AI in Kiwire

Three assistants, one credential chain, one door for outside agents, and a fixed list of things none of them will do.

The three surfaces

The reporting assistant sits at Reporting → AI and answers plain-language questions about a tenant's own reports — logins, sessions, devices, dwell time, reject reasons — by calling seven read-only report tools. It refuses anything that is not about Kiwire reporting, cannot change a setting, and will not quote a number it did not get from a tool.

KAIA — Kiwire AI Assistance runs on a schedule rather than on request: every 6, 12 or 24 hours it runs 13 detectors over the tenant, reconciles what it found against what was already open, and writes a plain-language digest when something changed. An admin can also chat with it, and it can propose a change from a small catalog — but every change waits for a human to click Apply.

KAIA marketing is KAIA's marketing mode: it computes guest segments and files proposals — a campaign draft, a segment tag, a message send — for an administrator to approve. Nothing reaches a guest without a consent record on that channel, and the gate fails closed.

What they share

All three resolve which provider key pays for a request the same way: the tenant's own key if it has brought one, otherwise Kiwire's platform key, otherwise .env. Requests on the platform key count against a monthly token cap per tenant; a tenant on its own key does not. Provider, model, key and cap are covered in Keys, models and limits.

An assistant that is not Kiwire's — Claude Desktop, Cursor, an agent of the tenant's own — reaches the same tools over MCP, on three HTTP endpoints behind the tenant's existing API keys and three separate on-switches. See MCP and external agents.

What AI in Kiwire never does

It never runs free-form SQL against a tenant's tables; it never sees an unmasked username, email, phone or full MAC; it never applies a change on its own; it never touches admins, roles, secrets, credentials, payments, the licence, RADIUS, MFA or connectors; it never deletes anything; it never messages a guest without a consent record; and it never reaches another tenant. Each of those is a mechanism in code, not a line in a prompt — Guardrails goes through them one at a time, and is the page to read if you are reviewing Kiwire's AI for security or privacy.