An event project starts with reading. A brand deck, a strategy document, a technical spec, notes from a briefing call, a transcript nobody has cleaned up. Somewhere in that pile is the running order. What the speaker says, what appears on the screen behind them, what gets cut because the stage time is not there. Finding it is slow, and the work tends to fall in the week with the least room, because the show date does not move. Warburg AI is our attempt to take that first pass off a person's desk.

It reads the source material and proposes a structure. Sections, a suggested running order, which points carry a visual and which are better left spoken. It also flags claims that need a source before they go near a screen. The output is an editable outline. Each line keeps a reference back to the page it came from, so a designer can check it rather than trust it. The name comes from Aby Warburg, who pinned images to panels and moved them until the relationships between them became visible. That is nearer to how we handle source material than any idea of automatic generation.

What did not work

Generating finished slides. Early versions produced layouts that looked like a deck and read like nothing. Summarising flattens emphasis, and emphasis is most of the job in keynote content. Deriving hierarchy from document headings failed as well. The line that matters is often in a footnote, a table cell or an image caption, and the heading above it says nothing useful. We also tried to have it judge what would hold up on a wide LED wall, and dropped that. It has no sense of physical scale, or of how a sentence reads from the back of a room. So it stops at structure. It runs on our own material inside the studio, and it takes on the part of the work that was only ever reading.