Dashboard manual · AUTOMAÇÕES

AI Connections: connect each AI provider with the company's key and choose the model for each task

AI connections is for connecting each of the company's AI providers once, with its own API key, and saying which model handles each service task, with a backup and rules by lead type. The platform's AI never answers in place of the company's models, and no AI usage is charged.

When it appears: Appears in the Automations group when AI Connections are released for the account; a new account is already created with Automations turned on, except Ads. On accounts that used own AI, the connections already registered come over on their own and the old address leads to this screen.

Who can use it

Role	What you see and do
Owner	Sees everything, connects, tests, replaces the key and removes connections, chooses the model for each task and chooses what is stored from the calls.
Admin	Sees and uses the screen like the Owner, within what the account allows non-owners to change.
Agent	Does not see the item in the menu; anyone who opens the direct address goes back to the "Painel" (Dashboard).

Screens

"Conexões de IA" (AI connections)

The AI Connections screen is for connecting the company's AI providers, testing and replacing the keys, choosing the primary model and the fallback for each task and defining what happens when the models fail.

Address in the dashboard: /app/automacoes/conexoes · Who uses it: Owner, Admin

How to get there: Via the menu Automations › AI Connections. Each tab has its own address, and the Platform agents link goes straight to the Model for each task tab, already in the right group.

Header and tabs

The title, the connect button and the screen's two tabs.

Buttons and actions

"Conectar provedor" (Connect provider): Opens the drawer to connect a new provider. In the header, the button appears on the "Conexões" (Connections) tab when there is already at least one connection.
"Conexões" (Connections): Tab with the connected keys, the failure path and data security. The number next to it is how many connections there are.
"Modelo de cada tarefa" (Model for each task): Tab with the primary model and the fallback of each group of tasks. The number beside it is the number of groups.

The sentence at the top reminds you that the usage cost is charged by the provider, contracted directly by the company.
The tabs move with the keyboard arrows, and Home and End go to the first and the last.

Migration notice

Appears once, for those who had own AI, and disappears after being dismissed.

Buttons and actions

"Entendi" (Got it): Dismisses the notice in this browser; it does not come back.
"Onde foi parar cada parte da tela antiga" (Where each part of the old screen went): Opens the list of where each part of the own AI went: models and keys, routing, what happens if the models fail and security, model health, fair use and terms.

The notice says how many connections came over on their own, with the keys, the tests and the models.

Connections tab: each connection's card

One line per provider, address and key.

Buttons and actions

"Testar conexão" (Test connection): Tests the key with the provider right away and tells you whether it answered, with the response time, or was rejected.
"Trocar a chave" (Change the key): Opens the Replace the key drawer. The old key stays valid until the new one passes the test.
"Remover" (Remove): Opens the confirmation to remove the connection.

The provider's name, the nickname (if any) and the masked key: only the start and the end show.
The status badge: "Testando…" (Testing…), "Falhou no teste" (Failed the test) or "Conectada · testada ‹quando›" (Connected · tested ‹when›).
Available models: one badge per model of the connection.
Used by: the tasks and agents that use the connection, or No agent uses this connection yet.
A connection with a failure shows The key was rejected by the provider. Replace it with a new one.
The box The key is encrypted and never shown again warns that the key does not appear again.

Connections tab: If your models fail

The order a message follows until it is answered; the lead is never left without a reply.

Step 1, The task backup: always on; tries the backup model chosen on the Model for each task tab, another of the company's models, preferably from another provider.
Step 2, Your team: the conversation goes to the queue with the notice that the models are down.
There is no switch: the platform's AI does not answer in place of the company's models, and no AI usage is charged. To have a backup, connect a second provider or choose a backup model for each task.

Connections tab: Security and data

What is stored from calls to the company's models.

Fields and filters

"Só números: tokens e tempo de resposta" (Numbers only: tokens and response time): Stores only the numbers of the calls, without the text.
"O texto das chamadas, por 30 dias, para conferir respostas" (Call text, for 30 days, to review responses): Stores the text of the calls for 30 days so the team can check what the model answered.

Connect provider drawer

Registers a provider with the company's key. Tests before saving.

