Rethinking enterprise event creation with AI
An agent that builds out a whole event, so admins don't set up every session by hand.
Problem
Learning admins at Highspot build events one piece at a time. They create an empty event, land in an overview, then set up each session separately. Same fields, again and again.
There's no view of the whole program until the end. Fine for a one-session webinar. Painful for a multi-region certification, where the same setup repeats for every session.
Approach
For a month, our design team built on Arctic AI, Highspot's new framework for AI agents. Each of us took the area where an agent could take the most work off someone's plate. I took event creation, because most of its friction is repeated setup, and an agent should take that over.
I didn't want an agent that fills in the same forms faster. Instead, admins describe the event once in chat, and the agent builds it: every session, and how they connect. Then admins finish on a canvas, dragging and opening nodes by hand, with chat open the whole time.
I owned the scope as well as the design, and I held one path from start to finish: prompt, draft, fix, publish. I prototyped it in code with Cursor instead of Figma, because an agent building an event only makes sense in motion. A simple parser stands in for the LLM.
Key decision
Once the agent drafted the event, admins needed a place to edit it. I prototyped three options. A Wizard was familiar, but it took a lot of clicks and never showed the whole event. A free-form Canvas showed everything, but it made admins zoom, pan and draw connections that events don't need. I chose a simple Node Map. Each part of the event is a node that admins can drag or open to edit. On small screens, an Outline shows the same structure as a list. Both open the same detail panel.
Agent drafts will sometimes be wrong, so the agent checks its own work before publish. If it finds an instructor double-booked across overlapping sessions, it suggests a new time and waits. The admin approves it or picks their own, then publishes.
Outcome
This didn't ship, and nothing was measured. The bet: setup goes from one form per session to one prompt and a review, so multi-region programs stop being the expensive ones to build. I'd measure time to first publish, and how many multi-session events actually get finished.
Running scope myself taught me to start with the one path that has to hold up: prompt, draft, fix, publish. Everything else waits until that works.