The Buzz-O-Meter was a fun activation for Dreamforce. Underneath it was a genuinely hard prediction problem.

Conference vocabulary moves fast. In 2023, "AI" was said eleven times as often as "agent" on Dreamforce stages. Two years later, "agent" was said nearly twice as often as "AI." Einstein, Salesforce's AI brand for a decade, lost 92% of its stage mentions in a single year.

This year's Dreamforce ran nearly 1,000 sessions over three days, each with its own speakers and no shared script. The day before it opened, we asked NEXUS to predict which words would get said most across all of them.

Its top two calls were "Agent" and "AI." Both finished exactly there. "Agent" was said 15,129 times, roughly once a minute across recorded sessions. "AI" came second at 6,088.

Of the 15 terms NEXUS ranked highest, 12 finished in the actual top 15. Across all 71 terms we tracked, its ranking matched reality with a 0.89 rank correlation, where 1.0 is a perfect match.

What we built

NEXUS worked from three sources:

• Four years of stage transcripts. 173 Dreamforce sessions from 2022 to 2025, about 890,000 words, normalized per 10,000 words so bigger years don't skew the counts.

• 46,000 tech news headlines. Compared year over year to measure which terms were gaining momentum and which were fading.

• This year's agenda. Titles and abstracts for 987 sessions.

It made a prediction for every term in every session, roughly 70,000 in total, and rolled them into one ranking, locked the day before Dreamforce began.

Each source does a different job: history sets the baseline, news momentum shows what's changing, and the agenda shows what speakers actually signed up to talk about. The call that shows why that matters is Claude.

The call history couldn't make

Claude wasn't said once on a Dreamforce stage before 2025, and it finished that year at #32. A model working from history alone would have left it around there.

NEXUS picked up its momentum in the news and flagged it as a riser. It finished #9, with 941 mentions, on a term the historical data barely registered.

Beyond Dreamforce

Predicting conference buzzwords is a bit of fun. The shape of the problem shows up everywhere: years of historical records, a stream of live signals, and a decision that depends on what happens next.

Swap the transcripts for sales history and the headlines for market signals, and you're forecasting demand. Swap them for meter readings and weather, and you're forecasting energy load. Most of the data behind decisions like those lives in tables, which is the data NEXUS was built for.

If it can call what 1,000 sessions of unscripted conversation will sound like, it's worth asking what else it could see coming.