Fields and filters

"Provedor" (Provider)required: OpenAI, Anthropic, Google, Azure OpenAI, Amazon Bedrock or Other OpenAI-compatible.
"Apelido (opcional)" (Nickname (optional)): A name to recognize the connection, such as primary. With no nickname, the model name is used.
"Endereço do serviço" (Service address)required: Only for Azure OpenAI and Other OpenAI-compatible: the service's https address.
"Primeiro modelo" (First model)required: The model to use first, written the way the provider names it.
"Chave de API" (API key)required: The provider's key. It is encrypted and never shown again. The screen warns when the start of the key looks like it belongs to another provider, but lets you test it anyway.

Buttons and actions

"Mostrar a chave" (Show the key): Shows what was typed in the key field; it becomes Hide the key.
"Testar e salvar" (Test and save): Tests the key with the provider and only saves it if it passes.
"Cancelar" (Cancel): Closes the drawer and erases what was typed.

Field errors: Paste the API key., The key looks incomplete. Check that you copied all of it. and Say which model to use first.
If the provider rejects it, the screen shows its response and asks you to check the key and the model.

Replace the key drawer

The same drawer, with just the key field.

Fields and filters

"Chave de API" (API key)required: The new key. If it is rejected, nothing changes and the old one stays valid.

Buttons and actions

"Testar e salvar" (Test and save): Tests the new key; if it passes, it replaces the old one.

Remove the connection confirmation

Says who stops answering before removing.

Buttons and actions

"Remover a conexão ‹provedor›" (Remove the ‹provedor› connection): Removes the connection and deletes the key from the platform's vault. It stays valid in the provider's account: revoke it there if you want.
"Cancelar" (Cancel): Closes without removing.

Lists the agents that use the connection and stop answering until another is chosen, or says No agent uses this connection.
If a task still uses a model from the connection, the removal is refused and the server's sentence appears in the confirmation.

Model for each task tab: shortcuts

They fill in the selectors at once. They stay disabled for those who cannot edit.

Buttons and actions

"Usar o mesmo modelo para todo o texto" (Use the same model for all text): Copies the conversation model to the light text tasks.
"Mais econômico" (More economical): Picks in each group the model with the lowest input price reported on the connection.
"Mais qualidade" (Higher quality): Picks in each group the model with the highest input price reported on the connection.

Search is not changed by the shortcuts, because changing the search model reindexes the knowledge base.

Model for each task tab: groups by capability

Six cards; each one saves the route for all the tasks in the group.

Fields and filters

"Modelo leve" (Light model): In the Text group, light tasks: the model for short support tasks.
"Modelo principal" (Primary model): In the Text group, conversation and auditor: the model for the conversation, the check and the follow-up.
"Modelo" (Model): In the Image reading, Audio transcription, Audio reply (voice) and Knowledge base search groups: the capability's model. Shows only the compatible ones from the connections.
"Reserva, se o principal falhar" (Backup, if the main one fails): The fallback model, always from another connected provider, or No fallback. In search, the fallback is keyword search, with no model.

Buttons and actions

"Testar" (Test): Tests the chosen model with a sample of the capability and shows Responded or Failed the test, with the time. The cost of the test goes to the company's account with the provider.
"Reindexar com ‹modelo›" (Reindex with ‹modelo›): In search, after choosing another model: confirms the change and reindexes the base. Search uses the old index until it finishes.
"Manter ‹modelo›" (Keep ‹model›): In search, gives up the change and stays with the current model.
"Conectar provedor" (Connect provider): When no connection has the capability, opens the connection drawer. The card explains what happens today without it (for example, the audio goes to the team without the text).

The groups are Text, light tasks; Text, conversation and auditor; Image reading; Audio transcription; Audio reply (voice); and Knowledge base search.
Below the backup field, a hint says how to have a backup: choose a backup model or connect a second provider; without a backup, if the primary fails, the conversation goes to the team.
If the chosen model did not pass the JSON test, the card warns that it answers anyway, asking for the JSON in the text, and suggests testing the connection again.
Each card lists the group's tasks; clicking a task takes you to Platform agents.
Capability badges: Text, Vision, Audio, Voice and Embedding.
Footer: tasks that work by rules do not use a model and cost nothing.

