How Salesforce deploys its AI agents internally

In a late June interview, Salesforce CEO Marc Benioff claimed that AI now does 30% to 50% of the work at Salesforce. But what does that actually mean for employees?

Salesforce hired its first chief digital officer, former Disney, Major League Baseball and SiriusXM executive Joe Inzerillo, earlier this year to co-manage the digital experience of Salesforce tools internally and on its customer-facing sites and apps. In his first interview with the press, we explore Benioff's claim with Inzerillo, how the company deploys agentic AI tools to its employees and customers, and much more.

Editor's note: This Q&A has been edited for clarity and brevity.

Would you consider yourself the manager of Salesforce's digital labor workforce?

Joe Inzerillo

Joe Inzerillo: Sort of. Jim Roth runs our support organization. I build the tools. I make sure those tools have the fit and finish. I do a lot of multi-variate testing, and things like that. Ultimately, I'm empowering other parts of our business to use those tools on behalf of the customer.

I don't think anybody fully understands how the digital workforce will come into being, but I have a feeling that it will have a bunch of parents at any given time. I don't think there will be that many instances where it's just one group. It takes a village to raise a child, so to speak, in the digital workforce.

Marc Benioff said that AI does 30% to 50% of the work at Salesforce. What does that mean to you, and is it 30% or is it 50%?

Inzerillo: It's a complicated question to answer, because AI is used all over the place, in all sorts of different tools. AI is embedded into things like Cursor. There's no question that we probably write about 30% more code than we did, and since a bunch of what we do is write code here, that's a pretty good anchor point.

We also use AI deployed in Agentforce, as well as other tools that have embedded AI. Every week, every day, we're seeing greater adoption of these things. So it's a defensible number for sure.

Many people think of it as, "I'm going to replace human beings with agents, I don't need as many humans." It's more nuanced than that. Agents are good at certain things, and humans are good at many other things. So I've found that it's less about the 30% of the work being done by agents and more that we're doing 30% more work. Maybe that would be work a fairly junior person inside the organization would wind up doing -- calling people back, scheduling meetings -- doing all those types of things that we can now really automate with an agent in a way that's satisfying for everybody. The humans then go on to other things.

Do you worry that your agents are getting rid of many low-level tasks and maybe entry-level jobs, too, raising barriers to entry into the workforce?

Inzerillo: It's a good question -- one of the things that's always one of our company values is trust. The human aspect of trust is how we have empathy. How are we compassionate to our customers and to ourselves internally? We're not super cavalier about what we're doing; we think about them.

But I think that jobs change. For example, when I was in college, you weren't allowed to use a programmable calculator. Now, kids bring laptops to class. The job of being a student has changed. Before, I had to be able to do arithmetic in my head.

I'm an optimist. I believe these tools are just bringing everybody up to a level where the new entry-level job is just a different job. It's not the exact same job that it was five years ago. Five years from now, it'll be radically different, I suspect.

One thing about the agentic revolution that I can say -- having lived through the web revolution, the mobile revolution, the streaming revolution and all those sorts of things -- is that this one is the fastest I've ever seen. But we're not seeing a displacement effect yet. We're seeing force multiplication.

Are you involved in governing agents, reeling them in if they behave badly, like the ones on X?

Inzerillo: Yes. We have an AI governance board, which is cross-discipline: Technologists, legal folks, employee success folks, people like that. We're thoughtful about where we decide to deploy agents, and we're also thoughtful about evaluating the efficacy of those agents, both objectively and also, are they doing the right thing?

Agentforce is different from other agents in that people take a raw large language model and just go do stuff with it, and it goes completely off the rails. Safeguards are built into Agentforce so that we don't have the thing where all of a sudden it goes anti-Semitic or something like that. That doesn't happen, because Agentforce is not just a large language model. It's wrapped in trust layers, data grounding and all these things that prevent it from getting that far out.

What tasks can Salesforce's service agent execute on its own, besides information lookup?

Inzerillo: There are certain things you can do -- reset passwords and things like that -- that agent could do on your behalf. The agent that's on Salesforce.com, for example, could give you a bunch of information, but it could also refer you to somebody to reach out to you as a prospect.

The more interesting, cool things they are doing are more back-of-house things. Those tend to be more complicated, where, for example, you can book travel and/or approve an expense in a natural language way. We're starting to use Tableau Next inside the company. Natural language prompting of Tableau is pretty powerful, where you can ask it a question, and you get a Tableau dashboard that you can then interact with.

Have you reduced the size of your contact center yet?

Inzerillo: This gets into the "jobs to be done" thing. There are definitely jobs the agent is now doing that people used to do. People are doing other things, engineering or different and deeper help center activities. We've shifted around resources because the agent is taking a lot of the load at this point. That first hit to the help site that used to be a human chatting is now the agent -- and the agent's better and faster than the humans were at it.

Salesforce's service agents handled their millionth support request last week. Do you see as much of the agentic AI being used on the sales side?

Inzerillo: Definitely. We have a lot widespread [deployment in sales], especially our integrations into Slack, where people use it every day. They don't even really think about it as being an agent doing the work anymore. They just think, "I go to this particular channel in Slack, and I get answers to the things I want to get answers to."

We also have a ton of pilots going on. Sales is much more complicated than service. The job to be done is more complicated, so it's taken us more time to deliver real value there. For example, if a salesperson spends 40% of their day researching their customers or researching things for their customers, and you could offload -- just picking numbers here -- 90% of that 40%, it gives them more throughput and allows them to spend more time with the customers.

We don't have any data to release at this point. I'm sure we will in the near future, probably at Dreamforce. But we're absolutely contributing to sales.

Don Fluckinger is a senior news writer for Informa TechTarget. He covers customer experience, digital experience management and end-user computing. Got a tip? Email him.