Glasshouse Live: How We Run NinjaCat on NinjaCat

NinjaCat is an AI operating system for marketing. So in an effort to stand by and test our own product, we run the company on it, now, out in the open.

That's the premise behind Glasshouse Live, a new webinar series where we open up the Ninjacat  factory floor. This isn’t a polished case study, or a curated highlight reel. Real workflows, real results, and the parts still being built.

The first session aired recently with NinjaCat leadership and team members walking through how NinjaCat uses its own platform across every function — marketing, sales, customer success, support, product, and operations. Together they covered what AI-native actually means as an operating model, ran three live factory demos, and shared what five factories running a real company look like from the inside.

Here's what was covered.

AI-Native Is an Operating Model, Not a Tool

Most companies bolt AI on top of existing processes. Copilots sit inside tools, helping individual people move faster. That's useful — but it doesn't compound.

An AI-native company is built differently. The work is designed from the outcome backward. Agents are named, scoped, and measured. There's an explicit line between what agents decide, what they propose, and what they never touch. Context — the data and memory every agent reads — is treated as infrastructure, not an afterthought.

The simplest test: delete the AI. If the company slows down, you're AI-assisted. If it stops, you're AI-native.

NinjaCat passes the second test. That's what this session was built to show.

From Chatbots to Factories

The session traced the evolution from early AI tools to the model NinjaCat runs on today.

Chatbots answer a question. Copilots draft something. Personal agents let individuals tune their own workflows — but those improvements stay siloed. When someone figures out a better prompt, nobody else benefits. When something fails, nobody sees it.

Factories are different. A factory is a shared system built for one kind of recurring work, owned by one person who watches every run. Effort goes deep rather than wide: one builder perfects one or two workflows instead of everyone spreading thin across ten. Every run is visible, scored, and fixable. When something breaks, it gets repaired — and the repair makes the whole factory better, not just one person's session.

This is the paradigm NinjaCat operates under today, and the one built into the platform customers buy.

The Three Factory Tours

Each Glasshouse session explores a different part of the company: the agents doing the work, the decisions people own, and what the team has changed along the way. In this session, Paul, Dustin, and Mark demoed three factories live.

GTM Factory: From Signal to Signed

Marketing and Sales run one shared pipeline. Intent signals, usage data, and events flag accounts. Agents research the company, map the buying committee, build on-brand outreach, prep for calls, and draft follow-ups. The one place a person sits: pricing. Proposals and quotes require human approval before they go out.

Product Factory: From Feedback to Feature

Support tickets, customer calls, usage patterns, and sales feedback all feed one backlog. Agents cluster themes and size them by revenue impact. The product lead sets the roadmap — that's the human gate. From there, agents write specs, acceptance criteria, and prototypes, then handle release notes and documentation on the other side.

Support Factory: From Question to Resolution

Nina, NinjaCat's support agent, resolves 69% of support tickets without human intervention. What she can't close gets routed and prioritized. The support team owns the hard cases. Bugs get reproduced and filed for engineering. Every fix updates the knowledge base automatically — so the next similar question costs less to answer.

Where Humans Belong

The session was direct about this: agents do the work, people own the decisions.

Every agent action is checked against policy before it runs — blocked, held for approval, or allowed. No agent reviews its own output; a second agent checks, and a person approves at every human gate. Each agent gets only the data and permissions its job actually requires.

The ladder of autonomy NinjaCat is climbing has five rungs, from assisting to owning outcomes. Today, NinjaCat operates at "act with approval" — agents take action, a person approves each step. The goal is policy-governed autonomy, where the policy engine handles routine approvals and people see only the exceptions.

‍