Dreamforce says business leaders have enough of old AI models

in sales force CEO Marc Benioff and executives keynote at annual Dreamforce mega conference in San Francisco this week human, OpenAI and Nvidia Delving into the heated debate about AI safety, executives weigh in on whether the pace of model development is too fast.

Nvidia CEO Jensen Huang joined Benioff on stage and urged Frontier Labs to “run as fast as we can.”

But on the ground more than a decade ago at a 50,000-person event dubbed the “Super Bowl of Software,” the discussion was much different. Participants spoke of their fear of artificial intelligence and the difficulty of even taking advantage of the technology that exists today, let alone dwelling on whether the rapid advancement of AI models threatens human existence.

“It’s already hard enough to keep up,” said Alec Bronston, senior director of Salesforce at Spinz, a Chicago-based retail data company. A potential slowdown would provide “many opportunities to catch up and gain ground,” he said.

Dreamforce landed in the midst of a dramatic moment in the AI ​​boom this year, and had a different vibe than previous years. A few days before the conference began, an anthropology researcher resigned, saying on his way home that his top lab was “putting our lives on the line.” For this reason, Anthropic’s Dario Amodei and OpenAI’s Sam Altman proposed safety initiatives and pushed for a slower pace of model development.

In the months leading up to the event, Benioff had been reinforcing Salesforce’s role in the AI ​​explosion, trying to shatter the “SaaSpocalypse” narrative and convince investors, customers and employees that AI is a business enabler, not a threat. Salesforce stock is down 8% this year despite a massive stock rally in August, while other software names such as adobe and autodesk has decreased even more significantly.

During its last earnings call in August, Salesforce announced “Claudeforce,” a new way for sales reps to access critical data directly from within Anthropic’s Claude chatbot.

On the show floor, where companies in the Salesforce ecosystem pay big bucks to promote their software and services, Anthropic’s was the busiest booth Tuesday. A young AI Lab employee wore an oversized white sweater and demonstrated to passersby.

Salesforce customers and partners at the conference told CNBC that older, cheaper AI models are powerful enough for day-to-day sales and customer service operations. Many users are still figuring out their AI budget and deciding whether it’s best to use an Anthropic or OpenAI model, or a cheaper open source alternative instead.

“The frontier model is already far advanced,” said Jaya Rohit Buyur, vice president at consulting firm SummitX. “A lot of the customer base is still in the blanks. There’s still a gap in where customers are still available and where agents can start to understand what they can do.”

Meanwhile, guests interacted with the cartoonish corporate mascot as they traversed the expansive Dreamforce campus (which Salesforce calls the campground), resulting in a viral meme on social media. The mascots were a nice little distraction from the AI ​​chatter that was going on during the event.

Some attendees said their employers are adopting Salesforce’s Agentforce tool, which allows them to respond to customer service questions and sales inquiries. According to the website’s support page, Salesforce does not rely on Anthropic’s state-of-the-art Claude Fable 5.1 or OpenAI’s new GPT-6 Astra for bots built with Agentforce.

“The majority of agent success is not driven by frontier capabilities,” said Tim Saunders, chief innovation officer at software review firm G2. “They are driven by the AI ​​of yesteryear, which has less to do with agent providers and certainly less with SaaS (software as a service).”

Kevin Lee, head of technology at cloud contact center software vendor Nice, said his company doesn’t rely on high-end models like Fable for most of its workloads.

“For the most part, the models that are already on sale today, even one generation from now, are very performant and effective in doing what customers need,” Lee said. “Once you get past this point, it’s as if everything is fine.”

This is not to underestimate the power of AI across the software space. Many executives say this has been transformative for their businesses.

electronic signature software developer docusignThe Dreamforce exhibitor uses “all Big Frontier models and some Open Class models,” CEO Allan Thygesen said.

According to Docusign’s website, “Large frontier models are reserved for judgment-intensive tasks such as complex clause analysis, multi-document reasoning, and summarization, and general reasoning capabilities are worth paying a higher cost per call.”

Frontier models are typically proprietary, meaning users cannot review data sets or fine-tune the technology to suit their own purposes. In contrast, open-weight models can be downloaded and run on any infrastructure.

DocuSign’s home base for this week’s three-day event was at the Canopy by Hilton across from San Francisco’s Moscone Center. The company rented out the entire hotel, with plenty of quiet meeting space. In an interview there, Thygesen said Docusign uses a technique called model routing to send each request to the most cost-effective AI system.

This is a tool that Nice and many other companies are currently using as one of the key ways to keep costs down while leveraging AI.

Nice’s Saunders said cloud software companies also need to figure out how to respond in a new token-based economy, where people pay based on their demand for AI usage and output, rather than on a subscription basis.

Saunders said that as customers migrate to tokens, software companies need to adapt both to their use of technology and how they deliver their services. One token represents about three-quarters of a word.

“Traditionally, SaaS didn’t have variable costs to provide a service,” Sanders said. “The profit margin structure has changed.”

Saunders said switching from a typical SaaS model to an agent model can reduce gross margins, or the profit left after cost of goods sold, from more than 85% to nearly 45%.

Several companies want to access the best capabilities that AI has to offer. Databricks, a data analytics software company with technology to build AI agents, this week released GPT-6 Astra to all 3,500 of its software developers.

“Astra clearly outperforms our previous top-of-the-line models (Opus 5, Sol 5.6) for highly complex tasks,” Patrick Wendell, vice president of engineering at Databricks, writes about X, referring to Anthropic and OpenAI models, respectively.

As a systems integrator, Nagarro is in a completely different position.

Ram Reddy, head of technology for consumer industries at Nagaro, was wandering around his company’s booth at Dreamforce this week. Following Mr. Amodei and Mr. Benioff’s conversation, some who stopped by discussed regulations.

But the hottest topic is Claude Force, not the future of the frontier, Reddy said. At Nagaro, Reddy said, engineers are waiting about three months to deploy the latest AI Lab model.

“We’re not one of the first early adopters to jump on it,” he said.

— CNBC’s Kif Leswing contributed to this report

WATCH: When is Anthropic IPO?