The software giant and Marc Benioff made a number of promises in AI on Monday, from new products to US$500 million to invest in startups.
On Monday, Marc Benioff was back onstage to present Salesforce’s role in helping customers “trust” generative AI.
Courtesy of Salesforce
Salesforce wants to be a player in generative AI, and chief executive Marc Benioff’s pitch hinges on one word: trust.
At an event in New York on Monday, Benioff and Salesforce leaders presented the company’s current AI strategy, updated for a business world gripped by the potential — and risks — of generative AI.
“The most important thing we are doing in AI is trusted and responsible AI,” Benioff told assembled partners, employees, analysts and press. Customers, particularly large enterprises, are equal parts intrigued and concerned by the technology’s potential, he added. “We’ve all seen the movies and where this can go. We have all these crazy ideas in our head of what can happen.”
To punctuate the point, Benioff told the story of a recent dinner meeting with OpenAI CEO Sam Altman, one of the faces of the current generative AI wave. On a recent earnings call, Benioff had alluded to the meeting — significant given OpenAI’s close alliance with Salesforce’s CRM rival, Microsoft — but had said that sharing details of what Altman demonstrated was not appropriate. On Monday, Benioff divulged more, revealing that Altman had used a newer, unpublished OpenAI model to simulate Benioff giving a 1960s speech from John F. Kennedy, produced from just a few seconds of Benioff’s recorded voice.
Impressed by the model’s accuracy, Benioff said he then asked where the audio sample of his voice was stored, and how the file was secured. Because it was input into the model, it no longer could be located or extracted, he said Altman replied. Not a problem for a parlor trick. But a concern if, say, a CEO of a bank needed to know where his data was for compliance reasons, Benioff explained. (Altman and OpenAI didn’t respond to requests for comment on that anecdote or tech.)
For Salesforce to use such AI tools with its own customers, Benioff argued, it would need more protections, a “trust layer” bridging corporate users to the model providers on the other end, without permanently sharing data. “We understand the burden there must be on us as we’re trying to take this forward.”
This is Salesforce’s angle, helping to explain why it’s invested in two foundation model challengers to OpenAI so far, Anthropic and Cohere. Back in 2016, when Forbes profiled Benioff for a cover story, Salesforce’s AI savior was a new product called Salesforce Einstein, complete with mascot, which would use machine learning to integrate predictive insights across Salesforce’s suite of sales, customer service and marketing software. Now, it’s generative AI, but Salesforce isn’t looking to win on building the models themselves (though it’s developing some). If generative AI is the most important tech of a lifetime, as Benioff said he now believed, his company’s role is as custodian and guide to make it useful and safe.
The ‘AI vacuum cleaner’
In a warm-up to Benioff’s address, a Salesforce emcee asked an employee from its customer Peloton what they thought about AI. Their answer — that they were excited but concerned about AI’s data risks — didn’t seem like a coincidence. And when Salesforce AI CEO Clara Shih revealed that Salesforce plans to ship pretty much exactly what you’d expect (think generated sales emails, recommended responses for service agents, code for developers and segmentation of customers), data protection was at the presentation’s core.
“Different models are going to have different strengths and weaknesses,” Benioff said in response to a question about his favorite large language model (LLM) providers. “I think every company is going to end up getting in their lane.” Salesforce will protect customers from plugging their data into any model with security concerns, he added, when asked if customers should trust sharing proprietary info with Salesforce’s AI partners. “Customers are not going to lose their data to an AI vacuum cleaner,” he claimed.
Salesforce AI CEO Clara Shih said Salesforce’s AI tools would learn and improve from business outcomes, not simply feedback on prompts.
Courtesy of Salesforce
In an interview, cofounder and CTO Parker Harris and Salesforce Research chief scientist Silvio Savarese said that Salesforce will deploy its own models alongside those others depending on how customers use such AI tools over the next six to 12 months. Generally speaking, they anticipate Salesforce’s models helping with use cases like generating new code in text or completing code, and specializing on customer use cases in specific areas such as healthcare, or the creation of visual marketing materials.
“OpenAI is trying to solve for all of human knowledge asking anything, which is beautiful, and we can use these models for our use cases,” Harris told Forbes. “But an email in a healthcare business to a customer is a very specific program.” Over time, customers will be able to train their own models within Salesforce’s AI capabilities, he added, when cost-effective. “We will get there, but we believe we can go a long way [with base models] first,” said Savarese.
