AI agents
Nine agents behind one gateway. They prepare texts and conclusions; nothing leaves without a human's approval. The analytical models — RFM, stages, engagement, ICP fit, attribution — are not generative AI: they compute numbers on your data inside the installation, without a language model.
| Agent | What it does | Tasks | Where in the product |
|---|---|---|---|
| Hypothesis generator | Proposes working hypotheses for a segment: whom, with which offer, through which channels, and how to tell it worked. | Formulate hypotheses for the campaign goal; pick channels only from the available ones; set success metrics | Segments → Hypotheses |
| Company brief | A summary for the manager: what is happening at the client, who decides, the next action, the risks. | Highlight signals; propose next steps; draft an email; name the risks | Account card → Brief |
| Company dossier | A note from CRM data and the company's website: what it does, what it sells, history, how to open the conversation. | Describe the company; list the products named on the site; suggest conversation topics | Account card → Dossier |
| Campaign emails | Several email variants for a goal, tone and segment — for mailings and playbooks. | Subject and body variants of a given length; an explanation of the tone | Playbooks and campaigns |
| Personal email to a decision maker | An email to one person for their role and interests, not a mailing with a name inserted. | Address the role's tasks; account for the company's activity and signals | Contact card |
| Segment from text | Turns a plain-language request into segment rules. | Parse the request; build rules only from allowed fields; ask when unclear | Segments → From text |
| Chart explanation | One sentence on what the chart shows, hypotheses and an action for the week. | Describe the trend; up to three hypotheses; state confidence; correlations only, never causes | Analytics |
| External signal filter | Decides whether a publication about a company — news, a tender, a vacancy — is worth showing to the manager. | Drop the irrelevant; give the reason in one sentence; when in doubt, show | External signals |
| Company initiatives | From the signals, works out the client's initiatives and links them to the supplier's products. | Identify initiatives; link them to products from a closed list; each with references to signals | Account card → Signals |
How they work
No personal data reaches the model
The gateway stops a request that contains an email or a phone; names are replaced by pseudonyms and substituted only when an email is sent.
A schema-checked answer or a refusal
Every answer is validated against a schema and a content validator; a bad one gets up to three attempts and a fallback provider; an invented answer is never shown.
Conclusions rest on facts
The brief and the initiatives cite facts and signals from the catalogue; a reference to a non-existent fact invalidates the answer.
AI does nothing on its own
An email goes out only after a human approves it; rights to AI actions depend on the user's role.
Budget and journal
The cost is reserved before the call and the organisation's limit is checked; every call records provider, model, template version, tokens, cost and the facts on input.
Quality is tested, not assumed
Reference cases, comparison of template versions and expert review of answers.
Models
- In the cloud service — YandexGPT and GigaChat.
- In a customer installation the administrator connects a model on the AI providers screen: a model in your network or a cloud one, at your discretion; an address outside the perimeter is enabled only with the administrator's explicit acknowledgement, logged with who and when.
- Without a connected model all nine agents plainly answer that AI is not connected; the rest of the product works.
- For production load we recommend models of 7B and up: small ones hold the answer format worse, and the gateway then honestly refuses.