Marc Benioff's Big Bet on AI Agents
1 billion digital agents in 15 months? Salesforce CEO calls strategy a "high wire act."
Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who worked in strategic comms at Oracle, IBM, and MongoDB. I invite you to subscribe, share, comment, and connect with me on LinkedIn. This post is sponsored by Salesforce.
Marc Benioff, this blog post is for you.
Following your Dreamforce ‘24 keynote, you met with the media and invited us to share our feedback. Here’s mine on the Agentforce strategy you laid out to the 45,000 Dreamforce attendees.
I see huge potential in the concept of autonomous, digital AI agents for customer-facing businesses. I’ve been writing about business initiatives aimed at improving the customer experience for more than 25 years. But it’s not working. In too many areas of everyday life, customer service and support is still woefully lacking.
Salesforce, as the CRM market leader, is ideally positioned to meet this challenge. Service agents, support agents, personal shopping agents, B2B buying agents, banking/health/finance agents — we need all of these and more.
But Salesforce and its customers still have a long way to go to get to the 1 billion agents you set as a stretch goal by the end of next year. As you said yourself, Marc, this kind of wholesale change and disruption is a high wire act.
There are naysayers and risks, and I’ll get to those, but let’s start with the upside. The payoff, if you get it right, is vastly-improved customer experiences, higher business productivity, and new opportunities and career paths for employees who are stretched thin in low-skill jobs.
Early movers can get a jumpstart with Salesforce’s out-of-the-box agents. We heard about one customer who stood up an Agentforce agent over the weekend.
It may be fast and easy to get started, but let’s be honest — there’s still a fair amount of fine tuning required to educate these agents on the idiosyncrasies of an organization’s products, processes, and customers. An agent may take a few minutes to activate, but days or weeks to bring up to full speed. Not unlike onboarding a new employee.
The Agentforce tech stack
These AI-powered digital agents are the topmost layer of Salesforce’s modern AI stack, which includes an agent builder, AI/LLM model builder, and prompt builder. There’s also a reasoning engine dubbed Atlas that’s designed to mimic how humans think. Here’s how Salesforce describes Altas:
“It starts by evaluating user queries, refining them for clarity and relevance. Next, it retrieves the most relevant data and builds a plan for execution. The process then refines the plan further, ensuring it’s accurate, relevant, and grounded in trusted data. This reasoning process allows Agentforce to autonomously reason, make decisions, and complete business tasks, all while delivering precise, factually accurate results.”
The catch: Good, trusted AI requires good, trusted data. Which means businesses need modern data management — with timely, well-orchestrated data pipelines; support for both structured and unstructured data; security and governance; and data catalogs to ensure data quality. This is where Salesforce Data Cloud comes in, an all-encompassing source of data that will drive the agents’ ability to make decisions and take actions.
Data-driven AI agents
In a recent LinkedIn post, I referred to Data Cloud as the superpower behind Agentforce. Data Cloud has grown 1,000 times over the past few years, making it Salesforce’s fastest growing organic innovation ever. That’s according to Rahul Auradkar, Exec VP and GM of Data Cloud, who gave me an overview and demo a few days before Dreamforce.
For years, organizations have managed CRM data using old-style Customer Data Platforms. Data Cloud is more comprehensive than that. It’s a data lake with integrations to sources outside of the Salesforce Platform — the superset of Salesforce technologies that include Tableau, Mulesoft, and Flow business-process automation software.
A key capability in Data Cloud is “zero copy,” which obviates the traditional extract/transform/load (ETL) process typical of data warehouses. Salesforce is building zero-copy integrations with Snowflake, AWS Redshift, Google BigQuery, Databricks, and others. The beauty of this approach is that it enables data access without copying, moving, or reformatting — providing near-real-time data to the reasoning engine and agents.
Data Cloud has more than 200 pre-built connectors and additional integrations through Mulesoft Direct, enabling access to widely used apps like Google Drive, Microsoft SharePoint, and other data sources.
At Dreamforce, Salesforce announced Data Cloud enhancements, including native processing of audio and video; a semantic data model that enables agents and humans to use data consistently; and real-time data activations.
Interest in Data Cloud is high, with 130% growth in the second quarter compared to a year ago. In fact, I couldn’t squeeze my way into the Data Cloud keynote. The session was filled to capacity, and I was stopped at the door.
This tells me that customers are looking to build and expand their data clouds. So they should be ready to go with Agentforce when the time comes.
Job growth or destruction?
How quickly will customers begin to deploy Agentforce? If the crowded Agent Launchzone was any indication, many are gearing up. You set a goal of helping attendees create 1,000 prototype AI agents during Dreamforce; there were more than ten times that number.
The idea is that AI agents will augment human workers by handling routine tasks. By one estimate, 41% of a worker’s time is spent on low-value, repetitive work. In an example of how agents can help, Salesforce demoed a digital agent handling a wrong-size merchandise exchange for retailer Saks.
Yet, what if these AI agents become so good at this kind of grunt work that human agents are no longer needed? There’s real worry about job destruction and layoffs.
Salesforce’s answer is re-skilling, redeploying, apprenticeships, and working with its ecosystem (customers, partners, communities) to create new jobs and opportunities. In one encouraging sign, Ben Kus, CTO of Box, said Box software engineers are using AI to build code. That brings efficiencies, he said, but not reason to downsize. “We don’t want fewer software engineers. We want more.” Reason for cautious optimism.
Hallucinations, bias, misuse
Salesforce isn’t the only game in town. Microsoft, Oracle, ServiceNow, Workday, SAP, and others are introducing their own AI-enabled agents and copilots. CIOs and CTOs will have to sort out the many and growing options.
What can go wrong?
The risks include hallucinations, bias, privacy, security, unpredictable behavior, misuse, and hacks. Some worry about agenic AI, where independent agents are able to act on their own. And the prospect of autonomous agents interacting with other autonomous agents feels eerily unpredictable. Early movers may need to tap the brakes as they work through all of this.
Salesforce has developed a “trust layer” to assuage such concerns. It bakes policy and principles into what AI agents are allowed to do and not do, a.k.a. guardrails. Capabilities include bias and toxicity detectors; security, privacy, and safety controls; audit trails; and the ability to “nudge” human intervention if needed. The trust layer serves as a gateway to LLMs to filter out unwanted behavior or inaccurate outputs. This helps reduce hallucinations.
The hoped-for result will be accurate, trusted AI agents that raise the quality of engagement for customers and the quality of life for employees. In the process, issue resolution can increase (to 90% or more, per Salesforce) and customer satisfaction can spike upward.
Marc, you summed it up this way: “You’re going to have that moment where you say, ‘Wow, I’m deploying autonomous agents for my company. And it’s working!’ ”
That’s the vision, the promise, the sought-after outcome. Yet, trust in AI agents must be earned, not given. Hopefully, frustrated end users will not be left on hold, pressing buttons, for much longer.
Thanks for sharing your thoughts John - great to meet you at Dreamforce :)