AI safety debate meets reality at Dreamforce as business leaders say last year's models are enough

At Salesforce's annual Dreamforce mega-conference in San Francisco this week, the keynote conversations between CEO Marc Benioff and the heads of Anthropic, OpenAI and Nvidia delved into the raging AI safety debate, with the execs opining on whether model development is moving too fast.

Nvidia CEO Jensen Huang, while on stage with Benioff, urged frontier labs to "run as fast as you can."

But on the grounds of the 50,000-person event, described over a decade ago as the "Super Bowl of software," the discussion was very different. Far from obsessing over the fears of artificial intelligence and whether the rapid advancements of AI models threaten human existence, attendees said they're having trouble even taking advantage of the technology as it exists today.

"It's already hard enough to keep up," said Alec Bronston, who works at Chicago-based retail data company Spins as senior Salesforce director. A potential slowdown would present "a lot of opportunity to just even catch up and get our feet wet," he said.

Dreamforce had a different tone this year, landing in the middle of a dramatic moment in the AI boom. Days before the conference's kickoff, an Anthropic researcher resigned and, on his way out, said the top labs were "gambling with our lives." That led Anthropic's Dario Amodei and OpenAI's Sam Altman to propose safety initiatives and to push 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 and batting down the "SaaSpocalypse" narrative, trying to convince investors, customers and employees that AI is an accelerant to its business rather than a threat. Salesforce shares have dropped 8% this year even after a massive pop in August, while other software names like Adobe and Autodesk are down significantly more.

During Salesforce's last earnings call in August, the company unveiled "Claudeforce," a new way for salespeople to access critical data directly from inside Anthropic's Claude chatbot.

On the exposition floor, where companies in the Salesforce ecosystem pay big bucks to tout their software and services, the most highly trafficked booth on Tuesday was Anthropic's. Youthful employees of the AI lab donned oversized white sweaters and carried out demonstrations for passersby.

Salesforce customers and partners at the conference told CNBC that older and cheaper AI models are plenty powerful for everyday sales and customer service work. Many users are still figuring out their AI budgets and determining if their companies are best served using models from Anthropic or OpenAI, or instead opting for cheaper open-source alternatives.

"The frontier models are way ahead already," said Jaya Rohit Vuyyuru, a vice president at consulting firm SummitX. "A lot of the customer base is still getting their feet wet. There's still that gap where the clients can still fulfill and start getting a sense of what agents can do."

Meanwhile, as guests traversed the sprawling Dreamforce campus, or what Salesforce calls the Campground, they interacted with cartoonish corporate mascots, resulting in viral memes on social media. The mascots offered just a slight distraction from the AI chatter that was taking place across the event.

Some attendees said their employers have adopted Salesforce's Agentforce tools, which can respond to customer service questions and sales inquiries. For bots built with Agentforce, Salesforce isn't relying on Anthropic's cutting-edge Claude Fable 5.1 or OpenAI's new GPT-6 Astra, according to a support page on its website.

"The majority of agentic outcomes aren't driven by frontier capabilities," said Tim Sanders, chief innovation officer at software reviewing company G2. "They're driven by last year's AI. It's not that relevant to agentic providers, and certainly not that relevant to SaaS," or software as a service.

Kevin Lee, technology chief at cloud contact center software vendor Nice, said his company doesn't count on high-end models such as Fable for most workloads.

"In large part, with the models that are out there already today, and even one generation behind, they are highly performant and effective at doing the things that our customers need," Lee said. "It's almost like everything beyond this point is icing on the cake."

That's not to undersell the power of AI across the software sector. Many executives say it's been transformational for their business.

Electronic signature software developer Docusign, a Dreamforce exhibitor, uses "all the big frontier models as well as some of the open-weight models," said CEO Allan Thygesen.

According to Docusign's website, "larger frontier models are reserved for judgment-intensive work — complex clause analysis, multi-document reasoning, summarization — where general reasoning capability is worth the higher per-call cost."

Frontier models are generally proprietary, preventing users from seeing their data sets or fine-tuning the technology for their own purposes. Open-weight models, by contrast, can be downloaded and run on any infrastructure.

Docusign's home for the three-day event this week was across the street from San Francisco's Moscone Center at the Canopy by Hilton. The company rented out the whole hotel, reserving plenty of quiet meeting space. Thygesen said in an interview there that Docusign uses a technique called model routing to direct each request to the most cost-effective AI system.

It's a tool that Nice and many other companies now use as one major way to keep costs in check while capitalizing on AI.

Nice's Sanders said cloud software companies also have to figure out how to play in the new token-based economy, where people pay based on their use of AI and what they demand by way of outputs, rather than on a subscription basis.

Software companies have to adapt both in their use of technology and in how they deliver their own services as customers shift to tokens, Sanders said. A single token represents about three-quarters of a word.

"SaaS up to now has had no variable cost to deliver services," Sanders said. "The profit margin structure changed."

Sanders said that switching from a typical SaaS model to an agentic model could pull down gross margins, or the profit left after accounting for the cost of goods sold, from upwards of 85% to closer to 45%.

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

"Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks," Patrick Wendell, engineering vice president at Databricks, wrote on X, referring to models from Anthropic and OpenAI, respectively.

Nagarro, a system integrator, is in a very different spot.

Ram Reddy, technology chief for consumer industries at Nagarro, hung around his company's booth this week at Dreamforce. Some who stopped by talked about regulation, following Amodei's chat with Benioff.

But the hottest topic, Reddy said, was Claudeforce and not the future of the frontier. Reddy said that at Nagarro, engineers wait about three months before plugging in the latest models from the AI labs.

"We are not one of those first early adopters jumping at it," he said.