Agentforce AI: From Confusion to Vision
Written in totality with time and effort by a human 100%
By: Ari Leeds, CEO, LeedsSource LLC
Agentforce has had my head cocked sideways like a confused labrador for quite some time. When it was first announced, it was exciting because it showed to Salesforce admins and developers like a set of fresh new flow features that were going to fit right into your existing build just like the erector-set like building bricks Salesforce is known for. Just a few simple prompt templates and instructions and “Bam”, you have entered the world of agentic AI! The marketing was really good, not to mention the amazing Agentforce drone show at Dreamforce to get us all locked in. I hit the trailhead training and got to work and that’s when the confusion began. I followed the instructions filled in the forms, got my first customer support agent up and running in my safe little dev org and I realized it was the first thing I ever built in Salesforce where I had no idea what the heck I just did. I couldn’t see the gears turning or the dominos falling. And yes, I understand the black-box nature of AI, but I couldn’t even see why any of what I did even produced the thing it produced. No problem, I figured it would click, so I did more in practice, bought some credits in my org and tried to get something off the ground. I had absorbed enough of the concepts to pass the Agentforce Specialist certification exam (the more advanced of the 2 available at the time). The reality in my own org was a different story. I could hardly figure out how to enable it. My sequence was something like this:
- Turn the thing on;
- OK, it needs a user;
- That user needs permissions, got it;
- Wait, I need permissions to build it…which ones exactly?;
- Does anyone need permission to interact with it?
- Do I need Data Cloud now too?
- It doesn’t seem to understand my build or schema;
- Why do I need to prebuild it with prompts? Isn’t that what the end user is going to do? Prompt it?
Bear in mind that this was not yesterday, this was basically the first generally available version of Agentforce but maybe some of you have had these feelings. A little further down the road and it started becoming:
- My clients are happy with Copilot, I’m crushing it with Claude, I can get this up and running but it sure burns a lot of calories.
But, I wasn’t going to give up quite yet because Salesforce had been putting the hard sell on my clients and they were starting to look to me for value and answers. Then more products rolled out like the SDR Agent and the Setup Agent and Agentforce Grid (not sure if this is even GA yet, but saw it at a conference a while ago). The developer and solution architect in me was still struggling to see what layer of the platform it even operated on and who it was for in concrete terms. I understood the value of Agentic AI, I just couldn’t tell where Agentforce lived. I understand Apex, flows, LWC, front-end, back-end, APIs and Objects - I can see them in my mind's eye like clockwork but these Agentforce tools and abilities just seemed strewn about. I started to question some of the product team in person about where these tools lived in the platform (not from a navigation standpoint, but from a build standpoint). Is the Grid a tool for admins? super users? Marketers? Sales Reps? Does it require you to build an Agent first? What technology building blocks is it consuming? Are these out-of-the-box agents something different or are they just pre-built agents providing a head start to building your own? Even the product marketers at Salesforce, demoing these things couldn’t provide a straight answer. It was frustrating and quite shocking.
I turned my back on it for a little bit and decided to let it ripen and mature more before revisiting. In the meantime Anthropics Claude entered my life and changed my business completely. Yes, I had been using AI in doses for a while to help with some stuff but suddenly I didn’t just have a chat bot or a helper to create content or do research, I had a new employee. I was blazing through work because I had Claude taking immense hours of typing and clicking off my plate. I instructed my project managers to give it a try and to start coming up with a framework to make sure it was creating operational consistency in our company and not becoming everyone's own personal 1-off assistant. This is where I think a lot of companies get stuck even if they are AI-savvy. All the employees are using AI but the company has to trust that each employee is following the guidelines. I wanted to do it differently and ensure the AI had the company policies built-in. We enforced our procedures for ticket/userstory writing (we use ClickUp), Code, Unit Testing, Deployment and Dev-Ops, Version Control with Github, and production data and security guards. We built our own software for centralizing Claude onboarding, configuration, documentation, adoption and centralized command over SOP updates. This meant that we could implement new company policies and mandates for AI into a central location and they would automatically be dispatched to all company users. Nobody could have their agents running rogue because each installation was connected back to our central command and integration center which we happened to build into our Salesforce org. This was a game-changer for the business. Our Kanban boards are now meticulous, our dev-ops are perfectly streamlined, and our QA is better than ever. As time went on, I began to envision a world of centralized AI command for enterprises to bring AI usage out of the world of Shadow IT and into controllable modules with business process and purpose in mind and really scale this AI powerhouse on rails!
That’s when it hit me - this is what Agentforce is for! It seems so obvious now. It’s a collection of AI tooling modules that is centrally controlled in your business operations system (Salesforce), defined by your business and doing the things that your business does. It is the enterprise version of AI that I had been imagining and building for my company. You have access to almost model for every purpose, the prompt templates exist because you have to bake in business operating procedures and proper behaviors that just work no matter who is using it, actions let you put an AI brain right in the middle of existing automation or let you put existing automation right inside of an AI Agent, retrievers and libraries let you define the operating landscape and context, and so on and so on. They did build all the stuff after all and it's not as simple as just chatting with a model out of the box because it's not supposed to be when it comes to enterprise. Just like training an employee on your processes to operate successfully in a business, you have to train your agents and they need to be tied back to a central system of controls. Yes, using Claude Code or OpenAI or Gemini is inherently just easier than Agentforce but it’s also ungoverned, doesn’t care about your company guidelines and the users using it may not even know what it’s doing to accomplish what it needs to accomplish. Agentforce is built for business plain and simple and if you're expecting to have a know-it-all bot right out the box, then you’re going to be disappointed. Unfortunately too many SMB Account Executives are selling it this way. Is it way easier to flip it on than before? Absolutely. I think Salesforce had to do this because consumers weren’t understanding the true value. It will give you the basics, but without a truly proper implementation where your Agent is trained on your business’ unique implementation of Salesforce, you’ll be wondering why this bot just isn’t giving you what you need.
Maybe you don’t really need it, maybe you do, but you need to understand what it’s really for and I don’t really think Salesforce has quite nailed the delivery on this. They’ve hit the high level sales pitch and the demos of possibility, but I feel that they’ve missed the mark on explaining why anyone should use this instead of all the other, more intuitive options.
We are now building out Agentforce implementations the right way, with purpose and deep understanding, but I personally like to make sure my clients truly understand what they are getting and why it’s valuable to their organization. For one client we allowed them to take PDFs provided by their suppliers, drop it into Salesforce, have an agent pick it up, read it, and fill in about 15 minutes of data entry in seconds. It then hands off the process like a person would, causing all the remaining downstream automations to run. It's so seamlessly integrated into their process that the end-users don’t even need to enter a prompt - it just acts like another doing a job. If you’re the type of user looking for help on your slide decks and spreadsheets and emails in a vacuum, then you’ll probably be disappointed. But, if you want enterprise AI that’s purpose built for scale and control at your company, and you're willing to invest in a trainer (implementor), then you might just be ready to truly take advantage. It’s not perfect and we are still going to continue building our special version of enterprise Claude capabilities, but we are definitely implementing some incredible Agentforce solutions too now that the vision is clear.
