Salesforce not ready to unleash generative AI on its customers

Salesforce was an early adopter of artificial intelligence (AI) with its Einstein recommendation tools, but it is taking a cautious approach to deploying the latest AI trend, generative AI.

It’s been a month since Salesforce CEO Marc Benioff tweeted, “Get ready to be wowed by Salesforce EinsteinGPT! It generates leads, closes deals, and even makes coffee (just kidding, but wouldn’t that be amazing?)”

On the eve of the company’s Trailblazer DX ’23 developer conference, though, Einstein GPT still wasn’t ready. “We’re still very early days,” Salesforce Service Cloud GM Clara Shih said in a conference call to demonstrate Einstein GPT to journalists. “We’re testing and validating and iterating before we launch these products as generally available.”

Salesforce has committed to building AI that is responsible, inclusive, accountable, transparent and empowering, a series of values similar to those espoused in the Rome Call for AI Ethics.

Interest in generative AI from customers is high, Shih said. “We have pilots across every single cloud that is happening, and many of them are oversubscribed,” she said.

It’s easy to see why they’re interested: Generative AI tools such as ChatGPT can write sales proposals or respond interactively to customer complaints far quicker and more cost-effectively than a person. But there are dangers: the models such generative AI tools use to decide which word to write next can occasionally “hallucinate,” generating wildly inaccurate or inappropriate recommendations, and making it risky for enterprises to use them unsupervised.

In a period of great economic uncertainty, enterprises are focusing on customer experience and customer retention, said Gartner distinguished analyst Gene Alvarez. “AI brings hopes of providing great customer experiences at scale while not driving costs higher,” he said.

But imperfections in the output of today’s generative AIs mean that human review is needed, he said. “Salesforce appears to be taking its time to navigate all these issues in advance of the release, which is a prudent approach,” he added.

What will Einstein GPT do?

Einstein GPT is not one product but five: Einstein GPT for Service, for Sales, for Marketing, for Slack and for Developers.

When they’re eventually released, they’ll offer enterprises ways to incorporate suggestions from a generative AI model into their Salesforce workflows. These suggestions will be based on the data used to train the generative AI model, and also on the data held in an enterprise’s Salesforce system.

“We’ve got the relevant context to really make the most out of generative AI,” said Shih. “Our Salesforce Data Cloud means we can allow the generative AI to continuously adapt as customer information and needs change.”

Einstein GPT for Service will prompt service agents with responses to customer questions and requests, summarize the interaction to create case notes, and even extract relevant information to generate new knowledge base articles, she said.

For sales staff, Einstein GPT for Sales will summarize news about accounts and identify key contacts, as well as generate drafts for sales emails. The creation of targeted email campaigns, ads and landing pages will be the preserve of Einstein GPT for Marketing.

Einstein GPT for Slack will provide natural language access to CRM functions from within the Slack messaging platform that Salesforce acquired in 2021, while Einstein GPT for Developers will automatically create code snippets and test cases based on comments in code.

For Gartner analyst Kyle Davis, there’s potential for generative AI to assist developers in writing better quality code.

“This is different from developing a whole application but provides valuable assistance in smaller scoped scenarios,” Davis said. “I see this with Microsoft using Power Apps Ideas to help developers create a Power Fx formula that would’ve been difficult to create on their own, especially if they’re new to Power Apps development. The same is now being incorporated into Salesforce with ApexGuru. Poorly written Apex code can cause low-performing applications.”

Einstein GPT: Not just GPT

Despite Einstein GPT’s name, GPT won’t be the only generative AI tool it uses. Salesforce plans to combine its own AI models with others from an open ecosystem of vetted partners, including those of its launch partner, GPT developer OpenAI, said  Shih.

“This means customers have choice,” she said. “They can bring their own generative AI models or choose one of our out-of-the-box generative AI options.”

In all the demonstrations of Einstein GPT during the conference call, one feature remained constant: Whenever Einstein GPT generated any text, another click was required before it was shared with another person.

Jayesh Govindarajan, SVP of engineering, Einstein and Bots, explained: “You saw generative AI operating within the trust boundary, with a human in the loop at all times. Another powerful benefit to having humans in the loop is that the edits and the refinements go back into reinforcing the model, and over time, improve their performance. The more you use, the better it gets.”

Requiring a person to click “OK” may not be sufficient to keep AI hallucinations from causing trouble, particularly if that person’s compensation depends on how many queries they respond to or how many lines of code they write, so CIOs will need to take a holistic view of the checks and balances they build into their processes, not just look at the user interface.

There are other UX challenges too, Gartner’s Alvarez notes: Just because a bot can tailor a response using a customer’s personal information quicker than a human operator could, doesn’t mean it should.

“There are ethical issues to be wrestled with, such as the thin line between being helpful to a customer and being invasive,” Alvarez said. “This is where ethics on customer etiquette are needed to avoid interactions that will drive customers away.”

While Salesforce is taking a cautious approach to introducing generative AI, its partner OpenAI has gone full speed ahead. OpenAI uses Slack internally, and saw early on the potential for plugging its ChatGPT bot right into its corporate Slack channels.

It has now released a beta version of its ChatGPT app for Slack that anyone can sign up to use. It will appear in the Slack App directory, subject to the usual corporate approvals for app integrations. Once installed, it adds two functions to Slack’s “More actions” button: “Summarize thread” and “Draft reply.” It’s also possible to use the app as a search tool by asking it questions. Salesforce wants other companies to pile in on generative AI, and its investment arm, Salesforce Ventures, has set up a new fund to support them. “We’re launching this $250 million AI investment fund to nurture the next generation of AI startups and to guide the ecosystem in developing generative AI in a responsible way,” Shih said.