As we inch towards the next Dreamforce conference, we also approach Agentforce’s second birthday. They grow up fast, don’t they?
The last two years, and the last year especially, have been critical for Agentforce in terms of development and trial and error. A lot of the updates haven’t been perfect, and although it can be argued that we’re either out of the hype cycle or easing out of it, Salesforce’s proprietary AI product has been taken a lot more seriously. But what have we actually learned in this time, and better still – what has Salesforce learned?
The Journey of Agentforce
Agentforce has gone through many different iterations (I think we can count five) in its short lifespan, bringing new features, agents, and advancements on board each time. The most recent version is Agentforce 360, announced at Dreamforce 2025, and is made up of four key components: the Agentforce 360 platform, Data 360, Customer 360 apps, and Slack.
The situation became slightly more confusing when Salesforce announced Headless 360 at this year’s TrailblazerDX conference in April – an AI layer within Salesforce that means almost everything on the platform becomes surfaced through MCP, which is usable by agents.
Since then, we’ve seen the addition of new tools and features to Agentforce and its connected platform, including Slackbot, Claude Tag, and Agentforce Coworker. Plus, just last week, we also saw Agentforce’s pricing change yet again to introduce its pay-per-resolution option.
Both investors and analysts have been harsh on the product, continuously scrutinizing it and what it means for Salesforce’s future in both the SaaS space and the AI space. It has been a notable driving force behind the CRM giant’s turbulent stock price, with Wall Street still not convinced it has what it takes to compete against other industry players, including some of the AI titans like Anthropic and OpenAI.
However, it is impossible to ignore the progress that has been made with Agentforce. From the development of Salesforce’s Help Agent to the company’s acknowledgement of community concerns, the Agentforce we have today is very different from the Agentforce of a year ago – even the Agentforce of six months ago. But what has Salesforce actually learned from all of this?
“Keeping Up With Everything Has Been a Challenge”
At this year’s Agentforce World Tour in London, I got the chance to sit down with Paul O’Sullivan, the SVP of Solution Engineering at Salesforce. Through our conversation, I was able to discern that Salesforce had indeed felt the enormity of its trial-and-error period with Agentforce, learning from mistakes so that its customer base didn’t necessarily need to.
“I think the pace of change, and making sure we can keep up with everything, has been a challenge,” he told SF Ben. “If we’re being honest, we’ve made mistakes, and we’ve learned. We’ve done what we always do, which is partner really closely with some of our customers to go through this change with them because we’re changing and they’re changing.”
This can be directly observed through the work that Salesforce’s Forward Deployed Engineers (FDEs) do. FDEs work at the intersection of Salesforce’s product teams and services teams, helping customers get their first few agents off the ground, and act as the intermediary between Salesforce’s very new AI technology and a customer’s existing tech stack. The FDE team is one of the fastest-growing teams at Salesforce.
Paul emphasized that Salesforce has had to go through a multi-faceted transformation when it comes to Agentforce due to the rapid speed of change. The risk has always been rushing ahead when its customers are still not ready.
“It’s not just the technology changing – it’s the people and operating model too,” he said. “So it’s a threefold transformation internally at Salesforce as much as it is in our customers’ domain as well. I think we didn’t really anticipate that.
“We know the frontier models are pushing boundaries, but the reality is our customers need something that’s grounded and contextual in their business and what they’re trying to achieve.”
Some of Salesforce’s internal changes to Agentforce include the shift away from LLMs to more deterministic models due to persisting hallucinations. The SaaS giant also learned a lot from its Help Agent, as aforementioned, with Paul explaining that the company used its internal learnings to help advise its customers.
Salesforce was initially happy with the Help Agent’s accuracy in early testing, but that accuracy level dropped when it became generally available.
“What I will say is, I see a pattern between some of the challenges that I think customers and Salesforce have dealt with,” he said. “That’s a common customer challenge as well, right, so there’s a pitfall there.
“I’ve seen customers put agents live, and typically they’ll test it, and it’ll have a high 80s-90% accuracy rate, and when they go live, it will drop to 50-60%.”
Agents that depend on a high accuracy level, like customer agents, are one of the most popular agents companies are deploying right now. With Salesforce learning from its customers and testing its agents to see where they can improve, the company is able to provide support where it’s needed without lengthy user surveys or extensive outreach.
Final Thoughts
Agentforce still isn’t perfect, and it likely never will be. But there is no denying that Salesforce has continuously been making an active effort to learn about the critiques and problems of its customers and community to bring Agentforce closer to the AI tool everyone is eager to use.
Perhaps Wall Street will never be satisfied, at least whilst other AI leaders continue to grow, but getting its customer base on board with the product is equally, if not more, important. There is still work to be done, but we could be looking at a very different Agentforce by the time Dreamforce rolls around in September.
