92conversations · 30 days
A revenue operator agent we architected and installed for RIVET, live on their own site: voice and chat, Arabic and English, instrumented, connected to the booking system, and scored on every conversation. Read what it does, what it measures, and the number we refuse to quote yet.
92
conversations in 30 days — real inbound traffic, no human present
4.9h
of revenue conversation handled across the window
2.6s
median response time — measured, not a vendor spec sheet
100%
compliance — no invented facts on any scored conversation
92%
experience quality — give-before-you-ask, scored automatically
$0.21
model cost per conversation — the commercial unit
The brief
RIVET needed a first-contact seat on their own site — one that could sell to the buy-ready, teach the curious, qualify and hand off a team buyer, route a partner and support an existing member, in Arabic or English, at any hour. We architected it, built it, instrumented it, and handed it over under our own governance law.
What it is
A single agent with multiple intent modes rather than a menu of bots. Within two turns it establishes whether it is talking to a member, a subscriber or a new visitor, then switches into the matching play.
Where it runs
Live on RIVET’s public site — one launcher, voice and chat, Arabic and English with dialect. Verifiable in ten seconds by anyone: open the page and press call. Sixty-one-message conversations in its history are not deflection; that is a discovery conversation running to depth.
The law it obeys
Give before you ask. One question per turn, never two in a row. Contact captured as a gift, never a toll. Every price and proof point answered from a governed knowledge base — it is forbidden from answering these from memory. And one honest record written at the end that a human can resume from.
How it is wired
Three tools, deliberately no more: a once-only lead record written at the end of every conversation, a calendar read before any time is offered, and a booking write. Each additional tool is another way for a model to leave the spine.
How it is judged
Automated evaluation criteria and twenty-five extracted data points per conversation — topic clustering, sentiment, and the buyer’s own stated complication. The transcript record is already producing a demand map for RIVET without anyone reading a transcript. An installation is scored; a demo is not.
We do not report a conversion rate for RIVA yet.
Part of her scoring rig is under repair, and an aggregate built on broken criteria is a number you cannot defend. When it publishes, it publishes with the before as well as the after — that is the same discipline we bring to your system. Every competitor quotes numbers they cannot show.
Built by Taken Cake for RIVET and running on their site. Deployed on AI-SCORE™, the communication model of the RIVET INFINITE GTM OS, under licence. Metrics read from the production dashboard, 30 days to 31 July 2026, and directional at this sample size. What never leaves the room: the prompt, the law set, the tool schemas and the knowledge base — the machine stays sealed; the results travel.
