Salesforce AI Cloud Seeks To Combat ‘Trust Gap,’ Hallucinations
Wade Tyler Millward
‘When we talk about privacy, when we talk about hallucinations, when we talk about data control, when we talk about bias, when we talk about toxicity … these are actual technical explanations of things that are happening inside these models,’ says Salesforce CEO Marc Benioff.
Salesforce’s latest move in the enterprise generative artificial intelligence arms race is AI Cloud, a series of capabilities that seeks to close a “trust gap” co-founder and CEO Marc Benioff said he sees in current generative AI offerings.
The San Francisco-based CRM software vendor unveiled the new capabilities Monday during an event in New York and streamed online. The AI Cloud “starter pack” is available for $360,000 a year with some AI Cloud features already generally available (GA), according to the vendor.
“When we look at these new models that everyone’s going to roll out, we know one thing, that is there’s a pretty big gap … it’s this AI trust gap,” Benioff said during the event. “When we talk about privacy, when we talk about hallucinations, when we talk about data control, when we talk about bias, when we talk about toxicity … these are actual technical explanations of things that are happening inside these models.”
Salesforce AI Cloud
The news builds on Salesforce’s generative AI announcements earlier this year around the EinsteinGPT generative AI offering, as well as a greater conversation around trust and security in sharing corporate data with generative AI tools for analysis.
During the Salesforce event Monday, Julie Sweet, CEO of Salesforce partner Accenture, told the audience that customers believe AI will transform work, but they are cautious about the dangers of this new technology.
“I would emphasize this technology is really early,” Sweet said. “And one of the big advantages that I think you have by working with Salesforce is that trust and security are being built in from the beginning and you’re bringing that together with the deep expertise around the use cases. And that is really resonating with our clients who want to make sure that they don’t get ahead of themselves before some of these things have been worked out.”
Combating AI Hallucinations
AI Cloud can host Large Language Models (LLMs) from Amazon Web Services, Anthropic, Cohere and other vendors within Salesforce’s infrastructure, according to the vendor.
Users of AI Cloud can leverage Salesforce’s own LLMs for code generation, business process automation and other capabilities. Salesforce LLMs include CodeGen, CodeT5+ and CodeTF.
Salesforce is also working on AI prompts to ground generated outputs and provide generated content without hallucinations, the term for when AI produces incorrect results despite the training data.
Two AI Cloud features go into GA this month. One is Einstein GPT Trust Layer, which promises to prevent LLMs from retaining sensitive customer data.
Einstein GPT Trust Layer promises to help customers maintain data governance controls while still using generative AI, according to Salesforce. Trust Layer will provide deployment capabilities for any relevant LLM while meeting data privacy, security, residency and compliance goals.
Users bringing their own domain-specific models through Amazon SageMaker, Google Vertex AI and other products can connect to AI Cloud through Einstein GPT Trust Layer and keep the customer data in their trust boundaries.
The other feature to go into GA this month is Service GPT, which combines real-time Data Cloud data and AI trust capabilities to automate mundane service worker tasks.
Service workers will have the ability to automatically create service briefings, case summaries and work orders based on case data and customer history, according to Salesforce.
Pilots Coming Soon
AI Cloud features launching in pilot this month are Marketing GPT, Apex GPT, Sales GPT and Commerce GPT. Marketing GPT should go into GA in February. Sales GPT and Commerce GPT should go into GA in July.
Marketing GPT promises marketers the ability to generate audience segments and better target messages and offers to customers, according to Salesforce.
Apex GPT aims to help developers scan for code vulnerabilities and receive inline code suggestions from within the Salesforce development environment, according to the vendor.
Sales GPT can auto-generate personalized emails based on CRM data, according to Salesforce. And Commerce GPT can generate product descriptions based on buyers’ customer data plus give commerce employees suggestions for unloading inventory, increasing average order value and accomplishing other commerce goals, among other use cases.
Pilots launching later this year include Flow GPT in October and Tableau GPT in November. Flow GPT should allow workflow creation with natural language prompts. An example of Flow GPT in action is sales representatives receiving notifications when a lead becomes an opportunity, according to Salesforce.
Meanwhile, Tableau GPT should allow sales employees to generate data visualizations with natural language prompts. Employees, for example, should have the ability to display real-time progress against quotas and receive recommendations for meeting goals, according to Salesforce.
Salesforce customers are already embracing its AI offerings, according to the vendor. Its Einstein AI product powers more than 1 trillion predictions a week across the Salesforce suite.
Benioff called Monday’s news a “breakthrough” in security and trust for enterprise-level generative AI. He said that generative AI is “one of the most important technologies of our lifetime” and “maybe it’s the most important technology of any lifetime.”
But businesses in highly regulated industries such as health care and banking can’t accept the liability that comes with generative AI that hallucinates.
“There’s no company that’s even close to what we’re doing in the customer relationship management area of artificial intelligence,” Benioff said. “And we understand the burden there that must be on us as we’re going to take this forward.”