Salesforce hosted the conference where three frontier lab CEOs argued about how fast AI should move.

Salesforce does not run Agentforce on the frontier. Its bots sit on earlier generations, not on Claude Fable 5.1 and not on GPT-6 Astra.

Start there, because everything else at Dreamforce this week follows from it.

The floor disagreed with the stage

Marc Benioff had Sam Altman, Dario Amodei and Jensen Huang on stage debating pace and safety. Huang’s position was the simplest of the three. Run as fast as you can.

Then CNBC went and talked to the 50,000 people who paid to attend, and got a different answer entirely.

Tim Sanders, chief innovation officer at G2: the majority of agentic outcomes are not driven by frontier capabilities. They are driven by last year’s AI.

Kevin Lee, technology chief at Nice: the models already deployed are highly performant and effective at doing the things their customers need. His company deliberately stays at least a generation back.

Jaya Rohit Vuyyuru, a VP at SummitX: the frontier models are way ahead already, and a lot of the customer base is still getting its feet wet.

And Alec Bronston, a senior Salesforce director at Spins, said the thing most people only say privately. It is already hard enough to keep up, and a slowdown would be a chance to even catch up.

For marketing, this is a budget story

The license is the cheap part of a model upgrade.

The expensive part is everything that follows. Prompts that worked last quarter get retested. Brand voice guidance gets recalibrated because the new model has different default rhythms. Internal documentation goes stale. The team gets retrained. Two or three weeks of senior attention go into a change nobody asked for.

Meanwhile, the thing that was actually broken is still broken, because it was rarely the model. An ad that does not convert is not waiting for a smarter model. It is waiting for a clearer offer. An email that reads like a template reads that way because nobody ever wrote down what the brand sounds like.

We made that case in marketing does not need its own AI model. Dreamforce is the same argument, told by the buyers instead of by us.

Steal the three-month rule

Nagarro’s policy is the most useful thing in the coverage, and it is boring, which is exactly why it works.

Three months. A model ships, the team keeps working, and somebody evaluates it when the calendar says to evaluate it.

The marketing version writes itself. One fixed window per quarter for reviewing tools and models. Outside that window, nobody swaps engines regardless of what the feed is doing. Inside it, you run the candidate against what you already have, using briefs you have already run, and you only switch on a clear win.

What that buys you is a quarter spent on the offer and the message instead of on relearning the tool. It buys brand voice enough time to stabilize. And it means you can compare Q3 to Q4 without half the variance coming from having changed models in November.

The one thing worth doing now

Waiting is not the whole answer, and I would be selling you short if I stopped there.

There is one thing worth looking at this week, and it has nothing to do with which model you run. Who reviews what goes out? If your team publishes AI-assisted work with no judgment step in the middle, the newest model on earth does not fix that, and last year’s model does not make it worse. That is where the brand risk actually lives, and we wrote it up in your AI marketing needs a verifier.

If you are putting the Q4 plan together right now, that is the order I would use. Verifier first. Quarterly window second. Frontier model when something genuinely requires it.

See how we run marketing with AI