AI
In AI you connect a model, teach it to talk like you and choose when it answers on its own and when it drafts. AI replies go out from the phone, from the same account. After this page the AI will be ready to handle lead replies in Chats.
The 8-step setup wizard
In AI press Start the wizard, and gGram walks you through eight steps.
| Step | What we do |
|---|---|
| 1. Provider and key | Choose what runs the model: the gGram balance (a shared key, gpt-4.1-mini), or your own OpenAI, Anthropic or OpenRouter key. DeepSeek models are available through OpenRouter. |
| 2. Request limit | For your own key choose a daily cap: 100, 300 or 1000. It counts across all accounts. Test runs count too. |
| 3. About you | Tell who you are and what you do. This becomes the facts. |
| 4. Conversation goal | What you want from the talk with a lead: a meeting, a request, a payment. |
| 5. Materials | Attach or paste texts: prices, terms, objection handling, common questions. |
| 6. Study | gGram reads the materials and shows the facts and rules it found. |
| 7. Clarifications | Answer gGram's questions: do not say, ready messages, your corrections. |
| 8. Check | Run a test of 26 typical questions and approve the result. |
Providers and models
| Provider | How to connect | Which models |
|---|---|---|
| gGram balance | Enter nothing. Usage is charged to the dashboard balance in rubles. | gpt-4.1-mini, through gGram's OpenRouter. |
| OpenAI | Paste your sk-... key. | The OpenAI catalogue. |
| Anthropic | Paste your sk-ant-... key. | The Claude catalogue. |
| OpenRouter | Paste your key. It gives access to DeepSeek, Mistral, Llama and others. | deepseek/deepseek-chat, deepseek/deepseek-reasoner and others. |
Funnels
A funnel is the script for one kind of conversation, for example cold outreach or an inbound request. Each funnel has seven blocks:
- Facts: what the model must know about you, the product and prices.
- Rules: how to answer, tone, length, required questions.
- Do not say: what must not be written (price before a request, deadline promises, competitors).
- Ready messages: templates for typical turns of the conversation.
- Your corrections: what the model should have done differently last time.
- Questions and answers: exact pairs, asked like this, answer like that.
- Situations: where to steer the conversation in rare cases.
Modes
| Mode | What it does |
|---|---|
| Off | The AI neither answers nor drafts. |
| Draft | Prepares a draft; you send it in Chats. |
| Auto-reply | Answers on its own, the reply goes out through the phone. You can set a from-to pause and working hours. |
The 26-question check
Before you turn it on, gGram runs 26 typical questions and shows the model's answer to each. Mark OK or Fix. Fixes go to the funnel's Your corrections block, and the model takes them into account next time. The check uses about 26 requests from the daily limit.
Working hours and pauses
- Working hours: the hours when the AI answers. At other times no draft is prepared and no auto-reply goes out.
- Auto-reply from-to pause: a random delay between the lead's message and the answer.
- Auto-enable in new chats: an account setting. When on, new chats are handled by the AI right away.
Hot and Needs a human marks
The AI marks a chat by itself when it sees a reason:
- Hot: the lead is ready to discuss terms, it is time to call or book a meeting.
- Needs a human: the knowledge has no answer, or the lead asked for a person.
7-day summary
In Summary you see how many chats the AI handled, how many drafts were sent as is, how many were edited and where the model got things wrong. The week's fixes can go straight into the funnel's Your corrections.