Catch

What Catch is building

Catch is an AI assistant for executive admin, which is a category most of the AI market has skipped in favour of coding, research and content generation. It syncs with an executive's calendar, email, Slack and WhatsApp, builds a profile of who they work with, how they prefer to meet and how they travel, and then does the work: scheduling and conflict resolution, inbox triage, briefing notes, follow-ups, travel and hotel booking, form filling, and outbound phone calls. The lightest-touch entry point is that you can CC an address on a scheduling thread and it takes over, no account required. The explicit product bet is against workflow builders. Sabato's framing is that executives do not want to build their own AI agents, and in the admin space they do not want to approve every step either, they want it done right in a way they can trust. So Catch is designed to observe rather than be configured, and to escalate only when a decision genuinely needs human judgment. The example the company gives is small and telling: if you book a flight without a hotel, Catch checks rates at the hotel you normally use and asks whether it should book. Under that is a real architecture argument. Ramon says Catch anonymises data and uses it to post-train open-source models through supervised learning and reinforcement learning, rather than routing everything to a frontier API. The runtime uses what the company calls splinter agents: each one restricted to a single task and given access only to the information required to complete it, which is a permissions design as much as a performance one. That matters for a product that is holding an executive's entire inbox. Traction so far: more than 12,000 meetings scheduled, more than 20,000 tasks delegated, and more than 200,000 emails handled.

Why this matters

The AI assistant category is crowded with demos and thin on products people keep using, and the reason is almost always the same. Most tools produce a draft and hand the work back, which means the executive still carries the task. Catch is one of the few attempts to close the loop and actually take the item off the list, including over voice, and the throughput numbers it has published are the kind that only accumulate if people stop supervising. The strategic point is the data. Every scheduling thread, travel booking and rescheduled call is a labelled trajectory for a task nobody has good training data for, because this work has historically happened inside human executive assistants' heads. Post-training open-source models on anonymised versions of that flow is a defensible position if the volume compounds, and it is why the round is funding model training as a line item rather than just headcount. The risk is equally legible: this product only works at a level of access most companies are nervous about granting, so trust and security posture are the go-to-market, not a footnote on it. That is presumably why they built splinter agents with least-privilege scoping before they built a growth motion.

Investors: Entrée Capital (co-lead), Pitango (co-lead), Seedcamp, Factorial Capital