REMY is our own AI product. A customer adds their pages, PDFs, Word documents and hand-written FAQs; REMY processes what has been approved, and the widget answers visitor questions from that content only, linking the sources it used. When it cannot ground an answer, it says so and offers the customer's own "talk to a human" contact instead of guessing.
It grows in three modes. REMY Knowledge answers from public content and is self-service once the domain is verified. REMY Connect retrieves authenticated, read-only data for signed-in visitors through capabilities the customer's team imports, tests and approves one at a time. REMY Actions performs approved changes — and only after the visitor confirms the exact payload and passes step-up authentication.
The design rule is that the model interprets and never authorises. The assistant can only call capabilities a team has configured, reviewed and activated, filtered through an output allowlist before the model sees anything. A production capability goes live only when REMY's own administrators approve it; the customer's workspace cannot switch it on alone. Saying "yes" in chat is not consent. Money movement, account closure and accepting legal terms are refused outright rather than gated.
The widget runs in an isolated iframe on its own domain, so it cannot read the host page, its forms or its visitors' cookies. Rate limits and a daily question cap are enforced before any AI call, so a capped request costs nothing. Behind it sit workspaces with roles, two-factor authentication, an append-only audit log, analytics with semantic clustering of unanswered questions, and payments through Stripe, Flutterwave or Pesapal — the last two because mobile money is the usual way to pay across much of Africa and Stripe cannot process it.