Editor’s Note: This excerpt is adapted from Agentforce: Harnessing the Agency of AI to Scale, Grow, and Lead Any Industry.
When Salesforce unveiled Agentforce in October 2024, we started getting candid reports from the field. Customers and partners were excited and optimistic about AI and the agentic future, but there were challenges. One of the most common questions was: Can I really get this to work?
Some reluctance was understandable. Even the iPhone was mistrusted at first. To address this challenge, product and engineering teams set to work building a portfolio of off-the-shelf agents that could be powered up quickly and adapted without too much angst. They focused on a handful of the most common, agent-ready workflow problems in the enterprise. Much like human labor, agentic labor was divided into functional areas where Salesforce customers and customer intelligence said they’d be the most useful. These areas were service, sales, marketing, commerce, and HR.
The first prebuilt agent was called Agentforce for Service. It was designed to engage with customers and complete customer-facing support tasks — 24 hours a day and in natural language, guided by brand policies and even tone of voice. In sales, we offered two prebuilt agents: Sales Development, which would nurture inbound leads 24/7, and Sales Coaching, which gave every rep a dedicated coach to improve productivity. In marketing, we developed a couple of prebuilt agents to support campaign creation and loyalty program management. We offered three types of prebuilt agents for commerce: Agentforce Merchant, which sets up online stores; Agentforce Personal Shopper, a 24/7 personal shopping assistant; and Agentforce Buyer, which helps B2B buyers locate products and check order status. And we teamed up with Workday to create a joint prebuilt agent to handle common employee HR tasks like onboarding and personalized self-development.
That was just the beginning. Salesforce now offers more than 200 prebuilt agents tackling everything from deep research to appointment scheduling.
That was just the beginning. Salesforce now offers more than 200 prebuilt agents tackling everything from deep research to appointment scheduling.
But most of these prebuilt agents are intended to be a start, a relatively easy way for Salesforce customers to get over any initial reluctance and envision what a custom AI agent could do for them. And it quickly became clear that, once our customers became comfortable with the idea, there was no limit to the creativity they could apply and the business problems they could address.
From prebuilt to custom
By talking to customers and observing early adopters, the Agentforce team pretty quickly saw some themes emerge: areas where AI agents were being actively used and proving useful in real situations. They were as follows:
- Deflector: Handles a large number of tactical tasks that aren’t necessarily big-picture but require continual work and attention to detail. These are particularly useful when teams are overwhelmed by simple tasks — like too many product returns.
- Co-worker: Augments a workforce with additional capacity — for instance, an extra hand on the help desk to give customers a better level of service.
- Advisor: Provides important advice at the right moment, either to customers or employees — for instance, to recommend cross-selling opportunities, identify potential breakpoints, or coach a sales team.
Successful agentic brainstorming processes applied these roles to their own team’s needs. In general, they focused on a few key areas:
- Communication: This is where the new generation of AI models excels. Most obviously, it’s essential for customer-facing functions like service, sales, and employee relations. But there are other forms of communication — doing analytical work behind the scenes, or communicating with other agents.
- Scaling fast: Anything that requires an army of humans accomplishing simple tasks is likely not a big revenue driver, and ripe for automation.
- Searching: Processes where a human or process has to search through vast troves of data, particularly unstructured data, can often be agentized.
- Screening: Agents can also unlock processes where a large number of entities go through a set of stages or gates for assessment, filtering, or triage (think: school admissions, applications, sales leads)
- Recommending: Machine learning technology has powered recommendation engines for decades, and agents can take this a step further — providing personal shoppers or in-store concierges.
- Digitizing: More broadly, agents can play a role in the digital realm that humans (or something else) play in real life — like content creation or psychotherapy.
Most companies deployed their initial agents as tests — both to validate the technology and to train their teams. New ideas emerged from this phase and led to larger deployments in more functional areas. But as they did so, another question emerged.
The human question
How “human” should agents be?
There is some debate about where to draw the lines. Because AI agents can now use some very natural language, it’s tempting to see them as faux humans — a troop of Mini-Mes. But this assumption has proven not only incorrect but misleading.
In the early days of chatbots, some builders assumed they would simply impersonate people until they could not, at which point a real person would pick up the thread — without the customer even noticing. But we humans are good at detecting nonhumans; by definition, we are experts in humanity. Indeed, there is a phenomenon called the “uncanny valley” – that realm lurking between our real world and a facsimile that’s almost right. This very quality of almost-ness disturbs us to our core processors. That is why Pixar films don’t try for absolute photorealism.
Very quickly, assumptions were changed. It became standard practice for chatbots to identify themselves as “AI assistants” or the equivalent, much as Siri or Alexa never claimed to be real people, as far as I know.
This proved to be a good approach for agents as well. First, customers are already beset by AI agents and are comfortable interacting with them. And having an AI agent announce itself upfront minimizes the risk of surprise or disappointment. For example, imagine talking to a customer service agent about a health condition, and they express some sympathy. You share some more. They’re patient and understanding, like Dr. Jennifer Melfi in The Sopranos. Then at some point, they say, “Thanks for sharing, I’m going to transfer you to a real person now.” How would you feel?
Likewise, it’s best to avoid the mistake made by an HR technology company that announced it would give digital workers personas, putting them on organization charts with managers and titles. Beyond the optics, this could impact the morale of a company’s human workforce.
That’s why AI agents should not be presented as robot-for-human replacements. They don’t take up space on an org chart. They are digital labor embedded in workflows, with the emphasis on the syllable work. They should not be referred to as a “customer service associate” or “rep” but rather simply labeled by their work function: “customer service,” “returns,” and so on.
Today, thousands of Agentforce agents are already on the job, with many more to come.
Today, thousands of Agentforce agents are already on the job, with many more to come. It’s easy to forget just how new this technology is. Our customers can sometimes feel like they’re learning how to drive an FI car that’s already on the track.
The good news is we’re all learning fast. With an enterprise-grade platform like Agentforce, almost anyone can build and deploy powerful AI agents to augment and empower their human workforce. We’re only just starting to unlock the power of the platform.
More information:
- Purchase the book on Amazon
- Learn more about Agentforce and read the Agentforce 3 announcement