AI System Card

Transparency about the AI system · Last updated: 9 June 2026

Contact addresses below are placeholders pending confirmation — owner to verify/create these inboxes.

Janiva uses AI to run your business back-office on WhatsApp. This card explains what the AI does, which providers it relies on, what data is sent to them, where a human stays in control, and the limits you should keep in mind.

In short: Janiva is an assistant. It drafts and proposes; you review and approve. AI output can be wrong and should always be checked before you rely on it — especially for financial, tax, and compliance records.

1. Purpose

Janiva interprets instructions you give in plain language over WhatsApp and turns them into back-office actions: posting ledger entries, creating GST invoices and Tally exports, managing virtual numbers, drafting and routing inbox replies, composing and scheduling social posts, and delegating and chasing tasks. The AI is a conversational and drafting layer over these features; it is not a system that makes autonomous decisions about people (it does not perform credit scoring, hiring, biometric identification, or similar high-stakes automated decision-making).

2. Model providers

Janiva does not host AI models itself. It sends requests to third-party large-language-model (LLM) providers and returns their output to you. The providers currently used are:

Google (Gemini) — primary models for most assistant tasks.
OpenRouter — routing to additional models for some tasks or as a fallback.
Anthropic (Claude) — used for the assistant on certain tasks.

These providers process requests on our behalf to return a result. Under their standard API terms they do not use data sent through the API to train their models. We do not use your business or customer data to train generalized AI models.

3. What data is sent to the AI

Only the data needed to fulfil the specific request is sent to the AI provider for that request.

Feature	Sent to the model	Not sent
Assistant chat	Your instruction and the relevant context for the task (e.g. the invoice or ledger details being discussed, recent turns of the conversation).	Passwords, access tokens, and unrelated records from other parts of your workspace.
Invoice / accounting drafting	The line items, amounts, and party details needed to draft the document.	Unrelated ledgers, other companies' data, credentials.
Inbox reply suggestions	Recent messages in the active conversation.	Other conversations and contacts' threads.
Social post drafting	The brief or prompt you provide for the post.	Your accounting and customer records.
Task delegation	The task description and the contact's name needed to draft the follow-up message.	Unrelated business data and credentials.

Authentication secrets (passwords, access tokens, integration credentials) are never sent to any AI model. See our Privacy Policy for how this data is handled overall.

4. Human oversight and approval

Janiva keeps a person in the loop at the points that matter:

Social posts have a review queue. AI-drafted posts are queued for review and approval before they are published to Instagram or Facebook.
Sensitive actions require confirmation. Actions such as creating or finalising invoices and posting accounting entries are surfaced for you to confirm rather than executed silently.
Drafts, not auto-sends. Inbox reply suggestions are drafts you can edit and choose to send.
You stay in control of scope. You decide which companies, numbers, and channels Janiva can act on.

5. Known limits and failure modes

AI output can be wrong. Language models can confidently produce incorrect figures, facts, or phrasing. Always review invoices, ledger entries, reports, and messages before relying on or sending them. AI output is not professional accounting, tax, or legal advice.
Not a substitute for your judgment on compliance. You remain responsible for the accuracy and legality of the financial and tax records you approve.
Language coverage. The assistant handles English and Hindi well; quality on other languages depends on the underlying provider.
Provider dependence. If an AI provider is unavailable, the affected feature may be temporarily degraded or fall back to another model, which can change the style or quality of output.

6. Reporting and contact

If Janiva produces output that looks wrong or harmful, please tell us so we can investigate. For privacy and data questions about the AI system, contact [email protected]; for general help, contact [email protected].