Agentforce is Salesforce’s new application platform that leverages AI to improve worker productivity and unlock new business opportunities. And at its recent Dreamforce event, Salesforce did what I thought was a great job grabbing first-mover status in the emerging AI agent space. It is indeed early times, and agents represent a new class of composite applications. So, although they have a great deal of potential, success with agents will also hinge on good governance and IT practices.
The Basics of Agents
An agent is a composite and autonomous application that enables humans and AI to collaborate to complete a task. By composite, we mean that the application uses both AI components such as LLMs and deterministic components such as workflows and APIs. Agents can act autonomously in the sense that a human may delegate most or all of a business process to an agent. However, on some level agents must also collaborate with humans in handling exceptions, getting updates and being monitored for compliance.
Agents are meaningful in a couple of very important ways. First, given that an agent will be created with an express functional purpose, its performance can be measured and a business value can be attributed to it. For instance, if an agent can execute a task in one-half or one-tenth of the time it takes a person, there is an expected cost savings. Salesforce also demonstrated many agents that might help increase revenues or customer satisfaction, so there is also a growth dimension that agents could achieve.
Second, an agent can help ease tensions about AI being used for a task. Unlike a chatbot, an agent works within the parameters of a process and is told which applications or services to use. This added context will help an agent deliver better and more precise results.
For more on this and why it’s important to enterprises, take a look at my deep dive on AI agents.
Salesforce Demonstrated The Power And Potential Of Agents
At Dreamforce, the power of agents was on display; by granting attendees easy access to test sandboxes and human experts, Salesforce helped accelerate the learning process for its install base. As Salesforce has arguably the largest low-code/no-code developer community, letting its users see and touch agents was the culmination of years of investment in these users. That investment included a grassroots AI world tour that I wrote about this summer.
What makes Agentforce especially interesting is how it has connected the various components of an AI app through a no-code Agent Builder. This helps a no-code developer assemble and configure an agent based upon numerous templates, workflows and APIs (via Salesforce’s MuleSoft integration layer).
There is also a Prompt Builder, which uses natural language to define how the agent will interface with AI or the outputs from Salesforce’s Data Cloud. Prompt Builder has some nice integration capabilities to assist in the task of prompt engineering, which is a welcome addition. When seen in action, this integration showed how users are able to build and deploy an agent in minutes. This level of integration and agility seemed to create a powerful positive impression for a group of users well versed in Salesforce’s existing data structures and product offerings.
Agents In Application Platforms Will Get Early Traction
It makes sense that application platforms such as Salesforce are coming out more aggressively in the AI agent arena. Agents are fairly complex, but application platforms have a lot of the necessary pieces for agents embedded within them. To wit, application platforms typically have:
- Low- or no-code developer tools and users
- Established use cases, workflows and business context
- A cross-platform data-sharing or integration platform
- Robust access control capabilities
- Workflow engines
All of these are necessary to build an AI agent. And while Salesforce has made the biggest and most vocal announcements so far, its competitors are also doing work in this arena. For example, in the weeks leading up to Dreamforce, both Oracle and ServiceNow made announcements specific to AI or AI agents. I also believe that over the next six months the AI agent space will heat up considerably across the entire tech industry.
AI Agents Have Great Potential, But Tread Slowly
While we are very bullish on AI agents, we also see this as a technology that needs careful consideration. So, a “walk before you run” approach is preferred. Specifically, this means:
- Experiment and play — The barriers to entry for agents on application platforms like Salesforce are relatively low, and there is a lot of help to get started. Your IT staff should try things out and begin to figure out what the possible business impacts could be.
- If you want to go faster, start internally — There are no best practices for agents yet in terms of testing, governance, legal considerations or deployment. While going light on some of these areas at first is okay within your own team, external usage is higher risk than other types of applications.
- Really consider the business model and justification — While there is a lot of excitement about AI, there is a question about the right business model to use with AI agents. If you are using a consumption model to deploy any part of your agent, there may be costs associated with how the agent performs. For instance, the more complex your agent, the more “thinking” the LLM may need to do. This can lead to costs that are different than what you may expect today. Also, it’s not clear yet how vendors plan to monetize agents. Will you pay by the agent? By the task? Will it be a dedicated or an on-demand service? These are the ins and outs you need to consider before you go too far down a road that doesn’t work for you.
- If you aren’t already doing observability and some FinOps, you need to get started — AI agents are simply different from what you’ve deployed before, and how they can scale and perform will be different. If you are not using more advanced management and monitoring tools now, AI may be the reason you need to consider a serious upgrade to ensure smooth operations and no surprise bills from your providers.
Where Salesforce Could Go From Here
What Salesforce announced at Dreamforce is impressive, and the company has done a very good job of demystifying AI and presenting an aligned product strategy. AI-curious Salesforce customers have a great place to get started. There are of course some areas where Salesforce could take things even further.
- Application platforms like Salesforce are typically used for internal applications. This is a function of pricing/packaging, the incumbent IT management stack, scaling and performance considerations. If Salesforce wants to make a big mark when it comes to customer service, for example, it may have to work hard to overcome the perception of internal use only.
- I’d like to see and hear more from Salesforce about testing, observability and governance. While I did tune into a few sessions on these topics, I left them all wanting more. This could become a major barrier in going from sandboxes to production.
- A very specific governance challenge I am wondering about is this: What will Salesforce do to prevent agent sprawl? Salesforce power users are notorious for creating data and dashboard sprawl, and AI applications will come with a higher cost in terms of infrastructure than a dashboard or custom form. Will every sales or marketing manager want to create their own agents? Companies will want some assurance that they will avoid getting into a bind over performance or cost due to sprawl.
- Many of Salesforce’s important tools such as MuleSoft, Tableau and Data Cloud already work very well outside of the Agentforce ecosystem. It might be worthwhile for Salesforce to speak to those use cases as well, especially in the early stages of the AI market where customers may elect a best-of-breed solution.
Based on what Marc Benioff and other Salesforce leaders said at Dreamforce, they believe they have a winner on their hands with Agentforce. I think they might—but the real wins and losses, in both business and technical terms, will require close attention to the details I’ve laid out here.
Read the full article here