Seeding startups
Key to Salesforce’s credibility in AI is a history of startup acquisitions, including last decade’s purchases of RelateIQ, PredictionIO and MetaMind. When Benioff revealed Einstein – unsuccessfully urging a pained Harris to dress up as the famed scientist – seven years ago, it was with MetaMind founder Richard Socher as his new AI guru. Socher’s long since left Salesforce and is now the CEO of startup You.com, one of six investments Salesforce said it has made in AI in recent months. On Monday, the company also announced it was doubling a $250 million commitment to invest in generative AI startups made in March to $500 million.
In an interview, John Somorjai, who oversees Salesforce Ventures and the company’s startup acquisitions, said Salesforce was doubling down because it was “just overwhelmed” by the number of investment opportunities it saw in the category. While it’s become a meme in venture capital circles how much attention and dollars are flowing into AI startups, Somorjai argued that cheaper compute costs and off-the-shelf models that are getting more powerful and cheaper to train and use are creating new opportunities for startups built on top of them.
“OpenAI’s Achilles heel is their relationship with Microsoft; OpenAI’s superpower is their relationship with Microsoft.”
Salesforce Ventures expects to close five more AI investments in June, adding to the six it’s made so far (Anthropic, Cohere, Hearth.ai, Humane, Tribble, You.com). It’s looking at several additional foundation model providers beyond Anthropic and Cohere, Somorjai said. He echoed Benioff’s argument that business leaders are “so worried about what AI can do,” and that Salesforce’s teams can help evaluate and propagate the “responsible” models when it comes to data privacy and security. “AI is only as good as the data it’s trained on. You want to make sure that the company that’s collecting the data is doing it in a really responsible framework,” he said. Any startup Salesforce backs in AI will be vetted by its own researchers to make sure they “apply our ethical framework,” he added.
Salesforce is far from the only corporation making strategic investments in generative AI startups, of course. Oracle and Nvidia also invested in Cohere, with Oracle touting the move in its Monday earnings call (founder Larry Ellison and Benioff, his one-time protégé, have a history of one-upping each other with simultaneous press events). For the startups, it’s a no-brainer to take Salesforce’s money along with potential access to its own huge employee base, as well as its portfolio of product offerings, which include Slack and Tableau, several industry experts said. Salesforce, meanwhile, gets a seat at the table with potential challengers to the OpenAI-Microsoft power couple, while bolstering its own line of business that has focused more on predictive AI in recent years. (An AI Starter Pack will run customers $360,000 for an annual contract including 50 CRM, Slack and Tableau licenses along with Einstein, MuleSoft and data cloud usage and access to Salesforce Services support.)
“The significant Microsoft investment in OpenAI set the stage for others to say, maybe that’s something we should do,” said one AI-focused investor who asked to speak anonymously because their portfolio startups do business with some of those players. “OpenAI’s Achilles heel is their relationship with Microsoft; OpenAI’s superpower is their relationship with Microsoft.”
Peanut gallery placated
When it comes to launching new products — especially ones that are still works in progress — Salesforce and Benioff have perfected the shtick. Einstein didn’t set the world on fire, but years later, Salesforce customers use it to make one trillion predictions per week, the company said.
Like back then, Benioff and co. appear to have found cautious optimism for their new generative AI “trust” offensive. Next-day reports from analysts at Cowen, Mizuho Americas and Stifel all adopted a cautiously optimistic stance on Salesforce’s ability to capitalize on the opportunity. “While Rome was not built in a day and neither will CRM’s AI strategy, now this gives Benioff & Co. a major AI horse in the race to cross-sell its installed base,” wrote Wedbush’s Dan Ives on Tuesday morning.
And the narrative of customer caution, and a brewing storm over data security in working with AI models, resonated with at least one former Salesforce exec whose new company is focused on the same problems. “Every data system we’ve ever used, there’s a delete button. But AI models don’t delete,” said Anshu Sharma, CEO of Skyflow, a startup looking to protect shared business data. “Everybody’s struggling to figure out what the right answer is.”
This story was first published on forbes.com and all figures are in USD.
