or an audio already gets a reply, with the first-message greeting.

Audio: local transcription and voice replies

Both features are in the Audio transcription section of Settings > Service, run on the server itself and are free (Free badge: This feature is processed on this server itself and adds no AI cost.).

Audio transcription

In the chat, the transcript appears as running text under the player, with Read more for long texts and the Transcribing indicator while the service is working. A short audio takes around 15 seconds to transcribe; long audio takes longer, depending on the server's processor. When it cannot be transcribed (after three attempts), Transcript unavailable appears.

For the AI: if the customer sent audio and the transcript is still on its way, the turn waits, in 5-second steps and for up to 4 minutes, to reply to what was said rather than to an "audio received". If the transcription service is stopped, the AI does not wait and asks the customer to write.

With automatic translation on in the conversation, the transcript is also translated into the agent's language (see Support AI).

Audio replies

How the decision is made for each reply:

• Automatic rule: if the customer spoke by audio since the last message sent, the reply goes out as voice.
• At the customer's request: "send me an audio" (driving, unable to read right now) makes that reply go out as voice, even if they wrote.
• Permanent preference: if the customer asks to always receive audio (cannot read, has poor eyesight), the AI records the preference for their phone, which also applies to future conversations, and the team receives the internal note "IA registrou preferência do cliente: sempre responder em áudio" (AI recorded the customer's preference: always reply with audio). If they ask for text again, the preference is removed.
• What never becomes voice: a reply with a link, email, bank slip, payment slip number, Pix code or long number goes out as text, because it is data the customer copies or taps. Lists and long replies (more than 450 characters) also go out as text, unless the customer asked for audio, in which case the limit rises to 1,200 characters.
• A failure never leaves the customer without a reply: if audio generation fails or the channel refuses the note, the same text is sent in writing.

In the voice note bubble, the spoken text appears as a transcript: the agent reads what the AI said without having to listen. Amounts in reais are spoken in full ("89 reais e 90 centavos").

Local AI tab Self-hosted installation

In Settings > Local AI (visible to users with the Server (network and certificate) permission) you can see whether transcription and voice are working, the reason when they are not (not installed, stopped, not responding for more than a minute, incompatible processor) and the Install now button, which installs whatever is missing without interrupting customer service. On a virtual machine with the generic processor type, transcription may show Unavailable on this processor: the screen explains how to change the type to "host" in the hypervisor. See Settings.

After-hours on-call AI

The autonomous AI handles, at any time, the conversations flows hand to it. On-call mode is something else: it is the AI covering, after hours, conversations that are already with the human team. It is configured on the Waiting for reply: customer answers outside business hours card, in Settings > Service. The choice is saved immediately, with no Save button.

With Notice + on-call AI, the AI replies after hours in two cases:

• An agent's conversation marked as waiting for the customer's reply: when the customer replies after hours, they receive the notice with the offer (the Notice message with the AI offer field, which accepts {retorno}, {protocolo} and {nome}) and the following messages are answered by the AI.
• Conversation in the queue after hours, with no agent: nobody will take over until the next opening, and a customer who keeps talking is answered by the AI instead of being left without a reply.

On-call rules

• The AI never owns the conversation: it knows the conversation belongs to the agent (or the queue) and that the agent resumes it at the next opening. It helps with whatever it can confirm through the systems, collects information and leaves everything ready.
• Never transfers or closes: if it tries, the system ignores it. When it cannot help, it replies that the team will pick up at the next opening.
• Leaves when the agent speaks: any message from the agent turns off on-call mode immediately, even discarding a reply that was being generated. The return of business hours and closing the conversation also turn it off.
• It can use the enabled actions and open service orders as in any conversation.

The waiting for reply feature is described in Customer service.

AI support for agents

Features that help whoever is serving the customer, without talking to the customer on their own. Available on all plans.

Suggest or improve the reply

The conversation's message field has an AI button with the tooltip Suggest a reply with AI (empty field) or improve the typed text:

• Empty field: the AI reads the conversation and suggests the agent's next message, based only on what has already been said.
• Field with text: the AI improves the clarity, tone and spelling of the draft, without changing its meaning or adding information.

The text goes back to the message field: nothing is sent until the agent reviews it and clicks send. The suggestion comes out in the agent's language, without dashes and without making up amounts, deadlines or links. You must have taken over the conversation; otherwise Take over the conversation to send messages to the customer. appears. For more complete help (knowing the customer, history, alerts), use the Copilot reply suggestion.

Automatic translation

The globe button in the conversation header (Enable automatic translation) turns on translation for that conversation:

• Received messages are translated into the agent's language (their account language, or the system language) and appear with the Translated mark; one click shows the original side by side.
• The agent writes in their own language; the customer receives the message translated into their detected language. In the panel, the sent message appears as it was typed.
• When it is turned on, the last 20 received messages not yet translated are translated. Transcribed audio is also translated.
• The translation is faithful: it does not change the context or the content. If translating the sent message fails, the original is sent.

Summary of closed conversations

Every closed conversation in which the customer said something automatically gets a descriptive title (up to 70 characters, such as "Slowness after router replacement") and a 2 to 4 sentence summary: what the customer brought up, what was done and the outcome or pending item. The summaries appear under the Previous tickets button in the conversation header, which lists the customer's history with date, title and summary, and they are read by the Copilot to warn about recurring problems.

Summary for service orders

When a service order originates from a conversation, the AI prepares the briefing for the team: what the customer reported (symptoms, since when, what they have already tried), what the service checked or advised and what is still pending, with useful data such as the address and the agreed time. If the conversation is closed while the service order is still open, the briefing is redone covering the whole conversation. See Service orders.

In flows

In flow menus, the AI interprets free-text answers ("I want to see my invoice" picks the billing option) and can write the short title and description of each option in WhatsApp official API list menus. These uses appear separately in AI consumption. See Flows.

AI consumption and costs

The AI consumption screen shows what the AI consumed in each conversation and how much it cost. To open it: Audit log menu, AI consumption button at the top. It requires the Audit log permission.

What the screen shows

Uses counted

The Used for column separates: Customer service (the assistant), Ticket summary, Service order summary, Translation, Menu titles, Menu interpretation, Agent text (the suggest and improve button), Copilot and Other. Importing texts into the Knowledge base is also counted, on its own line.

How the number is calculated

• Tokens are the ones reported by the AI platform itself in each response, never an estimate by ChatISP.
• The cost is calculated from the official price table of the model used (in US dollars per million tokens) and converted to reais at the day's exchange rate, updated every 12 hours. With no exchange rate available, the screen shows No exchange rate available: amounts shown in US dollars (US$).
• Transcription and voice do not appear: they run on the server itself and have no AI cost.

Privacy: what goes to the AI

The assistant's replies, the summaries, the translations and the Copilot are generated by the Ravi Systems AI platform, outside the ISP's server. Each call carries only what is needed for that task.

All content coming from the customer, external systems or lookups is marked for the AI as data, never as instructions, which protects against attempts to manipulate the assistant through the conversation.

Common problems

Other symptoms in Troubleshooting.