Model for each task tab: Rules by lead type

They change the conversation model according to the lead type; they apply from top to bottom.

Fields and filters

"Quando a regra vale" (When the rule applies): The condition: ideal profile, if there is an objection or if the card has the tag.
"Nota mínima do perfil ideal" (Minimum ideal-profile score): The minimum ideal-profile score for the rule to apply.
"Objeção" (Objection): Which objection triggers the rule.
"Etiqueta" (Tag): Which card tag triggers the rule.
"Modelo da regra" (Rule's model): The model that handles the conversation when the rule matches.

Buttons and actions

"Nova regra" (New rule): Creates a rule at the end of the list, starting with ideal profile 80 or more.
"Arrastar para mudar a ordem" (Drag to reorder): Reorders the rules, with mouse or keyboard; the first one that matches decides the model.
"Remover regra" (Remove rule): Opens the Remove this rule? confirmation; conversations go back to following the table.

With no connected model, creation is locked with Connect a model to create rules.
With no rules: "Nenhuma regra. Todas as conversas seguem a tabela acima." (No rules. All conversations follow the table above.)

Screen states

Loading: Gray blocks while the connections and routes load.
Nothing here yet: No provider connected, with the explanation that the key is encrypted and the Connect provider button.
If something goes wrong: An error notice with the try again button; the screen comes back when the read succeeds.
A connection's key was rejected: The card is highlighted as an error with Failed the test and the request to replace the key.
No connection has a group's capability: The card says what to do and what happens today (for example, audio goes to the team without the text, or the reply goes as text).
Search is reindexing: A bar with the percentage; search uses the old index until it finishes.
The account is resold: There is no migration notice.

Step by step

How to connect an AI provider's key

Open Automations › AI Connections and click Connect provider.
Choose the provider; for those that require it, fill in the Service address.
If you want, give it a nickname and write the First model, the way the provider names it.
Paste the "Chave de API" (API key) and click Test and save.
When it passes, the connection appears on the "Conexões" (Connections) tab as Connected.

How to replace a connection's key without stopping service

On the Connections tab, click Replace the key on the connection's card.
Paste the new key and click Test and save.
The old key stays valid until the new one passes the test; if it is rejected, nothing changes.

How to choose the model for each task and the fallback

Open the Model for each task tab.
In each group, choose the model in the selector; it saves right away.
In Fallback, if the primary fails, choose a model from another connected provider.
Click "Testar" (Test) to check the response.
For a quick change, use Most economical, Highest quality or Use the same model for all text.

How to change the search model

On the Model for each task tab, open the Knowledge base search group.
Choose the new model; the screen shows how many passages will be reindexed and the estimated time.
Click Reindex with the chosen model, or Keep to give up.
Follow the bar; search uses the old index until it finishes.

How to remove a connection

First, in Model for each task, change the models that still use the connection.
On the "Conexões" (Connections) tab, click "Remover" (Remove) on the card.
Check which agents stop answering and confirm in Remove the connection.
Revoke the key in the provider's account if you want it to stop being valid for good.

How to change the conversation model for a lead type

On the Model for each task tab, scroll down to Rules by lead type and click "Nova regra" (New rule).
Choose the condition, the detail (score, objection or tag) and the "Modelo da regra" (Rule's model).
Drag the rules into the order you want: the first one that matches decides.

Frequently asked questions

Who charges for the use of the connected models?

The provider, directly on the account the company has with it. With an active connection, AI usage by these models is not charged by the platform; the platform's AI does not answer in place of the company's models, so no AI usage is charged.

Can the key be seen after it is saved?

No. It is encrypted and the screen shows only the start and the end. To see the whole key, check the account with the provider.

The new key was rejected. Does service stop?

No. The old key stays valid until the new one passes the test.

What happens if the model goes offline?

The message tries the task backup (another of the company's models) and, without a backup, goes to the team with the notice that the models are down. The platform's AI does not answer in their place. To have a backup, connect a second provider. The lead is never left without an answer.

Why does changing the search model take a while?

Because the knowledge base is reindexed with the new model. Search uses the old index until it finishes, without interrupting service.

Where did the Own AI screen go?

On accounts with AI Connections released, it became this screen: models and keys on the Connections tab, routing on the Model for each task tab, and the old address leads here.

Help center guide

"Conexões de IA" (AI connections)

Connect each AI provider once and choose which model does each task. Usage is billed by your provider, which you contract directly.

In "Automações → Conexões de IA" (Automations → AI connections) you connect your company's AI providers and choose which model does each task. The screen has two tabs: "Conexões" (Connections) and "Modelo de cada tarefa" (Model for each task). Anyone who used "IA própria" (Own AI) finds everything here: the connections already registered came over on their own, with the keys, the tests and the models, and the old address brings you to this screen.

"Conexões" (Connections)

Connect a provider: Click "Conectar provedor" (Connect provider), choose the provider (OpenAI, Anthropic, Google, Azure OpenAI, Amazon Bedrock or another OpenAI-compatible one), add a nickname if you want, name the first model and paste the key. "Testar e salvar" (Test and save) tests the key with the provider before storing it.
If the test fails: The screen shows what the provider answered and warns you when the key looks like it belongs to another provider. On Google, the key starts with AIza (with a capital i) or, on newer keys, with AQ.: both work.
Check the status: Each connection shows the start and end of the key, whether it is "Conectada" (Connected) (and when it was tested), the available models and who uses the connection. "Testar conexão" (Test connection) tests it again right away.
Change the key when you need to: In "Trocar a chave" (Change the key), paste the new one. The old key keeps working until the new one passes the test: if the new one is refused, nothing changes.

"Remover" (Remove) asks for confirmation and says which agents use the connection and stop answering. While a task uses a model from the connection, change the model in "Modelo de cada tarefa" (Model for each task) before removing it. The key leaves our vault but still works on your provider account: revoke it there if you want.

If your models fail: The message tries the task backup (another model of yours, chosen in the Model for each task tab) and, without a backup, goes to your team with the notice that the models are down. The platform's AI doesn't step in. To have a backup, connect a second provider. The lead is never left without an answer.
A support task with no answer: If only a support task has no model for that message (the answer check, intent, qualification, sentiment or data extraction), it uses the local rules and the reply your model wrote goes out as usual. The conversation only goes to your team when the reply to the lead itself has no model.
What is charged: With an active connection, AI usage by your models is not charged by the platform: messages, hot and warm leads, qualification adjustment, segment and form generation, cadences, campaigns, transcribed audio and read images. SMS and email are not charged either. The platform's AI does not answer in place of your models, so nothing for AI is charged. Meta templates stay on the prepaid balance.
Model that didn't pass the JSON check: The conversation with the lead works with it. Intent and qualification, which need the answer in a fixed format, are requested from the same model as text; if it gets the format wrong, the message follows the same path as when your models fail. For the best result, choose a model that passed the JSON check for these tasks.
"Segurança e dados" (Security and data): Choose what is stored from the calls: numbers only (tokens and response time) or the text of the calls for 30 days, to check answers.

"Modelo de cada tarefa" (Model for each task)

The platform agents' tasks are grouped by capability: "texto, tarefas leves" (text, light tasks); "texto, conversa e auditor" (text, conversation and auditor); "leitura de imagem" (image reading); "transcrição de áudio" (audio transcription); "resposta em áudio (voz)" (audio reply (voice)); and "busca na base de conhecimento" (knowledge base search). Each selector shows only the compatible models from your connections, and the backup comes from another connected provider.

Yes: "Usar o mesmo modelo para todo o texto" (Use the same model for all text), "Mais econômico" (Cheaper) and "Mais qualidade" (Higher quality) fill in the selectors at once.
Yes: Changing the "busca" (search) model reindexes your base: the screen tells you the size and the time, and search keeps using the old index until it finishes, with live progress.
Yes: When no connection has the capability, the group says what happens today and offers "Conectar provedor" (Connect provider). With no backup chosen, the hint under the field suggests connecting a second provider or choosing a backup model. A model that didn't pass the JSON test still counts: it answers by asking for the JSON in the text, and the group suggests testing the connection again.
Yes: Each of your agents can change only the main and the light model in the editor; image, audio and voice follow this tab.

In the dashboard: "Abrir Conexões de IA" (Open AI connections)